annotate cabac.h @ 5083:ce36118abbbb libavcodec

rename attribute_unused to av_unused and moves its declaration to common.h patch by Carl Eugen Hoyos cehoyos chez ag or at original thread: [FFmpeg-devel] [PATCH] attribute_unused -> av_unused date: 05/29/2007 01:23 PM
author benoit
date Wed, 30 May 2007 09:32:25 +0000
parents 9a6a0818e93f
children 4dbe6578f811
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
1287
michaelni
parents:
diff changeset
1 /*
michaelni
parents:
diff changeset
2 * H.26L/H.264/AVC/JVT/14496-10/... encoder/decoder
michaelni
parents:
diff changeset
3 * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
michaelni
parents:
diff changeset
4 *
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3946
diff changeset
5 * This file is part of FFmpeg.
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3946
diff changeset
6 *
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3946
diff changeset
7 * FFmpeg is free software; you can redistribute it and/or
1287
michaelni
parents:
diff changeset
8 * modify it under the terms of the GNU Lesser General Public
michaelni
parents:
diff changeset
9 * License as published by the Free Software Foundation; either
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3946
diff changeset
10 * version 2.1 of the License, or (at your option) any later version.
1287
michaelni
parents:
diff changeset
11 *
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3946
diff changeset
12 * FFmpeg is distributed in the hope that it will be useful,
1287
michaelni
parents:
diff changeset
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
michaelni
parents:
diff changeset
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
michaelni
parents:
diff changeset
15 * Lesser General Public License for more details.
michaelni
parents:
diff changeset
16 *
michaelni
parents:
diff changeset
17 * You should have received a copy of the GNU Lesser General Public
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3946
diff changeset
18 * License along with FFmpeg; if not, write to the Free Software
3036
0b546eab515d Update licensing information: The FSF changed postal address.
diego
parents: 2967
diff changeset
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
1287
michaelni
parents:
diff changeset
20 *
michaelni
parents:
diff changeset
21 */
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
22
1287
michaelni
parents:
diff changeset
23 /**
michaelni
parents:
diff changeset
24 * @file cabac.h
michaelni
parents:
diff changeset
25 * Context Adaptive Binary Arithmetic Coder.
michaelni
parents:
diff changeset
26 */
michaelni
parents:
diff changeset
27
4975
9a6a0818e93f split h264.c to move parser in its own file
aurel
parents: 4908
diff changeset
28 #ifndef CABAC_H
9a6a0818e93f split h264.c to move parser in its own file
aurel
parents: 4908
diff changeset
29 #define CABAC_H
9a6a0818e93f split h264.c to move parser in its own file
aurel
parents: 4908
diff changeset
30
9a6a0818e93f split h264.c to move parser in its own file
aurel
parents: 4908
diff changeset
31 #include "bitstream.h"
1287
michaelni
parents:
diff changeset
32
3284
a224d9752912 don't force asserts in release builds. 2% faster h264.
lorenm
parents: 3036
diff changeset
33 //#undef NDEBUG
1287
michaelni
parents:
diff changeset
34 #include <assert.h>
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
35 #ifdef ARCH_X86
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
36 #include "x86_cpu.h"
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
37 #endif
1287
michaelni
parents:
diff changeset
38
3948
3edbf131ee44 refill cabac variables in 16bit steps, 3% faster get_cabac()
michael
parents: 3947
diff changeset
39 #define CABAC_BITS 16
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
40 #define CABAC_MASK ((1<<CABAC_BITS)-1)
3984
bb186452e7da BRANCHLESS_CABAD --> BRANCHLESS_CABAC_DECODER
diego
parents: 3982
diff changeset
41 #define BRANCHLESS_CABAC_DECODER 1
4039
866a83726985 Kill a warning with MSVC
gpoirier
parents: 4037
diff changeset
42 //#define ARCH_X86_DISABLED 1
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
43
1287
michaelni
parents:
diff changeset
44 typedef struct CABACContext{
michaelni
parents:
diff changeset
45 int low;
michaelni
parents:
diff changeset
46 int range;
michaelni
parents:
diff changeset
47 int outstanding_count;
michaelni
parents:
diff changeset
48 #ifdef STRICT_LIMITS
michaelni
parents:
diff changeset
49 int symCount;
michaelni
parents:
diff changeset
50 #endif
2024
f65d87bfdd5a some of the warning fixes by (Michael Roitzsch <mroi at users dot sourceforge dot net>)
michael
parents: 1787
diff changeset
51 const uint8_t *bytestream_start;
f65d87bfdd5a some of the warning fixes by (Michael Roitzsch <mroi at users dot sourceforge dot net>)
michael
parents: 1787
diff changeset
52 const uint8_t *bytestream;
2116
48d9f86fb047 overread fix
michael
parents: 2024
diff changeset
53 const uint8_t *bytestream_end;
1287
michaelni
parents:
diff changeset
54 PutBitContext pb;
michaelni
parents:
diff changeset
55 }CABACContext;
michaelni
parents:
diff changeset
56
4014
b2582438effe dehack *ps_state indexing in the branchless decoder
michael
parents: 4012
diff changeset
57 extern uint8_t ff_h264_mlps_state[4*64];
4039
866a83726985 Kill a warning with MSVC
gpoirier
parents: 4037
diff changeset
58 extern uint8_t ff_h264_lps_range[4*2*64]; ///< rangeTabLPS
3993
8b7c59b7af01 make state transition tables global as they are constant and the code is slightly faster that way
michael
parents: 3992
diff changeset
59 extern uint8_t ff_h264_mps_state[2*64]; ///< transIdxMPS
8b7c59b7af01 make state transition tables global as they are constant and the code is slightly faster that way
michael
parents: 3992
diff changeset
60 extern uint8_t ff_h264_lps_state[2*64]; ///< transIdxLPS
4024
d550343b5dac shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents: 4014
diff changeset
61 extern const uint8_t ff_h264_norm_shift[512];
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
62
1287
michaelni
parents:
diff changeset
63
michaelni
parents:
diff changeset
64 void ff_init_cabac_encoder(CABACContext *c, uint8_t *buf, int buf_size);
2024
f65d87bfdd5a some of the warning fixes by (Michael Roitzsch <mroi at users dot sourceforge dot net>)
michael
parents: 1787
diff changeset
65 void ff_init_cabac_decoder(CABACContext *c, const uint8_t *buf, int buf_size);
3993
8b7c59b7af01 make state transition tables global as they are constant and the code is slightly faster that way
michael
parents: 3992
diff changeset
66 void ff_init_cabac_states(CABACContext *c);
1287
michaelni
parents:
diff changeset
67
michaelni
parents:
diff changeset
68
michaelni
parents:
diff changeset
69 static inline void put_cabac_bit(CABACContext *c, int b){
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
70 put_bits(&c->pb, 1, b);
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
71 for(;c->outstanding_count; c->outstanding_count--){
1287
michaelni
parents:
diff changeset
72 put_bits(&c->pb, 1, 1-b);
michaelni
parents:
diff changeset
73 }
michaelni
parents:
diff changeset
74 }
michaelni
parents:
diff changeset
75
michaelni
parents:
diff changeset
76 static inline void renorm_cabac_encoder(CABACContext *c){
michaelni
parents:
diff changeset
77 while(c->range < 0x100){
michaelni
parents:
diff changeset
78 //FIXME optimize
michaelni
parents:
diff changeset
79 if(c->low<0x100){
michaelni
parents:
diff changeset
80 put_cabac_bit(c, 0);
michaelni
parents:
diff changeset
81 }else if(c->low<0x200){
michaelni
parents:
diff changeset
82 c->outstanding_count++;
michaelni
parents:
diff changeset
83 c->low -= 0x100;
michaelni
parents:
diff changeset
84 }else{
michaelni
parents:
diff changeset
85 put_cabac_bit(c, 1);
michaelni
parents:
diff changeset
86 c->low -= 0x200;
michaelni
parents:
diff changeset
87 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
88
1287
michaelni
parents:
diff changeset
89 c->range+= c->range;
michaelni
parents:
diff changeset
90 c->low += c->low;
michaelni
parents:
diff changeset
91 }
michaelni
parents:
diff changeset
92 }
michaelni
parents:
diff changeset
93
3928
987fffdf6ae7 don't try to inline cabac functions. gcc ignored the hint anyway, and forcing it would make h264 slower.
lorenm
parents: 3642
diff changeset
94 static void put_cabac(CABACContext *c, uint8_t * const state, int bit){
4024
d550343b5dac shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents: 4014
diff changeset
95 int RangeLPS= ff_h264_lps_range[2*(c->range&0xC0) + *state];
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
96
1287
michaelni
parents:
diff changeset
97 if(bit == ((*state)&1)){
michaelni
parents:
diff changeset
98 c->range -= RangeLPS;
3993
8b7c59b7af01 make state transition tables global as they are constant and the code is slightly faster that way
michael
parents: 3992
diff changeset
99 *state= ff_h264_mps_state[*state];
1287
michaelni
parents:
diff changeset
100 }else{
michaelni
parents:
diff changeset
101 c->low += c->range - RangeLPS;
michaelni
parents:
diff changeset
102 c->range = RangeLPS;
3993
8b7c59b7af01 make state transition tables global as they are constant and the code is slightly faster that way
michael
parents: 3992
diff changeset
103 *state= ff_h264_lps_state[*state];
1287
michaelni
parents:
diff changeset
104 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
105
1287
michaelni
parents:
diff changeset
106 renorm_cabac_encoder(c);
michaelni
parents:
diff changeset
107
michaelni
parents:
diff changeset
108 #ifdef STRICT_LIMITS
michaelni
parents:
diff changeset
109 c->symCount++;
michaelni
parents:
diff changeset
110 #endif
michaelni
parents:
diff changeset
111 }
michaelni
parents:
diff changeset
112
3928
987fffdf6ae7 don't try to inline cabac functions. gcc ignored the hint anyway, and forcing it would make h264 slower.
lorenm
parents: 3642
diff changeset
113 static void put_cabac_static(CABACContext *c, int RangeLPS, int bit){
1287
michaelni
parents:
diff changeset
114 assert(c->range > RangeLPS);
michaelni
parents:
diff changeset
115
michaelni
parents:
diff changeset
116 if(!bit){
michaelni
parents:
diff changeset
117 c->range -= RangeLPS;
michaelni
parents:
diff changeset
118 }else{
michaelni
parents:
diff changeset
119 c->low += c->range - RangeLPS;
michaelni
parents:
diff changeset
120 c->range = RangeLPS;
michaelni
parents:
diff changeset
121 }
michaelni
parents:
diff changeset
122
michaelni
parents:
diff changeset
123 renorm_cabac_encoder(c);
michaelni
parents:
diff changeset
124
michaelni
parents:
diff changeset
125 #ifdef STRICT_LIMITS
michaelni
parents:
diff changeset
126 c->symCount++;
michaelni
parents:
diff changeset
127 #endif
michaelni
parents:
diff changeset
128 }
michaelni
parents:
diff changeset
129
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
130 /**
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
131 * @param bit 0 -> write zero bit, !=0 write one bit
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
132 */
3928
987fffdf6ae7 don't try to inline cabac functions. gcc ignored the hint anyway, and forcing it would make h264 slower.
lorenm
parents: 3642
diff changeset
133 static void put_cabac_bypass(CABACContext *c, int bit){
1287
michaelni
parents:
diff changeset
134 c->low += c->low;
michaelni
parents:
diff changeset
135
michaelni
parents:
diff changeset
136 if(bit){
michaelni
parents:
diff changeset
137 c->low += c->range;
michaelni
parents:
diff changeset
138 }
michaelni
parents:
diff changeset
139 //FIXME optimize
michaelni
parents:
diff changeset
140 if(c->low<0x200){
michaelni
parents:
diff changeset
141 put_cabac_bit(c, 0);
michaelni
parents:
diff changeset
142 }else if(c->low<0x400){
michaelni
parents:
diff changeset
143 c->outstanding_count++;
michaelni
parents:
diff changeset
144 c->low -= 0x200;
michaelni
parents:
diff changeset
145 }else{
michaelni
parents:
diff changeset
146 put_cabac_bit(c, 1);
michaelni
parents:
diff changeset
147 c->low -= 0x400;
michaelni
parents:
diff changeset
148 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
149
1287
michaelni
parents:
diff changeset
150 #ifdef STRICT_LIMITS
michaelni
parents:
diff changeset
151 c->symCount++;
michaelni
parents:
diff changeset
152 #endif
michaelni
parents:
diff changeset
153 }
michaelni
parents:
diff changeset
154
1300
e18667d1e94d FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents: 1298
diff changeset
155 /**
e18667d1e94d FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents: 1298
diff changeset
156 *
e18667d1e94d FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents: 1298
diff changeset
157 * @return the number of bytes written
e18667d1e94d FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents: 1298
diff changeset
158 */
3928
987fffdf6ae7 don't try to inline cabac functions. gcc ignored the hint anyway, and forcing it would make h264 slower.
lorenm
parents: 3642
diff changeset
159 static int put_cabac_terminate(CABACContext *c, int bit){
1287
michaelni
parents:
diff changeset
160 c->range -= 2;
michaelni
parents:
diff changeset
161
michaelni
parents:
diff changeset
162 if(!bit){
michaelni
parents:
diff changeset
163 renorm_cabac_encoder(c);
michaelni
parents:
diff changeset
164 }else{
michaelni
parents:
diff changeset
165 c->low += c->range;
michaelni
parents:
diff changeset
166 c->range= 2;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
167
1287
michaelni
parents:
diff changeset
168 renorm_cabac_encoder(c);
michaelni
parents:
diff changeset
169
michaelni
parents:
diff changeset
170 assert(c->low <= 0x1FF);
michaelni
parents:
diff changeset
171 put_cabac_bit(c, c->low>>9);
michaelni
parents:
diff changeset
172 put_bits(&c->pb, 2, ((c->low>>7)&3)|1);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
173
1287
michaelni
parents:
diff changeset
174 flush_put_bits(&c->pb); //FIXME FIXME FIXME XXX wrong
michaelni
parents:
diff changeset
175 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
176
1287
michaelni
parents:
diff changeset
177 #ifdef STRICT_LIMITS
michaelni
parents:
diff changeset
178 c->symCount++;
michaelni
parents:
diff changeset
179 #endif
1300
e18667d1e94d FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents: 1298
diff changeset
180
1787
752b51a3c8ed get_bit_count -> put_bits_count
alex
parents: 1301
diff changeset
181 return (put_bits_count(&c->pb)+7)>>3;
1287
michaelni
parents:
diff changeset
182 }
michaelni
parents:
diff changeset
183
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
184 /**
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
185 * put (truncated) unary binarization.
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
186 */
3928
987fffdf6ae7 don't try to inline cabac functions. gcc ignored the hint anyway, and forcing it would make h264 slower.
lorenm
parents: 3642
diff changeset
187 static void put_cabac_u(CABACContext *c, uint8_t * state, int v, int max, int max_index, int truncated){
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
188 int i;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
189
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
190 assert(v <= max);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
191
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
192 #if 1
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
193 for(i=0; i<v; i++){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
194 put_cabac(c, state, 1);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
195 if(i < max_index) state++;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
196 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
197 if(truncated==0 || v<max)
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
198 put_cabac(c, state, 0);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
199 #else
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
200 if(v <= max_index){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
201 for(i=0; i<v; i++){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
202 put_cabac(c, state+i, 1);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
203 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
204 if(truncated==0 || v<max)
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
205 put_cabac(c, state+i, 0);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
206 }else{
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
207 for(i=0; i<=max_index; i++){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
208 put_cabac(c, state+i, 1);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
209 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
210 for(; i<v; i++){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
211 put_cabac(c, state+max_index, 1);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
212 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
213 if(truncated==0 || v<max)
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
214 put_cabac(c, state+max_index, 0);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
215 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
216 #endif
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
217 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
218
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
219 /**
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
220 * put unary exp golomb k-th order binarization.
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
221 */
3928
987fffdf6ae7 don't try to inline cabac functions. gcc ignored the hint anyway, and forcing it would make h264 slower.
lorenm
parents: 3642
diff changeset
222 static void put_cabac_ueg(CABACContext *c, uint8_t * state, int v, int max, int is_signed, int k, int max_index){
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
223 int i;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
224
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
225 if(v==0)
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
226 put_cabac(c, state, 0);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
227 else{
1298
5bc3184810dc cleanup
michaelni
parents: 1290
diff changeset
228 const int sign= v < 0;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
229
4001
34fdffe98bd0 Rename ABS macro to FFABS.
diego
parents: 3999
diff changeset
230 if(is_signed) v= FFABS(v);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
231
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
232 if(v<max){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
233 for(i=0; i<v; i++){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
234 put_cabac(c, state, 1);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
235 if(i < max_index) state++;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
236 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
237
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
238 put_cabac(c, state, 0);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
239 }else{
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
240 int m= 1<<k;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
241
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
242 for(i=0; i<max; i++){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
243 put_cabac(c, state, 1);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
244 if(i < max_index) state++;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
245 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
246
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
247 v -= max;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
248 while(v >= m){ //FIXME optimize
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
249 put_cabac_bypass(c, 1);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
250 v-= m;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
251 m+= m;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
252 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
253 put_cabac_bypass(c, 0);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
254 while(m>>=1){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
255 put_cabac_bypass(c, v&m);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
256 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
257 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
258
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
259 if(is_signed)
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
260 put_cabac_bypass(c, sign);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
261 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
262 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
263
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
264 static void refill(CABACContext *c){
1c39d9786efd optimization
michael
parents: 2116
diff changeset
265 #if CABAC_BITS == 16
3946
michael
parents: 3943
diff changeset
266 c->low+= (c->bytestream[0]<<9) + (c->bytestream[1]<<1);
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
267 #else
1c39d9786efd optimization
michael
parents: 2116
diff changeset
268 c->low+= c->bytestream[0]<<1;
1c39d9786efd optimization
michael
parents: 2116
diff changeset
269 #endif
1c39d9786efd optimization
michael
parents: 2116
diff changeset
270 c->low -= CABAC_MASK;
1c39d9786efd optimization
michael
parents: 2116
diff changeset
271 c->bytestream+= CABAC_BITS/8;
1c39d9786efd optimization
michael
parents: 2116
diff changeset
272 }
1c39d9786efd optimization
michael
parents: 2116
diff changeset
273
1c39d9786efd optimization
michael
parents: 2116
diff changeset
274 static void refill2(CABACContext *c){
1c39d9786efd optimization
michael
parents: 2116
diff changeset
275 int i, x;
1c39d9786efd optimization
michael
parents: 2116
diff changeset
276
1c39d9786efd optimization
michael
parents: 2116
diff changeset
277 x= c->low ^ (c->low-1);
4024
d550343b5dac shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents: 4014
diff changeset
278 i= 7 - ff_h264_norm_shift[x>>(CABAC_BITS-1)];
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
279
1c39d9786efd optimization
michael
parents: 2116
diff changeset
280 x= -CABAC_MASK;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
281
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
282 #if CABAC_BITS == 16
1c39d9786efd optimization
michael
parents: 2116
diff changeset
283 x+= (c->bytestream[0]<<9) + (c->bytestream[1]<<1);
1c39d9786efd optimization
michael
parents: 2116
diff changeset
284 #else
1c39d9786efd optimization
michael
parents: 2116
diff changeset
285 x+= c->bytestream[0]<<1;
1c39d9786efd optimization
michael
parents: 2116
diff changeset
286 #endif
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
287
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
288 c->low += x<<i;
1c39d9786efd optimization
michael
parents: 2116
diff changeset
289 c->bytestream+= CABAC_BITS/8;
1c39d9786efd optimization
michael
parents: 2116
diff changeset
290 }
1c39d9786efd optimization
michael
parents: 2116
diff changeset
291
1287
michaelni
parents:
diff changeset
292 static inline void renorm_cabac_decoder(CABACContext *c){
4024
d550343b5dac shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents: 4014
diff changeset
293 while(c->range < 0x100){
1287
michaelni
parents:
diff changeset
294 c->range+= c->range;
michaelni
parents:
diff changeset
295 c->low+= c->low;
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
296 if(!(c->low & CABAC_MASK))
1c39d9786efd optimization
michael
parents: 2116
diff changeset
297 refill(c);
1287
michaelni
parents:
diff changeset
298 }
michaelni
parents:
diff changeset
299 }
michaelni
parents:
diff changeset
300
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
301 static inline void renorm_cabac_decoder_once(CABACContext *c){
3951
751bfc30df72 disable benchmarking code
michael
parents: 3950
diff changeset
302 #ifdef ARCH_X86_DISABLED
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
303 int temp;
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
304 #if 0
3950
900d21b85dd6 renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents: 3948
diff changeset
305 //P3:683 athlon:475
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
306 asm(
4024
d550343b5dac shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents: 4014
diff changeset
307 "lea -0x100(%0), %2 \n\t"
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
308 "shr $31, %2 \n\t" //FIXME 31->63 for x86-64
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
309 "shl %%cl, %0 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
310 "shl %%cl, %1 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
311 : "+r"(c->range), "+r"(c->low), "+c"(temp)
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
312 );
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
313 #elif 0
3950
900d21b85dd6 renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents: 3948
diff changeset
314 //P3:680 athlon:474
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
315 asm(
4024
d550343b5dac shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents: 4014
diff changeset
316 "cmp $0x100, %0 \n\t"
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
317 "setb %%cl \n\t" //FIXME 31->63 for x86-64
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
318 "shl %%cl, %0 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
319 "shl %%cl, %1 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
320 : "+r"(c->range), "+r"(c->low), "+c"(temp)
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
321 );
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
322 #elif 1
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
323 int temp2;
3950
900d21b85dd6 renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents: 3948
diff changeset
324 //P3:665 athlon:517
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
325 asm(
4024
d550343b5dac shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents: 4014
diff changeset
326 "lea -0x100(%0), %%eax \n\t"
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
327 "cdq \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
328 "mov %0, %%eax \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
329 "and %%edx, %0 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
330 "and %1, %%edx \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
331 "add %%eax, %0 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
332 "add %%edx, %1 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
333 : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2)
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
334 );
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
335 #elif 0
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
336 int temp2;
3950
900d21b85dd6 renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents: 3948
diff changeset
337 //P3:673 athlon:509
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
338 asm(
4024
d550343b5dac shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents: 4014
diff changeset
339 "cmp $0x100, %0 \n\t"
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
340 "sbb %%edx, %%edx \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
341 "mov %0, %%eax \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
342 "and %%edx, %0 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
343 "and %1, %%edx \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
344 "add %%eax, %0 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
345 "add %%edx, %1 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
346 : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2)
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
347 );
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
348 #else
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
349 int temp2;
3950
900d21b85dd6 renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents: 3948
diff changeset
350 //P3:677 athlon:511
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
351 asm(
4024
d550343b5dac shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents: 4014
diff changeset
352 "cmp $0x100, %0 \n\t"
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
353 "lea (%0, %0), %%eax \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
354 "lea (%1, %1), %%edx \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
355 "cmovb %%eax, %0 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
356 "cmovb %%edx, %1 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
357 : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2)
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
358 );
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
359 #endif
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
360 #else
3950
900d21b85dd6 renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents: 3948
diff changeset
361 //P3:675 athlon:476
4024
d550343b5dac shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents: 4014
diff changeset
362 int shift= (uint32_t)(c->range - 0x100)>>31;
3642
0efda682253c tweak cabac. 0.5% faster h264.
lorenm
parents: 3284
diff changeset
363 c->range<<= shift;
0efda682253c tweak cabac. 0.5% faster h264.
lorenm
parents: 3284
diff changeset
364 c->low <<= shift;
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
365 #endif
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
366 if(!(c->low & CABAC_MASK))
1c39d9786efd optimization
michael
parents: 2116
diff changeset
367 refill(c);
1c39d9786efd optimization
michael
parents: 2116
diff changeset
368 }
1c39d9786efd optimization
michael
parents: 2116
diff changeset
369
4908
777f250df232 Fix multiple "¡Æinline/static¡Ç is not at beginning of declaration" warnings.
diego
parents: 4882
diff changeset
370 static av_always_inline int get_cabac_inline(CABACContext *c, uint8_t * const state){
3642
0efda682253c tweak cabac. 0.5% faster h264.
lorenm
parents: 3284
diff changeset
371 //FIXME gcc generates duplicate load/stores for c->low and c->range
3969
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
372 #define LOW "0"
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
373 #define RANGE "4"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
374 #ifdef ARCH_X86_64
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
375 #define BYTESTART "16"
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
376 #define BYTE "24"
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
377 #define BYTEEND "32"
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
378 #else
3993
8b7c59b7af01 make state transition tables global as they are constant and the code is slightly faster that way
michael
parents: 3992
diff changeset
379 #define BYTESTART "12"
8b7c59b7af01 make state transition tables global as they are constant and the code is slightly faster that way
michael
parents: 3992
diff changeset
380 #define BYTE "16"
8b7c59b7af01 make state transition tables global as they are constant and the code is slightly faster that way
michael
parents: 3992
diff changeset
381 #define BYTEEND "20"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
382 #endif
4882
8131ccb4ea72 Mark code parts that cannot work on AMD64 due to broken relocations as such.
diego
parents: 4881
diff changeset
383 #if defined(ARCH_X86) && defined(CONFIG_7REGS) && defined(HAVE_EBX_AVAILABLE) && !defined(BROKEN_RELOCATIONS)
4044
5ccdefd60f61 Fix PIC compilation, some defines were under #ifdef !PIC but used
diego
parents: 4043
diff changeset
384 int bit;
5ccdefd60f61 Fix PIC compilation, some defines were under #ifdef !PIC but used
diego
parents: 4043
diff changeset
385
3984
bb186452e7da BRANCHLESS_CABAD --> BRANCHLESS_CABAC_DECODER
diego
parents: 3982
diff changeset
386 #ifndef BRANCHLESS_CABAC_DECODER
3969
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
387 asm volatile(
4035
b7f31a32bb30 cosmetic (%%eax->%0)
michael
parents: 4034
diff changeset
388 "movzbl (%1), %0 \n\t"
3969
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
389 "movl "RANGE "(%2), %%ebx \n\t"
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
390 "movl "RANGE "(%2), %%edx \n\t"
4024
d550343b5dac shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents: 4014
diff changeset
391 "andl $0xC0, %%ebx \n\t"
4035
b7f31a32bb30 cosmetic (%%eax->%0)
michael
parents: 4034
diff changeset
392 "movzbl "MANGLE(ff_h264_lps_range)"(%0, %%ebx, 2), %%esi\n\t"
3969
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
393 "movl "LOW "(%2), %%ebx \n\t"
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
394 //eax:state ebx:low, edx:range, esi:RangeLPS
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
395 "subl %%esi, %%edx \n\t"
4024
d550343b5dac shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents: 4014
diff changeset
396 "movl %%edx, %%ecx \n\t"
d550343b5dac shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents: 4014
diff changeset
397 "shll $17, %%ecx \n\t"
d550343b5dac shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents: 4014
diff changeset
398 "cmpl %%ecx, %%ebx \n\t"
3969
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
399 " ja 1f \n\t"
3999
6cbad3675632 slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents: 3996
diff changeset
400
6cbad3675632 slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents: 3996
diff changeset
401 #if 1
6cbad3675632 slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents: 3996
diff changeset
402 //athlon:4067 P3:4110
4024
d550343b5dac shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents: 4014
diff changeset
403 "lea -0x100(%%edx), %%ecx \n\t"
3999
6cbad3675632 slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents: 3996
diff changeset
404 "shr $31, %%ecx \n\t"
6cbad3675632 slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents: 3996
diff changeset
405 "shl %%cl, %%edx \n\t"
6cbad3675632 slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents: 3996
diff changeset
406 "shl %%cl, %%ebx \n\t"
6cbad3675632 slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents: 3996
diff changeset
407 #else
6cbad3675632 slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents: 3996
diff changeset
408 //athlon:4057 P3:4130
4024
d550343b5dac shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents: 4014
diff changeset
409 "cmp $0x100, %%edx \n\t" //FIXME avoidable
3969
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
410 "setb %%cl \n\t"
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
411 "shl %%cl, %%edx \n\t"
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
412 "shl %%cl, %%ebx \n\t"
3999
6cbad3675632 slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents: 3996
diff changeset
413 #endif
4035
b7f31a32bb30 cosmetic (%%eax->%0)
michael
parents: 4034
diff changeset
414 "movzbl "MANGLE(ff_h264_mps_state)"(%0), %%ecx \n\t"
3969
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
415 "movb %%cl, (%1) \n\t"
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
416 //eax:state ebx:low, edx:range, esi:RangeLPS
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
417 "test %%bx, %%bx \n\t"
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
418 " jnz 2f \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
419 "mov "BYTE "(%2), %%"REG_S" \n\t"
3969
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
420 "subl $0xFFFF, %%ebx \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
421 "movzwl (%%"REG_S"), %%ecx \n\t"
3969
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
422 "bswap %%ecx \n\t"
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
423 "shrl $15, %%ecx \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
424 "add $2, %%"REG_S" \n\t"
3969
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
425 "addl %%ecx, %%ebx \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
426 "mov %%"REG_S", "BYTE "(%2) \n\t"
3969
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
427 "jmp 2f \n\t"
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
428 "1: \n\t"
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
429 //eax:state ebx:low, edx:range, esi:RangeLPS
4024
d550343b5dac shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents: 4014
diff changeset
430 "subl %%ecx, %%ebx \n\t"
3969
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
431 "movl %%esi, %%edx \n\t"
3979
ce16f66a48ad reading 8bit mem into a 8bit register needs 2 uops on P4, 8bit->32bit with zero extension needs just 1
michael
parents: 3978
diff changeset
432 "movzbl " MANGLE(ff_h264_norm_shift) "(%%esi), %%ecx \n\t"
3969
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
433 "shll %%cl, %%ebx \n\t"
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
434 "shll %%cl, %%edx \n\t"
4035
b7f31a32bb30 cosmetic (%%eax->%0)
michael
parents: 4034
diff changeset
435 "movzbl "MANGLE(ff_h264_lps_state)"(%0), %%ecx \n\t"
4024
d550343b5dac shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents: 4014
diff changeset
436 "movb %%cl, (%1) \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
437 "add $1, %0 \n\t"
3969
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
438 "test %%bx, %%bx \n\t"
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
439 " jnz 2f \n\t"
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
440
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
441 "mov "BYTE "(%2), %%"REG_c" \n\t"
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
442 "movzwl (%%"REG_c"), %%esi \n\t"
3969
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
443 "bswap %%esi \n\t"
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
444 "shrl $15, %%esi \n\t"
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
445 "subl $0xFFFF, %%esi \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
446 "add $2, %%"REG_c" \n\t"
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
447 "mov %%"REG_c", "BYTE "(%2) \n\t"
3969
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
448
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
449 "leal -1(%%ebx), %%ecx \n\t"
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
450 "xorl %%ebx, %%ecx \n\t"
4024
d550343b5dac shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents: 4014
diff changeset
451 "shrl $15, %%ecx \n\t"
3979
ce16f66a48ad reading 8bit mem into a 8bit register needs 2 uops on P4, 8bit->32bit with zero extension needs just 1
michael
parents: 3978
diff changeset
452 "movzbl " MANGLE(ff_h264_norm_shift) "(%%ecx), %%ecx \n\t"
3994
2734b228fc87 use ecx instead of cl (no speed change on P3 but might avoid partial register stalls on some cpus)
michael
parents: 3993
diff changeset
453 "neg %%ecx \n\t"
2734b228fc87 use ecx instead of cl (no speed change on P3 but might avoid partial register stalls on some cpus)
michael
parents: 3993
diff changeset
454 "add $7, %%ecx \n\t"
3969
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
455
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
456 "shll %%cl , %%esi \n\t"
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
457 "addl %%esi, %%ebx \n\t"
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
458 "2: \n\t"
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
459 "movl %%edx, "RANGE "(%2) \n\t"
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
460 "movl %%ebx, "LOW "(%2) \n\t"
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
461 :"=&a"(bit) //FIXME this is fragile gcc either runs out of registers or misscompiles it (for example if "+a"(bit) or "+m"(*state) is used
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
462 :"r"(state), "r"(c)
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
463 : "%"REG_c, "%ebx", "%edx", "%"REG_S, "memory"
3969
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
464 );
3982
af16271634c2 moving another bit&1 out, this is as fast as with it in there, but it makes more sense with it outside of the loop
michael
parents: 3981
diff changeset
465 bit&=1;
4002
ec426fa57dfe adds some useful comments after some of the #else, #elseif,
gpoirier
parents: 4001
diff changeset
466 #else /* BRANCHLESS_CABAC_DECODER */
4046
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
467
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
468
4418
4cceb7c877af rename CMOV_IS_FAST to HAVE_FAST_CMOV and simplify configure
mru
parents: 4345
diff changeset
469 #if defined HAVE_FAST_CMOV
4046
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
470 #define BRANCHLESS_GET_CABAC_UPDATE(ret, cabac, statep, low, lowword, range, tmp, tmpbyte)\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
471 "mov "tmp" , %%ecx \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
472 "shl $17 , "tmp" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
473 "cmp "low" , "tmp" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
474 "cmova %%ecx , "range" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
475 "sbb %%ecx , %%ecx \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
476 "and %%ecx , "tmp" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
477 "sub "tmp" , "low" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
478 "xor %%ecx , "ret" \n\t"
4418
4cceb7c877af rename CMOV_IS_FAST to HAVE_FAST_CMOV and simplify configure
mru
parents: 4345
diff changeset
479 #else /* HAVE_FAST_CMOV */
4046
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
480 #define BRANCHLESS_GET_CABAC_UPDATE(ret, cabac, statep, low, lowword, range, tmp, tmpbyte)\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
481 "mov "tmp" , %%ecx \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
482 "shl $17 , "tmp" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
483 "sub "low" , "tmp" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
484 "sar $31 , "tmp" \n\t" /*lps_mask*/\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
485 "sub %%ecx , "range" \n\t" /*RangeLPS - range*/\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
486 "and "tmp" , "range" \n\t" /*(RangeLPS - range)&lps_mask*/\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
487 "add %%ecx , "range" \n\t" /*new range*/\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
488 "shl $17 , %%ecx \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
489 "and "tmp" , %%ecx \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
490 "sub %%ecx , "low" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
491 "xor "tmp" , "ret" \n\t"
4418
4cceb7c877af rename CMOV_IS_FAST to HAVE_FAST_CMOV and simplify configure
mru
parents: 4345
diff changeset
492 #endif /* HAVE_FAST_CMOV */
3975
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
493
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
494
4046
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
495 #define BRANCHLESS_GET_CABAC(ret, cabac, statep, low, lowword, range, tmp, tmpbyte)\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
496 "movzbl "statep" , "ret" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
497 "mov "range" , "tmp" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
498 "and $0xC0 , "range" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
499 "movzbl "MANGLE(ff_h264_lps_range)"("ret", "range", 2), "range" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
500 "sub "range" , "tmp" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
501 BRANCHLESS_GET_CABAC_UPDATE(ret, cabac, statep, low, lowword, range, tmp, tmpbyte)\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
502 "movzbl " MANGLE(ff_h264_norm_shift) "("range"), %%ecx \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
503 "shl %%cl , "range" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
504 "movzbl "MANGLE(ff_h264_mlps_state)"+128("ret"), "tmp" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
505 "mov "tmpbyte" , "statep" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
506 "shl %%cl , "low" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
507 "test "lowword" , "lowword" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
508 " jnz 1f \n\t"\
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
509 "mov "BYTE"("cabac"), %%"REG_c" \n\t"\
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
510 "movzwl (%%"REG_c") , "tmp" \n\t"\
4046
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
511 "bswap "tmp" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
512 "shr $15 , "tmp" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
513 "sub $0xFFFF , "tmp" \n\t"\
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
514 "add $2 , %%"REG_c" \n\t"\
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
515 "mov %%"REG_c" , "BYTE "("cabac") \n\t"\
4046
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
516 "lea -1("low") , %%ecx \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
517 "xor "low" , %%ecx \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
518 "shr $15 , %%ecx \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
519 "movzbl " MANGLE(ff_h264_norm_shift) "(%%ecx), %%ecx \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
520 "neg %%ecx \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
521 "add $7 , %%ecx \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
522 "shl %%cl , "tmp" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
523 "add "tmp" , "low" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
524 "1: \n\t"
3975
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
525
4046
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
526 asm volatile(
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
527 "movl "RANGE "(%2), %%esi \n\t"
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
528 "movl "LOW "(%2), %%ebx \n\t"
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
529 BRANCHLESS_GET_CABAC("%0", "%2", "(%1)", "%%ebx", "%%bx", "%%esi", "%%edx", "%%dl")
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
530 "movl %%esi, "RANGE "(%2) \n\t"
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
531 "movl %%ebx, "LOW "(%2) \n\t"
3975
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
532
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
533 :"=&a"(bit)
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
534 :"r"(state), "r"(c)
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
535 : "%"REG_c, "%ebx", "%edx", "%esi", "memory"
3975
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
536 );
3981
9854f686ba79 move the &1 out of the asm so gcc can optimize it away in inlined cases (yes this is slightly faster)
michael
parents: 3980
diff changeset
537 bit&=1;
4002
ec426fa57dfe adds some useful comments after some of the #else, #elseif,
gpoirier
parents: 4001
diff changeset
538 #endif /* BRANCHLESS_CABAC_DECODER */
4882
8131ccb4ea72 Mark code parts that cannot work on AMD64 due to broken relocations as such.
diego
parents: 4881
diff changeset
539 #else /* defined(ARCH_X86) && defined(CONFIG_7REGS) && defined(HAVE_EBX_AVAILABLE) && !defined(BROKEN_RELOCATIONS) */
3642
0efda682253c tweak cabac. 0.5% faster h264.
lorenm
parents: 3284
diff changeset
540 int s = *state;
4039
866a83726985 Kill a warning with MSVC
gpoirier
parents: 4037
diff changeset
541 int RangeLPS= ff_h264_lps_range[2*(c->range&0xC0) + s];
5083
ce36118abbbb rename attribute_unused to av_unused and moves its declaration to common.h
benoit
parents: 4975
diff changeset
542 int bit, lps_mask av_unused;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
543
1287
michaelni
parents:
diff changeset
544 c->range -= RangeLPS;
3984
bb186452e7da BRANCHLESS_CABAD --> BRANCHLESS_CABAC_DECODER
diego
parents: 3982
diff changeset
545 #ifndef BRANCHLESS_CABAC_DECODER
4345
88967250d718 replace a few hardcoded numbers with their correct named ones
michael
parents: 4283
diff changeset
546 if(c->low < (c->range<<(CABAC_BITS+1))){
3642
0efda682253c tweak cabac. 0.5% faster h264.
lorenm
parents: 3284
diff changeset
547 bit= s&1;
3993
8b7c59b7af01 make state transition tables global as they are constant and the code is slightly faster that way
michael
parents: 3992
diff changeset
548 *state= ff_h264_mps_state[s];
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
549 renorm_cabac_decoder_once(c);
1287
michaelni
parents:
diff changeset
550 }else{
4024
d550343b5dac shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents: 4014
diff changeset
551 bit= ff_h264_norm_shift[RangeLPS];
4345
88967250d718 replace a few hardcoded numbers with their correct named ones
michael
parents: 4283
diff changeset
552 c->low -= (c->range<<(CABAC_BITS+1));
3993
8b7c59b7af01 make state transition tables global as they are constant and the code is slightly faster that way
michael
parents: 3992
diff changeset
553 *state= ff_h264_lps_state[s];
3956
0910f2844f9a branchless renormalization (1% faster get_cabac) old branchless renormalization wasnt faster because gcc was scared of the shift variable (missusing bit variable now)
michael
parents: 3955
diff changeset
554 c->range = RangeLPS<<bit;
0910f2844f9a branchless renormalization (1% faster get_cabac) old branchless renormalization wasnt faster because gcc was scared of the shift variable (missusing bit variable now)
michael
parents: 3955
diff changeset
555 c->low <<= bit;
0910f2844f9a branchless renormalization (1% faster get_cabac) old branchless renormalization wasnt faster because gcc was scared of the shift variable (missusing bit variable now)
michael
parents: 3955
diff changeset
556 bit= (s&1)^1;
0910f2844f9a branchless renormalization (1% faster get_cabac) old branchless renormalization wasnt faster because gcc was scared of the shift variable (missusing bit variable now)
michael
parents: 3955
diff changeset
557
4345
88967250d718 replace a few hardcoded numbers with their correct named ones
michael
parents: 4283
diff changeset
558 if(!(c->low & CABAC_MASK)){
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
559 refill2(c);
3956
0910f2844f9a branchless renormalization (1% faster get_cabac) old branchless renormalization wasnt faster because gcc was scared of the shift variable (missusing bit variable now)
michael
parents: 3955
diff changeset
560 }
1287
michaelni
parents:
diff changeset
561 }
4002
ec426fa57dfe adds some useful comments after some of the #else, #elseif,
gpoirier
parents: 4001
diff changeset
562 #else /* BRANCHLESS_CABAC_DECODER */
4345
88967250d718 replace a few hardcoded numbers with their correct named ones
michael
parents: 4283
diff changeset
563 lps_mask= ((c->range<<(CABAC_BITS+1)) - c->low)>>31;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
564
4345
88967250d718 replace a few hardcoded numbers with their correct named ones
michael
parents: 4283
diff changeset
565 c->low -= (c->range<<(CABAC_BITS+1)) & lps_mask;
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
566 c->range += (RangeLPS - c->range) & lps_mask;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
567
3974
93746612bc78 optimize branchless C CABAC decoder
michael
parents: 3972
diff changeset
568 s^=lps_mask;
4014
b2582438effe dehack *ps_state indexing in the branchless decoder
michael
parents: 4012
diff changeset
569 *state= (ff_h264_mlps_state+128)[s];
3974
93746612bc78 optimize branchless C CABAC decoder
michael
parents: 3972
diff changeset
570 bit= s&1;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
571
4024
d550343b5dac shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents: 4014
diff changeset
572 lps_mask= ff_h264_norm_shift[c->range];
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
573 c->range<<= lps_mask;
1c39d9786efd optimization
michael
parents: 2116
diff changeset
574 c->low <<= lps_mask;
1c39d9786efd optimization
michael
parents: 2116
diff changeset
575 if(!(c->low & CABAC_MASK))
1c39d9786efd optimization
michael
parents: 2116
diff changeset
576 refill2(c);
4002
ec426fa57dfe adds some useful comments after some of the #else, #elseif,
gpoirier
parents: 4001
diff changeset
577 #endif /* BRANCHLESS_CABAC_DECODER */
4882
8131ccb4ea72 Mark code parts that cannot work on AMD64 due to broken relocations as such.
diego
parents: 4881
diff changeset
578 #endif /* defined(ARCH_X86) && defined(CONFIG_7REGS) && defined(HAVE_EBX_AVAILABLE) && !defined(BROKEN_RELOCATIONS) */
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
579 return bit;
1287
michaelni
parents:
diff changeset
580 }
michaelni
parents:
diff changeset
581
4579
03f7feb14dd3 use av_noinline instead of __attribute((noinline))
aurel
parents: 4418
diff changeset
582 static int av_noinline get_cabac_noinline(CABACContext *c, uint8_t * const state){
4008
b636f3d59283 prevent "mb level" get_cabac() calls from being inlined (3% faster decode_mb_cabac() on P3)
michael
parents: 4002
diff changeset
583 return get_cabac_inline(c,state);
b636f3d59283 prevent "mb level" get_cabac() calls from being inlined (3% faster decode_mb_cabac() on P3)
michael
parents: 4002
diff changeset
584 }
b636f3d59283 prevent "mb level" get_cabac() calls from being inlined (3% faster decode_mb_cabac() on P3)
michael
parents: 4002
diff changeset
585
b636f3d59283 prevent "mb level" get_cabac() calls from being inlined (3% faster decode_mb_cabac() on P3)
michael
parents: 4002
diff changeset
586 static int get_cabac(CABACContext *c, uint8_t * const state){
b636f3d59283 prevent "mb level" get_cabac() calls from being inlined (3% faster decode_mb_cabac() on P3)
michael
parents: 4002
diff changeset
587 return get_cabac_inline(c,state);
b636f3d59283 prevent "mb level" get_cabac() calls from being inlined (3% faster decode_mb_cabac() on P3)
michael
parents: 4002
diff changeset
588 }
b636f3d59283 prevent "mb level" get_cabac() calls from being inlined (3% faster decode_mb_cabac() on P3)
michael
parents: 4002
diff changeset
589
3928
987fffdf6ae7 don't try to inline cabac functions. gcc ignored the hint anyway, and forcing it would make h264 slower.
lorenm
parents: 3642
diff changeset
590 static int get_cabac_bypass(CABACContext *c){
4040
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
591 #if 0 //not faster
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
592 int bit;
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
593 asm volatile(
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
594 "movl "RANGE "(%1), %%ebx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
595 "movl "LOW "(%1), %%eax \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
596 "shl $17, %%ebx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
597 "add %%eax, %%eax \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
598 "sub %%ebx, %%eax \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
599 "cdq \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
600 "and %%edx, %%ebx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
601 "add %%ebx, %%eax \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
602 "test %%ax, %%ax \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
603 " jnz 1f \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
604 "movl "BYTE "(%1), %%"REG_b" \n\t"
4040
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
605 "subl $0xFFFF, %%eax \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
606 "movzwl (%%"REG_b"), %%ecx \n\t"
4040
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
607 "bswap %%ecx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
608 "shrl $15, %%ecx \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
609 "addl $2, %%"REG_b" \n\t"
4040
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
610 "addl %%ecx, %%eax \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
611 "movl %%"REG_b", "BYTE "(%1) \n\t"
4040
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
612 "1: \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
613 "movl %%eax, "LOW "(%1) \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
614
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
615 :"=&d"(bit)
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
616 :"r"(c)
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
617 : "%eax", "%"REG_b, "%ecx", "memory"
4040
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
618 );
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
619 return bit+1;
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
620 #else
4024
d550343b5dac shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents: 4014
diff changeset
621 int range;
1287
michaelni
parents:
diff changeset
622 c->low += c->low;
michaelni
parents:
diff changeset
623
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
624 if(!(c->low & CABAC_MASK))
1c39d9786efd optimization
michael
parents: 2116
diff changeset
625 refill(c);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
626
4345
88967250d718 replace a few hardcoded numbers with their correct named ones
michael
parents: 4283
diff changeset
627 range= c->range<<(CABAC_BITS+1);
4024
d550343b5dac shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents: 4014
diff changeset
628 if(c->low < range){
1287
michaelni
parents:
diff changeset
629 return 0;
michaelni
parents:
diff changeset
630 }else{
4024
d550343b5dac shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents: 4014
diff changeset
631 c->low -= range;
1287
michaelni
parents:
diff changeset
632 return 1;
michaelni
parents:
diff changeset
633 }
4040
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
634 #endif
1287
michaelni
parents:
diff changeset
635 }
4040
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
636
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
637
4283
d6f83e2f8804 rename always_inline to av_always_inline and move to common.h
mru
parents: 4241
diff changeset
638 static av_always_inline int get_cabac_bypass_sign(CABACContext *c, int val){
4241
92f773cfebf5 PIC fix
michael
parents: 4113
diff changeset
639 #if defined(ARCH_X86) && !(defined(PIC) && defined(__GNUC__))
4040
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
640 asm volatile(
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
641 "movl "RANGE "(%1), %%ebx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
642 "movl "LOW "(%1), %%eax \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
643 "shl $17, %%ebx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
644 "add %%eax, %%eax \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
645 "sub %%ebx, %%eax \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
646 "cdq \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
647 "and %%edx, %%ebx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
648 "add %%ebx, %%eax \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
649 "xor %%edx, %%ecx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
650 "sub %%edx, %%ecx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
651 "test %%ax, %%ax \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
652 " jnz 1f \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
653 "mov "BYTE "(%1), %%"REG_b" \n\t"
4040
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
654 "subl $0xFFFF, %%eax \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
655 "movzwl (%%"REG_b"), %%edx \n\t"
4040
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
656 "bswap %%edx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
657 "shrl $15, %%edx \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
658 "add $2, %%"REG_b" \n\t"
4040
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
659 "addl %%edx, %%eax \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
660 "mov %%"REG_b", "BYTE "(%1) \n\t"
4040
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
661 "1: \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
662 "movl %%eax, "LOW "(%1) \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
663
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
664 :"+c"(val)
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
665 :"r"(c)
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
666 : "%eax", "%"REG_b, "%edx", "memory"
4040
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
667 );
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
668 return val;
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
669 #else
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
670 int range, mask;
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
671 c->low += c->low;
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
672
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
673 if(!(c->low & CABAC_MASK))
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
674 refill(c);
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
675
4345
88967250d718 replace a few hardcoded numbers with their correct named ones
michael
parents: 4283
diff changeset
676 range= c->range<<(CABAC_BITS+1);
4040
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
677 c->low -= range;
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
678 mask= c->low >> 31;
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
679 range &= mask;
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
680 c->low += range;
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
681 return (val^mask)-mask;
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
682 #endif
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
683 }
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
684
4037
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
685 //FIXME the x86 code from this file should be moved into i386/h264 or cabac something.c/h (note ill kill you if you move my code away from under my fingers before iam finished with it!)
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
686 //FIXME use some macros to avoid duplicatin get_cabac (cant be done yet as that would make optimization work hard)
4882
8131ccb4ea72 Mark code parts that cannot work on AMD64 due to broken relocations as such.
diego
parents: 4881
diff changeset
687 #if defined(ARCH_X86) && defined(CONFIG_7REGS) && defined(HAVE_EBX_AVAILABLE) && !defined(BROKEN_RELOCATIONS)
4037
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
688 static int decode_significance_x86(CABACContext *c, int max_coeff, uint8_t *significant_coeff_ctx_base, int *index){
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
689 void *end= significant_coeff_ctx_base + max_coeff - 1;
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
690 int minusstart= -(int)significant_coeff_ctx_base;
4049
8c1a5ed03a00 another instruction less in decode_significance_x86() -> 1% faster ion P3
michael
parents: 4048
diff changeset
691 int minusindex= 4-(int)index;
4037
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
692 int coeff_count;
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
693 asm volatile(
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
694 "movl "RANGE "(%3), %%esi \n\t"
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
695 "movl "LOW "(%3), %%ebx \n\t"
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
696
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
697 "2: \n\t"
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
698
4047
61a4e7218a45 reordering instructions a little in decode_significance_x86() -> 2 instructions less / 1% faster decode_residual on P3
michael
parents: 4046
diff changeset
699 BRANCHLESS_GET_CABAC("%%edx", "%3", "(%1)", "%%ebx", "%%bx", "%%esi", "%%eax", "%%al")
4037
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
700
4047
61a4e7218a45 reordering instructions a little in decode_significance_x86() -> 2 instructions less / 1% faster decode_residual on P3
michael
parents: 4046
diff changeset
701 "test $1, %%edx \n\t"
4037
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
702 " jz 3f \n\t"
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
703
4047
61a4e7218a45 reordering instructions a little in decode_significance_x86() -> 2 instructions less / 1% faster decode_residual on P3
michael
parents: 4046
diff changeset
704 BRANCHLESS_GET_CABAC("%%edx", "%3", "61(%1)", "%%ebx", "%%bx", "%%esi", "%%eax", "%%al")
61a4e7218a45 reordering instructions a little in decode_significance_x86() -> 2 instructions less / 1% faster decode_residual on P3
michael
parents: 4046
diff changeset
705
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
706 "mov %2, %%"REG_a" \n\t"
4037
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
707 "movl %4, %%ecx \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
708 "add %1, %%"REG_c" \n\t"
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
709 "movl %%ecx, (%%"REG_a") \n\t"
4037
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
710
4047
61a4e7218a45 reordering instructions a little in decode_significance_x86() -> 2 instructions less / 1% faster decode_residual on P3
michael
parents: 4046
diff changeset
711 "test $1, %%edx \n\t"
4037
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
712 " jnz 4f \n\t"
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
713
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
714 "add $4, %%"REG_a" \n\t"
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
715 "mov %%"REG_a", %2 \n\t"
4048
bf6791303fa0 1 instruction less
michael
parents: 4047
diff changeset
716
4037
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
717 "3: \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
718 "add $1, %1 \n\t"
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
719 "cmp %5, %1 \n\t"
4037
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
720 " jb 2b \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
721 "mov %2, %%"REG_a" \n\t"
4037
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
722 "movl %4, %%ecx \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
723 "add %1, %%"REG_c" \n\t"
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
724 "movl %%ecx, (%%"REG_a") \n\t"
4037
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
725 "4: \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
726 "add %6, %%eax \n\t"
4037
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
727 "shr $2, %%eax \n\t"
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
728
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
729 "movl %%esi, "RANGE "(%3) \n\t"
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
730 "movl %%ebx, "LOW "(%3) \n\t"
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
731 :"=&a"(coeff_count), "+r"(significant_coeff_ctx_base), "+m"(index)\
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
732 :"r"(c), "m"(minusstart), "m"(end), "m"(minusindex)\
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
733 : "%"REG_c, "%ebx", "%edx", "%esi", "memory"\
4037
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
734 );
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
735 return coeff_count;
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
736 }
4051
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
737
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
738 static int decode_significance_8x8_x86(CABACContext *c, uint8_t *significant_coeff_ctx_base, int *index, uint8_t *sig_off){
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
739 int minusindex= 4-(int)index;
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
740 int coeff_count;
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
741 long last=0;
4051
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
742 asm volatile(
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
743 "movl "RANGE "(%3), %%esi \n\t"
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
744 "movl "LOW "(%3), %%ebx \n\t"
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
745
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
746 "mov %1, %%"REG_D" \n\t"
4051
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
747 "2: \n\t"
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
748
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
749 "mov %6, %%"REG_a" \n\t"
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
750 "movzbl (%%"REG_a", %%"REG_D"), %%edi \n\t"
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
751 "add %5, %%"REG_D" \n\t"
4051
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
752
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
753 BRANCHLESS_GET_CABAC("%%edx", "%3", "(%%"REG_D")", "%%ebx", "%%bx", "%%esi", "%%eax", "%%al")
4051
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
754
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
755 "mov %1, %%edi \n\t"
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
756 "test $1, %%edx \n\t"
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
757 " jz 3f \n\t"
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
758
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
759 "movzbl "MANGLE(last_coeff_flag_offset_8x8)"(%%edi), %%edi\n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
760 "add %5, %%"REG_D" \n\t"
4051
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
761
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
762 BRANCHLESS_GET_CABAC("%%edx", "%3", "15(%%"REG_D")", "%%ebx", "%%bx", "%%esi", "%%eax", "%%al")
4051
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
763
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
764 "mov %2, %%"REG_a" \n\t"
4051
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
765 "mov %1, %%edi \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
766 "movl %%edi, (%%"REG_a") \n\t"
4051
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
767
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
768 "test $1, %%edx \n\t"
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
769 " jnz 4f \n\t"
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
770
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
771 "add $4, %%"REG_a" \n\t"
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
772 "mov %%"REG_a", %2 \n\t"
4051
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
773
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
774 "3: \n\t"
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
775 "addl $1, %%edi \n\t"
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
776 "mov %%edi, %1 \n\t"
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
777 "cmpl $63, %%edi \n\t"
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
778 " jb 2b \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
779 "mov %2, %%"REG_a" \n\t"
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
780 "movl %%edi, (%%"REG_a") \n\t"
4051
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
781 "4: \n\t"
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
782 "addl %4, %%eax \n\t"
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
783 "shr $2, %%eax \n\t"
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
784
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
785 "movl %%esi, "RANGE "(%3) \n\t"
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
786 "movl %%ebx, "LOW "(%3) \n\t"
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
787 :"=&a"(coeff_count),"+m"(last), "+m"(index)\
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
788 :"r"(c), "m"(minusindex), "m"(significant_coeff_ctx_base), "m"(sig_off)\
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
789 : "%"REG_c, "%ebx", "%edx", "%esi", "%"REG_D, "memory"\
4051
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
790 );
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
791 return coeff_count;
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
792 }
4882
8131ccb4ea72 Mark code parts that cannot work on AMD64 due to broken relocations as such.
diego
parents: 4881
diff changeset
793 #endif /* defined(ARCH_X86) && && defined(CONFIG_7REGS) && defined(HAVE_EBX_AVAILABLE) && !defined(BROKEN_RELOCATIONS) */
1287
michaelni
parents:
diff changeset
794
1300
e18667d1e94d FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents: 1298
diff changeset
795 /**
e18667d1e94d FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents: 1298
diff changeset
796 *
e18667d1e94d FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents: 1298
diff changeset
797 * @return the number of bytes read or 0 if no end
e18667d1e94d FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents: 1298
diff changeset
798 */
3928
987fffdf6ae7 don't try to inline cabac functions. gcc ignored the hint anyway, and forcing it would make h264 slower.
lorenm
parents: 3642
diff changeset
799 static int get_cabac_terminate(CABACContext *c){
4024
d550343b5dac shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents: 4014
diff changeset
800 c->range -= 2;
4345
88967250d718 replace a few hardcoded numbers with their correct named ones
michael
parents: 4283
diff changeset
801 if(c->low < c->range<<(CABAC_BITS+1)){
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
802 renorm_cabac_decoder_once(c);
1287
michaelni
parents:
diff changeset
803 return 0;
michaelni
parents:
diff changeset
804 }else{
1300
e18667d1e94d FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents: 1298
diff changeset
805 return c->bytestream - c->bytestream_start;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
806 }
1287
michaelni
parents:
diff changeset
807 }
michaelni
parents:
diff changeset
808
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
809 /**
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
810 * get (truncated) unnary binarization.
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
811 */
3928
987fffdf6ae7 don't try to inline cabac functions. gcc ignored the hint anyway, and forcing it would make h264 slower.
lorenm
parents: 3642
diff changeset
812 static int get_cabac_u(CABACContext *c, uint8_t * state, int max, int max_index, int truncated){
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
813 int i;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
814
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
815 for(i=0; i<max; i++){
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
816 if(get_cabac(c, state)==0)
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
817 return i;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
818
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
819 if(i< max_index) state++;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
820 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
821
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
822 return truncated ? max : -1;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
823 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
824
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
825 /**
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
826 * get unary exp golomb k-th order binarization.
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
827 */
3928
987fffdf6ae7 don't try to inline cabac functions. gcc ignored the hint anyway, and forcing it would make h264 slower.
lorenm
parents: 3642
diff changeset
828 static int get_cabac_ueg(CABACContext *c, uint8_t * state, int max, int is_signed, int k, int max_index){
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
829 int i, v;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
830 int m= 1<<k;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
831
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
832 if(get_cabac(c, state)==0)
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
833 return 0;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
834
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
835 if(0 < max_index) state++;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
836
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
837 for(i=1; i<max; i++){
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
838 if(get_cabac(c, state)==0){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
839 if(is_signed && get_cabac_bypass(c)){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
840 return -i;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
841 }else
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
842 return i;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
843 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
844
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
845 if(i < max_index) state++;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
846 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
847
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
848 while(get_cabac_bypass(c)){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
849 i+= m;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
850 m+= m;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
851 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
852
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
853 v=0;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
854 while(m>>=1){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
855 v+= v + get_cabac_bypass(c);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
856 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
857 i += v;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
858
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
859 if(is_signed && get_cabac_bypass(c)){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
860 return -i;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
861 }else
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
862 return i;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
863 }
4975
9a6a0818e93f split h264.c to move parser in its own file
aurel
parents: 4908
diff changeset
864
9a6a0818e93f split h264.c to move parser in its own file
aurel
parents: 4908
diff changeset
865 #endif /* CABAC_H */