annotate cabac.h @ 4064:d3afe4505951 libavcodec

CABAC assembler optimizations ported to AMD64
author reimar
date Mon, 23 Oct 2006 16:54:21 +0000
parents 19f07b651d79
children 8b1819f82c4a
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
michaelni
parents:
diff changeset
28
3284
a224d9752912 don't force asserts in release builds. 2% faster h264.
lorenm
parents: 3036
diff changeset
29 //#undef NDEBUG
1287
michaelni
parents:
diff changeset
30 #include <assert.h>
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
31 #ifdef ARCH_X86_64
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
32 #define ARCH_X86
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
33 #endif
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
34 #ifdef ARCH_X86
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
35 #include "x86_cpu.h"
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
36 #endif
1287
michaelni
parents:
diff changeset
37
3948
3edbf131ee44 refill cabac variables in 16bit steps, 3% faster get_cabac()
michael
parents: 3947
diff changeset
38 #define CABAC_BITS 16
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
39 #define CABAC_MASK ((1<<CABAC_BITS)-1)
3984
bb186452e7da BRANCHLESS_CABAD --> BRANCHLESS_CABAC_DECODER
diego
parents: 3982
diff changeset
40 #define BRANCHLESS_CABAC_DECODER 1
4039
866a83726985 Kill a warning with MSVC
gpoirier
parents: 4037
diff changeset
41 //#define ARCH_X86_DISABLED 1
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
42
1287
michaelni
parents:
diff changeset
43 typedef struct CABACContext{
michaelni
parents:
diff changeset
44 int low;
michaelni
parents:
diff changeset
45 int range;
michaelni
parents:
diff changeset
46 int outstanding_count;
michaelni
parents:
diff changeset
47 #ifdef STRICT_LIMITS
michaelni
parents:
diff changeset
48 int symCount;
michaelni
parents:
diff changeset
49 #endif
2024
f65d87bfdd5a some of the warning fixes by (Michael Roitzsch <mroi at users dot sourceforge dot net>)
michael
parents: 1787
diff changeset
50 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
51 const uint8_t *bytestream;
2116
48d9f86fb047 overread fix
michael
parents: 2024
diff changeset
52 const uint8_t *bytestream_end;
1287
michaelni
parents:
diff changeset
53 PutBitContext pb;
michaelni
parents:
diff changeset
54 }CABACContext;
michaelni
parents:
diff changeset
55
4014
b2582438effe dehack *ps_state indexing in the branchless decoder
michael
parents: 4012
diff changeset
56 extern uint8_t ff_h264_mlps_state[4*64];
4039
866a83726985 Kill a warning with MSVC
gpoirier
parents: 4037
diff changeset
57 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
58 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
59 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
60 extern const uint8_t ff_h264_norm_shift[512];
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
61
1287
michaelni
parents:
diff changeset
62
michaelni
parents:
diff changeset
63 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
64 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
65 void ff_init_cabac_states(CABACContext *c);
1287
michaelni
parents:
diff changeset
66
michaelni
parents:
diff changeset
67
michaelni
parents:
diff changeset
68 static inline void put_cabac_bit(CABACContext *c, int b){
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
69 put_bits(&c->pb, 1, b);
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
70 for(;c->outstanding_count; c->outstanding_count--){
1287
michaelni
parents:
diff changeset
71 put_bits(&c->pb, 1, 1-b);
michaelni
parents:
diff changeset
72 }
michaelni
parents:
diff changeset
73 }
michaelni
parents:
diff changeset
74
michaelni
parents:
diff changeset
75 static inline void renorm_cabac_encoder(CABACContext *c){
michaelni
parents:
diff changeset
76 while(c->range < 0x100){
michaelni
parents:
diff changeset
77 //FIXME optimize
michaelni
parents:
diff changeset
78 if(c->low<0x100){
michaelni
parents:
diff changeset
79 put_cabac_bit(c, 0);
michaelni
parents:
diff changeset
80 }else if(c->low<0x200){
michaelni
parents:
diff changeset
81 c->outstanding_count++;
michaelni
parents:
diff changeset
82 c->low -= 0x100;
michaelni
parents:
diff changeset
83 }else{
michaelni
parents:
diff changeset
84 put_cabac_bit(c, 1);
michaelni
parents:
diff changeset
85 c->low -= 0x200;
michaelni
parents:
diff changeset
86 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
87
1287
michaelni
parents:
diff changeset
88 c->range+= c->range;
michaelni
parents:
diff changeset
89 c->low += c->low;
michaelni
parents:
diff changeset
90 }
michaelni
parents:
diff changeset
91 }
michaelni
parents:
diff changeset
92
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
93 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
94 int RangeLPS= ff_h264_lps_range[2*(c->range&0xC0) + *state];
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
95
1287
michaelni
parents:
diff changeset
96 if(bit == ((*state)&1)){
michaelni
parents:
diff changeset
97 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
98 *state= ff_h264_mps_state[*state];
1287
michaelni
parents:
diff changeset
99 }else{
michaelni
parents:
diff changeset
100 c->low += c->range - RangeLPS;
michaelni
parents:
diff changeset
101 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
102 *state= ff_h264_lps_state[*state];
1287
michaelni
parents:
diff changeset
103 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
104
1287
michaelni
parents:
diff changeset
105 renorm_cabac_encoder(c);
michaelni
parents:
diff changeset
106
michaelni
parents:
diff changeset
107 #ifdef STRICT_LIMITS
michaelni
parents:
diff changeset
108 c->symCount++;
michaelni
parents:
diff changeset
109 #endif
michaelni
parents:
diff changeset
110 }
michaelni
parents:
diff changeset
111
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
112 static void put_cabac_static(CABACContext *c, int RangeLPS, int bit){
1287
michaelni
parents:
diff changeset
113 assert(c->range > RangeLPS);
michaelni
parents:
diff changeset
114
michaelni
parents:
diff changeset
115 if(!bit){
michaelni
parents:
diff changeset
116 c->range -= RangeLPS;
michaelni
parents:
diff changeset
117 }else{
michaelni
parents:
diff changeset
118 c->low += c->range - RangeLPS;
michaelni
parents:
diff changeset
119 c->range = RangeLPS;
michaelni
parents:
diff changeset
120 }
michaelni
parents:
diff changeset
121
michaelni
parents:
diff changeset
122 renorm_cabac_encoder(c);
michaelni
parents:
diff changeset
123
michaelni
parents:
diff changeset
124 #ifdef STRICT_LIMITS
michaelni
parents:
diff changeset
125 c->symCount++;
michaelni
parents:
diff changeset
126 #endif
michaelni
parents:
diff changeset
127 }
michaelni
parents:
diff changeset
128
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
129 /**
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
130 * @param bit 0 -> write zero bit, !=0 write one bit
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
131 */
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
132 static void put_cabac_bypass(CABACContext *c, int bit){
1287
michaelni
parents:
diff changeset
133 c->low += c->low;
michaelni
parents:
diff changeset
134
michaelni
parents:
diff changeset
135 if(bit){
michaelni
parents:
diff changeset
136 c->low += c->range;
michaelni
parents:
diff changeset
137 }
michaelni
parents:
diff changeset
138 //FIXME optimize
michaelni
parents:
diff changeset
139 if(c->low<0x200){
michaelni
parents:
diff changeset
140 put_cabac_bit(c, 0);
michaelni
parents:
diff changeset
141 }else if(c->low<0x400){
michaelni
parents:
diff changeset
142 c->outstanding_count++;
michaelni
parents:
diff changeset
143 c->low -= 0x200;
michaelni
parents:
diff changeset
144 }else{
michaelni
parents:
diff changeset
145 put_cabac_bit(c, 1);
michaelni
parents:
diff changeset
146 c->low -= 0x400;
michaelni
parents:
diff changeset
147 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
148
1287
michaelni
parents:
diff changeset
149 #ifdef STRICT_LIMITS
michaelni
parents:
diff changeset
150 c->symCount++;
michaelni
parents:
diff changeset
151 #endif
michaelni
parents:
diff changeset
152 }
michaelni
parents:
diff changeset
153
1300
e18667d1e94d FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents: 1298
diff changeset
154 /**
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 * @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
157 */
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
158 static int put_cabac_terminate(CABACContext *c, int bit){
1287
michaelni
parents:
diff changeset
159 c->range -= 2;
michaelni
parents:
diff changeset
160
michaelni
parents:
diff changeset
161 if(!bit){
michaelni
parents:
diff changeset
162 renorm_cabac_encoder(c);
michaelni
parents:
diff changeset
163 }else{
michaelni
parents:
diff changeset
164 c->low += c->range;
michaelni
parents:
diff changeset
165 c->range= 2;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
166
1287
michaelni
parents:
diff changeset
167 renorm_cabac_encoder(c);
michaelni
parents:
diff changeset
168
michaelni
parents:
diff changeset
169 assert(c->low <= 0x1FF);
michaelni
parents:
diff changeset
170 put_cabac_bit(c, c->low>>9);
michaelni
parents:
diff changeset
171 put_bits(&c->pb, 2, ((c->low>>7)&3)|1);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
172
1287
michaelni
parents:
diff changeset
173 flush_put_bits(&c->pb); //FIXME FIXME FIXME XXX wrong
michaelni
parents:
diff changeset
174 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
175
1287
michaelni
parents:
diff changeset
176 #ifdef STRICT_LIMITS
michaelni
parents:
diff changeset
177 c->symCount++;
michaelni
parents:
diff changeset
178 #endif
1300
e18667d1e94d FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents: 1298
diff changeset
179
1787
752b51a3c8ed get_bit_count -> put_bits_count
alex
parents: 1301
diff changeset
180 return (put_bits_count(&c->pb)+7)>>3;
1287
michaelni
parents:
diff changeset
181 }
michaelni
parents:
diff changeset
182
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
183 /**
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
184 * put (truncated) unary binarization.
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
185 */
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
186 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
187 int i;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
188
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
189 assert(v <= max);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
190
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
191 #if 1
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
192 for(i=0; i<v; i++){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
193 put_cabac(c, state, 1);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
194 if(i < max_index) state++;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
195 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
196 if(truncated==0 || v<max)
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
197 put_cabac(c, state, 0);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
198 #else
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
199 if(v <= max_index){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
200 for(i=0; i<v; i++){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
201 put_cabac(c, state+i, 1);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
202 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
203 if(truncated==0 || v<max)
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
204 put_cabac(c, state+i, 0);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
205 }else{
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
206 for(i=0; i<=max_index; i++){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
207 put_cabac(c, state+i, 1);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
208 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
209 for(; i<v; i++){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
210 put_cabac(c, state+max_index, 1);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
211 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
212 if(truncated==0 || v<max)
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
213 put_cabac(c, state+max_index, 0);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
214 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
215 #endif
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
216 }
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 * put unary exp golomb k-th order binarization.
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
220 */
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
221 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
222 int i;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
223
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
224 if(v==0)
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
225 put_cabac(c, state, 0);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
226 else{
1298
5bc3184810dc cleanup
michaelni
parents: 1290
diff changeset
227 const int sign= v < 0;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
228
4001
34fdffe98bd0 Rename ABS macro to FFABS.
diego
parents: 3999
diff changeset
229 if(is_signed) v= FFABS(v);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
230
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
231 if(v<max){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
232 for(i=0; i<v; i++){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
233 put_cabac(c, state, 1);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
234 if(i < max_index) state++;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
235 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
236
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
237 put_cabac(c, state, 0);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
238 }else{
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
239 int m= 1<<k;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
240
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
241 for(i=0; i<max; i++){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
242 put_cabac(c, state, 1);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
243 if(i < max_index) state++;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
244 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
245
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
246 v -= max;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
247 while(v >= m){ //FIXME optimize
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
248 put_cabac_bypass(c, 1);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
249 v-= m;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
250 m+= m;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
251 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
252 put_cabac_bypass(c, 0);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
253 while(m>>=1){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
254 put_cabac_bypass(c, v&m);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
255 }
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 if(is_signed)
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
259 put_cabac_bypass(c, sign);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
260 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
261 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
262
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
263 static void refill(CABACContext *c){
1c39d9786efd optimization
michael
parents: 2116
diff changeset
264 #if CABAC_BITS == 16
3946
michael
parents: 3943
diff changeset
265 c->low+= (c->bytestream[0]<<9) + (c->bytestream[1]<<1);
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
266 #else
1c39d9786efd optimization
michael
parents: 2116
diff changeset
267 c->low+= c->bytestream[0]<<1;
1c39d9786efd optimization
michael
parents: 2116
diff changeset
268 #endif
1c39d9786efd optimization
michael
parents: 2116
diff changeset
269 c->low -= CABAC_MASK;
1c39d9786efd optimization
michael
parents: 2116
diff changeset
270 c->bytestream+= CABAC_BITS/8;
1c39d9786efd optimization
michael
parents: 2116
diff changeset
271 }
1c39d9786efd optimization
michael
parents: 2116
diff changeset
272
1c39d9786efd optimization
michael
parents: 2116
diff changeset
273 static void refill2(CABACContext *c){
1c39d9786efd optimization
michael
parents: 2116
diff changeset
274 int i, x;
1c39d9786efd optimization
michael
parents: 2116
diff changeset
275
1c39d9786efd optimization
michael
parents: 2116
diff changeset
276 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
277 i= 7 - ff_h264_norm_shift[x>>(CABAC_BITS-1)];
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
278
1c39d9786efd optimization
michael
parents: 2116
diff changeset
279 x= -CABAC_MASK;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
280
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
281 #if CABAC_BITS == 16
1c39d9786efd optimization
michael
parents: 2116
diff changeset
282 x+= (c->bytestream[0]<<9) + (c->bytestream[1]<<1);
1c39d9786efd optimization
michael
parents: 2116
diff changeset
283 #else
1c39d9786efd optimization
michael
parents: 2116
diff changeset
284 x+= c->bytestream[0]<<1;
1c39d9786efd optimization
michael
parents: 2116
diff changeset
285 #endif
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
286
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
287 c->low += x<<i;
1c39d9786efd optimization
michael
parents: 2116
diff changeset
288 c->bytestream+= CABAC_BITS/8;
1c39d9786efd optimization
michael
parents: 2116
diff changeset
289 }
1c39d9786efd optimization
michael
parents: 2116
diff changeset
290
1287
michaelni
parents:
diff changeset
291 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
292 while(c->range < 0x100){
1287
michaelni
parents:
diff changeset
293 c->range+= c->range;
michaelni
parents:
diff changeset
294 c->low+= c->low;
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
295 if(!(c->low & CABAC_MASK))
1c39d9786efd optimization
michael
parents: 2116
diff changeset
296 refill(c);
1287
michaelni
parents:
diff changeset
297 }
michaelni
parents:
diff changeset
298 }
michaelni
parents:
diff changeset
299
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
300 static inline void renorm_cabac_decoder_once(CABACContext *c){
3951
751bfc30df72 disable benchmarking code
michael
parents: 3950
diff changeset
301 #ifdef ARCH_X86_DISABLED
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
302 int temp;
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
303 #if 0
3950
900d21b85dd6 renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents: 3948
diff changeset
304 //P3:683 athlon:475
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
305 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
306 "lea -0x100(%0), %2 \n\t"
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
307 "shr $31, %2 \n\t" //FIXME 31->63 for x86-64
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
308 "shl %%cl, %0 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
309 "shl %%cl, %1 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
310 : "+r"(c->range), "+r"(c->low), "+c"(temp)
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
311 );
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
312 #elif 0
3950
900d21b85dd6 renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents: 3948
diff changeset
313 //P3:680 athlon:474
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
314 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
315 "cmp $0x100, %0 \n\t"
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
316 "setb %%cl \n\t" //FIXME 31->63 for x86-64
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
317 "shl %%cl, %0 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
318 "shl %%cl, %1 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
319 : "+r"(c->range), "+r"(c->low), "+c"(temp)
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
320 );
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
321 #elif 1
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
322 int temp2;
3950
900d21b85dd6 renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents: 3948
diff changeset
323 //P3:665 athlon:517
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
324 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
325 "lea -0x100(%0), %%eax \n\t"
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
326 "cdq \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
327 "mov %0, %%eax \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
328 "and %%edx, %0 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
329 "and %1, %%edx \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
330 "add %%eax, %0 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
331 "add %%edx, %1 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
332 : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2)
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
333 );
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
334 #elif 0
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
335 int temp2;
3950
900d21b85dd6 renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents: 3948
diff changeset
336 //P3:673 athlon:509
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
337 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
338 "cmp $0x100, %0 \n\t"
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
339 "sbb %%edx, %%edx \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
340 "mov %0, %%eax \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
341 "and %%edx, %0 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
342 "and %1, %%edx \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
343 "add %%eax, %0 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
344 "add %%edx, %1 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
345 : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2)
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
346 );
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
347 #else
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
348 int temp2;
3950
900d21b85dd6 renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents: 3948
diff changeset
349 //P3:677 athlon:511
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
350 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
351 "cmp $0x100, %0 \n\t"
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
352 "lea (%0, %0), %%eax \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
353 "lea (%1, %1), %%edx \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
354 "cmovb %%eax, %0 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
355 "cmovb %%edx, %1 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
356 : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2)
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
357 );
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
358 #endif
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
359 #else
3950
900d21b85dd6 renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents: 3948
diff changeset
360 //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
361 int shift= (uint32_t)(c->range - 0x100)>>31;
3642
0efda682253c tweak cabac. 0.5% faster h264.
lorenm
parents: 3284
diff changeset
362 c->range<<= shift;
0efda682253c tweak cabac. 0.5% faster h264.
lorenm
parents: 3284
diff changeset
363 c->low <<= shift;
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
364 #endif
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
365 if(!(c->low & CABAC_MASK))
1c39d9786efd optimization
michael
parents: 2116
diff changeset
366 refill(c);
1c39d9786efd optimization
michael
parents: 2116
diff changeset
367 }
1c39d9786efd optimization
michael
parents: 2116
diff changeset
368
4008
b636f3d59283 prevent "mb level" get_cabac() calls from being inlined (3% faster decode_mb_cabac() on P3)
michael
parents: 4002
diff changeset
369 static int always_inline get_cabac_inline(CABACContext *c, uint8_t * const state){
3642
0efda682253c tweak cabac. 0.5% faster h264.
lorenm
parents: 3284
diff changeset
370 //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
371 #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
372 #define RANGE "4"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
373 #ifdef ARCH_X86_64
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
374 #define BYTESTART "16"
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
375 #define BYTE "24"
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
376 #define BYTEEND "32"
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
377 #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
378 #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
379 #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
380 #define BYTEEND "20"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
381 #endif
4044
5ccdefd60f61 Fix PIC compilation, some defines were under #ifdef !PIC but used
diego
parents: 4043
diff changeset
382 #if defined(ARCH_X86) && !(defined(PIC) && defined(__GNUC__))
5ccdefd60f61 Fix PIC compilation, some defines were under #ifdef !PIC but used
diego
parents: 4043
diff changeset
383 int bit;
5ccdefd60f61 Fix PIC compilation, some defines were under #ifdef !PIC but used
diego
parents: 4043
diff changeset
384
3984
bb186452e7da BRANCHLESS_CABAD --> BRANCHLESS_CABAC_DECODER
diego
parents: 3982
diff changeset
385 #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
386 asm volatile(
4035
b7f31a32bb30 cosmetic (%%eax->%0)
michael
parents: 4034
diff changeset
387 "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
388 "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
389 "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
390 "andl $0xC0, %%ebx \n\t"
4035
b7f31a32bb30 cosmetic (%%eax->%0)
michael
parents: 4034
diff changeset
391 "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
392 "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
393 //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
394 "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
395 "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
396 "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
397 "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
398 " ja 1f \n\t"
3999
6cbad3675632 slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents: 3996
diff changeset
399
6cbad3675632 slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents: 3996
diff changeset
400 #if 1
6cbad3675632 slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents: 3996
diff changeset
401 //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
402 "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
403 "shr $31, %%ecx \n\t"
6cbad3675632 slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents: 3996
diff changeset
404 "shl %%cl, %%edx \n\t"
6cbad3675632 slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents: 3996
diff changeset
405 "shl %%cl, %%ebx \n\t"
6cbad3675632 slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents: 3996
diff changeset
406 #else
6cbad3675632 slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents: 3996
diff changeset
407 //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
408 "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
409 "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
410 "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
411 "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
412 #endif
4035
b7f31a32bb30 cosmetic (%%eax->%0)
michael
parents: 4034
diff changeset
413 "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
414 "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
415 //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
416 "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
417 " jnz 2f \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
418 "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
419 "subl $0xFFFF, %%ebx \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
420 "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
421 "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
422 "shrl $15, %%ecx \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
423 "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
424 "addl %%ecx, %%ebx \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
425 "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
426 "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
427 "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
428 //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
429 "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
430 "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
431 "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
432 "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
433 "shll %%cl, %%edx \n\t"
4035
b7f31a32bb30 cosmetic (%%eax->%0)
michael
parents: 4034
diff changeset
434 "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
435 "movb %%cl, (%1) \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
436 "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
437 "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
438 " 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
439
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
440 "mov "BYTE "(%2), %%"REG_c" \n\t"
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
441 "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
442 "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
443 "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
444 "subl $0xFFFF, %%esi \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
445 "add $2, %%"REG_c" \n\t"
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
446 "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
447
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 "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
449 "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
450 "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
451 "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
452 "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
453 "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
454
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 "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
456 "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
457 "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
458 "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
459 "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
460 :"=&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
461 :"r"(state), "r"(c)
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
462 : "%"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
463 );
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
464 bit&=1;
4002
ec426fa57dfe adds some useful comments after some of the #else, #elseif,
gpoirier
parents: 4001
diff changeset
465 #else /* BRANCHLESS_CABAC_DECODER */
4046
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
466
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
467
4050
2c79a8281cb6 Protect code that uses CMOV instructions with HAVE_CMOV,
gpoirier
parents: 4049
diff changeset
468 #if defined CMOV_IS_FAST
4046
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
469 #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
470 "mov "tmp" , %%ecx \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
471 "shl $17 , "tmp" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
472 "cmp "low" , "tmp" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
473 "cmova %%ecx , "range" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
474 "sbb %%ecx , %%ecx \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
475 "and %%ecx , "tmp" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
476 "sub "tmp" , "low" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
477 "xor %%ecx , "ret" \n\t"
4002
ec426fa57dfe adds some useful comments after some of the #else, #elseif,
gpoirier
parents: 4001
diff changeset
478 #else /* CMOV_IS_FAST */
4046
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
479 #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
480 "mov "tmp" , %%ecx \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
481 "shl $17 , "tmp" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
482 "sub "low" , "tmp" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
483 "sar $31 , "tmp" \n\t" /*lps_mask*/\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
484 "sub %%ecx , "range" \n\t" /*RangeLPS - range*/\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
485 "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
486 "add %%ecx , "range" \n\t" /*new range*/\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
487 "shl $17 , %%ecx \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
488 "and "tmp" , %%ecx \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
489 "sub %%ecx , "low" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
490 "xor "tmp" , "ret" \n\t"
4002
ec426fa57dfe adds some useful comments after some of the #else, #elseif,
gpoirier
parents: 4001
diff changeset
491 #endif /* CMOV_IS_FAST */
3975
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
492
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
493
4046
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
494 #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
495 "movzbl "statep" , "ret" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
496 "mov "range" , "tmp" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
497 "and $0xC0 , "range" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
498 "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
499 "sub "range" , "tmp" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
500 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
501 "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
502 "shl %%cl , "range" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
503 "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
504 "mov "tmpbyte" , "statep" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
505 "shl %%cl , "low" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
506 "test "lowword" , "lowword" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
507 " jnz 1f \n\t"\
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
508 "mov "BYTE"("cabac"), %%"REG_c" \n\t"\
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
509 "movzwl (%%"REG_c") , "tmp" \n\t"\
4046
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
510 "bswap "tmp" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
511 "shr $15 , "tmp" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
512 "sub $0xFFFF , "tmp" \n\t"\
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
513 "add $2 , %%"REG_c" \n\t"\
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
514 "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
515 "lea -1("low") , %%ecx \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
516 "xor "low" , %%ecx \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
517 "shr $15 , %%ecx \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
518 "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
519 "neg %%ecx \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
520 "add $7 , %%ecx \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
521 "shl %%cl , "tmp" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
522 "add "tmp" , "low" \n\t"\
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
523 "1: \n\t"
3975
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
524
4046
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
525 asm volatile(
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
526 "movl "RANGE "(%2), %%esi \n\t"
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
527 "movl "LOW "(%2), %%ebx \n\t"
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
528 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
529 "movl %%esi, "RANGE "(%2) \n\t"
8bbc695c9603 factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents: 4044
diff changeset
530 "movl %%ebx, "LOW "(%2) \n\t"
3975
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
531
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
532 :"=&a"(bit)
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
533 :"r"(state), "r"(c)
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
534 : "%"REG_c, "%ebx", "%edx", "%esi", "memory"
3975
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
535 );
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
536 bit&=1;
4002
ec426fa57dfe adds some useful comments after some of the #else, #elseif,
gpoirier
parents: 4001
diff changeset
537 #endif /* BRANCHLESS_CABAC_DECODER */
4033
f7a6b2bb3a2f Expand some #endif comments.
diego
parents: 4027
diff changeset
538 #else /* defined(ARCH_X86) && !(defined(PIC) && defined(__GNUC__)) */
3642
0efda682253c tweak cabac. 0.5% faster h264.
lorenm
parents: 3284
diff changeset
539 int s = *state;
4039
866a83726985 Kill a warning with MSVC
gpoirier
parents: 4037
diff changeset
540 int RangeLPS= ff_h264_lps_range[2*(c->range&0xC0) + s];
2522
e25782262d7d kill warnings patch by (M«©ns Rullg«©rd <mru inprovide com>)
michael
parents: 2323
diff changeset
541 int bit, lps_mask attribute_unused;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
542
1287
michaelni
parents:
diff changeset
543 c->range -= RangeLPS;
3984
bb186452e7da BRANCHLESS_CABAD --> BRANCHLESS_CABAC_DECODER
diego
parents: 3982
diff changeset
544 #ifndef BRANCHLESS_CABAC_DECODER
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
545 if(c->low < (c->range<<17)){
3642
0efda682253c tweak cabac. 0.5% faster h264.
lorenm
parents: 3284
diff changeset
546 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
547 *state= ff_h264_mps_state[s];
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
548 renorm_cabac_decoder_once(c);
1287
michaelni
parents:
diff changeset
549 }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
550 bit= ff_h264_norm_shift[RangeLPS];
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 c->low -= (c->range<<17);
3993
8b7c59b7af01 make state transition tables global as they are constant and the code is slightly faster that way
michael
parents: 3992
diff changeset
552 *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
553 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
554 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
555 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
556
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
557 if(!(c->low & 0xFFFF)){
1c39d9786efd optimization
michael
parents: 2116
diff changeset
558 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
559 }
1287
michaelni
parents:
diff changeset
560 }
4002
ec426fa57dfe adds some useful comments after some of the #else, #elseif,
gpoirier
parents: 4001
diff changeset
561 #else /* BRANCHLESS_CABAC_DECODER */
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
562 lps_mask= ((c->range<<17) - c->low)>>31;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
563
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
564 c->low -= (c->range<<17) & lps_mask;
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
565 c->range += (RangeLPS - c->range) & lps_mask;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
566
3974
93746612bc78 optimize branchless C CABAC decoder
michael
parents: 3972
diff changeset
567 s^=lps_mask;
4014
b2582438effe dehack *ps_state indexing in the branchless decoder
michael
parents: 4012
diff changeset
568 *state= (ff_h264_mlps_state+128)[s];
3974
93746612bc78 optimize branchless C CABAC decoder
michael
parents: 3972
diff changeset
569 bit= s&1;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
570
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
571 lps_mask= ff_h264_norm_shift[c->range];
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
572 c->range<<= lps_mask;
1c39d9786efd optimization
michael
parents: 2116
diff changeset
573 c->low <<= lps_mask;
1c39d9786efd optimization
michael
parents: 2116
diff changeset
574 if(!(c->low & CABAC_MASK))
1c39d9786efd optimization
michael
parents: 2116
diff changeset
575 refill2(c);
4002
ec426fa57dfe adds some useful comments after some of the #else, #elseif,
gpoirier
parents: 4001
diff changeset
576 #endif /* BRANCHLESS_CABAC_DECODER */
4033
f7a6b2bb3a2f Expand some #endif comments.
diego
parents: 4027
diff changeset
577 #endif /* defined(ARCH_X86) && !(defined(PIC) && defined(__GNUC__)) */
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
578 return bit;
1287
michaelni
parents:
diff changeset
579 }
michaelni
parents:
diff changeset
580
4008
b636f3d59283 prevent "mb level" get_cabac() calls from being inlined (3% faster decode_mb_cabac() on P3)
michael
parents: 4002
diff changeset
581 static int __attribute((noinline)) get_cabac_noinline(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
582 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
583 }
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 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
586 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
587 }
b636f3d59283 prevent "mb level" get_cabac() calls from being inlined (3% faster decode_mb_cabac() on P3)
michael
parents: 4002
diff changeset
588
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
589 static int get_cabac_bypass(CABACContext *c){
4040
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
590 #if 0 //not faster
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
591 int bit;
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
592 asm volatile(
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
593 "movl "RANGE "(%1), %%ebx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
594 "movl "LOW "(%1), %%eax \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
595 "shl $17, %%ebx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
596 "add %%eax, %%eax \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
597 "sub %%ebx, %%eax \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
598 "cdq \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
599 "and %%edx, %%ebx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
600 "add %%ebx, %%eax \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
601 "test %%ax, %%ax \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
602 " jnz 1f \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
603 "movl "BYTE "(%1), %%"REG_b" \n\t"
4040
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
604 "subl $0xFFFF, %%eax \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
605 "movzwl (%%"REG_b"), %%ecx \n\t"
4040
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
606 "bswap %%ecx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
607 "shrl $15, %%ecx \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
608 "addl $2, %%"REG_b" \n\t"
4040
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
609 "addl %%ecx, %%eax \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
610 "movl %%"REG_b", "BYTE "(%1) \n\t"
4040
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
611 "1: \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
612 "movl %%eax, "LOW "(%1) \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
613
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
614 :"=&d"(bit)
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
615 :"r"(c)
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
616 : "%eax", "%"REG_b, "%ecx", "memory"
4040
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
617 );
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
618 return bit+1;
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
619 #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
620 int range;
1287
michaelni
parents:
diff changeset
621 c->low += c->low;
michaelni
parents:
diff changeset
622
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
623 if(!(c->low & CABAC_MASK))
1c39d9786efd optimization
michael
parents: 2116
diff changeset
624 refill(c);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
625
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
626 range= c->range<<17;
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
627 if(c->low < range){
1287
michaelni
parents:
diff changeset
628 return 0;
michaelni
parents:
diff changeset
629 }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
630 c->low -= range;
1287
michaelni
parents:
diff changeset
631 return 1;
michaelni
parents:
diff changeset
632 }
4040
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
633 #endif
1287
michaelni
parents:
diff changeset
634 }
4040
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
635
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 static always_inline int get_cabac_bypass_sign(CABACContext *c, int val){
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
638 #ifdef ARCH_X86
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
639 asm volatile(
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
640 "movl "RANGE "(%1), %%ebx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
641 "movl "LOW "(%1), %%eax \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
642 "shl $17, %%ebx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
643 "add %%eax, %%eax \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
644 "sub %%ebx, %%eax \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
645 "cdq \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
646 "and %%edx, %%ebx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
647 "add %%ebx, %%eax \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
648 "xor %%edx, %%ecx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
649 "sub %%edx, %%ecx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
650 "test %%ax, %%ax \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
651 " jnz 1f \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
652 "mov "BYTE "(%1), %%"REG_b" \n\t"
4040
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
653 "subl $0xFFFF, %%eax \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
654 "movzwl (%%"REG_b"), %%edx \n\t"
4040
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
655 "bswap %%edx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
656 "shrl $15, %%edx \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
657 "add $2, %%"REG_b" \n\t"
4040
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
658 "addl %%edx, %%eax \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
659 "mov %%"REG_b", "BYTE "(%1) \n\t"
4040
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
660 "1: \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
661 "movl %%eax, "LOW "(%1) \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
662
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
663 :"+c"(val)
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
664 :"r"(c)
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
665 : "%eax", "%"REG_b, "%edx", "memory"
4040
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
666 );
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
667 return val;
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
668 #else
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
669 int range, mask;
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
670 c->low += c->low;
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
671
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
672 if(!(c->low & CABAC_MASK))
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
673 refill(c);
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
674
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
675 range= c->range<<17;
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
676 c->low -= range;
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
677 mask= c->low >> 31;
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
678 range &= mask;
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
679 c->low += range;
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
680 return (val^mask)-mask;
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
681 #endif
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
682 }
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
683
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
684 //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
685 //FIXME use some macros to avoid duplicatin get_cabac (cant be done yet as that would make optimization work hard)
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 #ifdef ARCH_X86
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
687 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
688 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
689 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
690 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
691 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
692 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
693 "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
694 "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
695
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 "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
697
4047
61a4e7218a45 reordering instructions a little in decode_significance_x86() -> 2 instructions less / 1% faster decode_residual on P3
michael
parents: 4046
diff changeset
698 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
699
4047
61a4e7218a45 reordering instructions a little in decode_significance_x86() -> 2 instructions less / 1% faster decode_residual on P3
michael
parents: 4046
diff changeset
700 "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
701 " 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
702
4047
61a4e7218a45 reordering instructions a little in decode_significance_x86() -> 2 instructions less / 1% faster decode_residual on P3
michael
parents: 4046
diff changeset
703 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
704
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
705 "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
706 "movl %4, %%ecx \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
707 "add %1, %%"REG_c" \n\t"
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
708 "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
709
4047
61a4e7218a45 reordering instructions a little in decode_significance_x86() -> 2 instructions less / 1% faster decode_residual on P3
michael
parents: 4046
diff changeset
710 "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
711 " 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
712
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
713 "add $4, %%"REG_a" \n\t"
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
714 "mov %%"REG_a", %2 \n\t"
4048
bf6791303fa0 1 instruction less
michael
parents: 4047
diff changeset
715
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
716 "3: \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
717 "add $1, %1 \n\t"
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
718 "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
719 " jb 2b \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
720 "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
721 "movl %4, %%ecx \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
722 "add %1, %%"REG_c" \n\t"
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
723 "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
724 "4: \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
725 "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
726 "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
727
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 "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
729 "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
730 :"=&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
731 :"r"(c), "m"(minusstart), "m"(end), "m"(minusindex)\
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
732 : "%"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
733 );
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 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
735 }
4051
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
736
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
737 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
738 int minusindex= 4-(int)index;
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
739 int coeff_count;
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
740 long last=0;
4051
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
741 asm volatile(
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
742 "movl "RANGE "(%3), %%esi \n\t"
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
743 "movl "LOW "(%3), %%ebx \n\t"
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
744
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
745 "mov %1, %%"REG_D" \n\t"
4051
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
746 "2: \n\t"
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
747
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
748 "mov %6, %%"REG_a" \n\t"
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
749 "movzbl (%%"REG_a", %%"REG_D"), %%edi \n\t"
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
750 "add %5, %%"REG_D" \n\t"
4051
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
751
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
752 BRANCHLESS_GET_CABAC("%%edx", "%3", "(%%"REG_D")", "%%ebx", "%%bx", "%%esi", "%%eax", "%%al")
4051
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
753
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
754 "mov %1, %%edi \n\t"
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
755 "test $1, %%edx \n\t"
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
756 " jz 3f \n\t"
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
757
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
758 "movzbl "MANGLE(last_coeff_flag_offset_8x8)"(%%edi), %%edi\n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
759 "add %5, %%"REG_D" \n\t"
4051
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
760
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
761 BRANCHLESS_GET_CABAC("%%edx", "%3", "15(%%"REG_D")", "%%ebx", "%%bx", "%%esi", "%%eax", "%%al")
4051
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
762
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
763 "mov %2, %%"REG_a" \n\t"
4051
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
764 "mov %1, %%edi \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
765 "movl %%edi, (%%"REG_a") \n\t"
4051
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
766
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
767 "test $1, %%edx \n\t"
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
768 " jnz 4f \n\t"
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
769
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
770 "add $4, %%"REG_a" \n\t"
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
771 "mov %%"REG_a", %2 \n\t"
4051
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
772
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
773 "3: \n\t"
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
774 "addl $1, %%edi \n\t"
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
775 "mov %%edi, %1 \n\t"
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
776 "cmpl $63, %%edi \n\t"
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
777 " jb 2b \n\t"
4064
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
778 "mov %2, %%"REG_a" \n\t"
d3afe4505951 CABAC assembler optimizations ported to AMD64
reimar
parents: 4051
diff changeset
779 "movl %%edi, (%%"REG_a") \n\t"
4051
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
780 "4: \n\t"
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
781 "addl %4, %%eax \n\t"
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
782 "shr $2, %%eax \n\t"
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
783
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
784 "movl %%esi, "RANGE "(%3) \n\t"
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
785 "movl %%ebx, "LOW "(%3) \n\t"
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
786 :"=&a"(coeff_count),"+m"(last), "+m"(index)\
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
787 :"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
788 : "%"REG_c, "%ebx", "%edx", "%esi", "%"REG_D, "memory"\
4051
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
789 );
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
790 return coeff_count;
19f07b651d79 decode_significance_8x8_x86()
michael
parents: 4050
diff changeset
791 }
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
792 #endif
1287
michaelni
parents:
diff changeset
793
1300
e18667d1e94d FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents: 1298
diff changeset
794 /**
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 * @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
797 */
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
798 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
799 c->range -= 2;
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 if(c->low < c->range<<17){
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
801 renorm_cabac_decoder_once(c);
1287
michaelni
parents:
diff changeset
802 return 0;
michaelni
parents:
diff changeset
803 }else{
1300
e18667d1e94d FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents: 1298
diff changeset
804 return c->bytestream - c->bytestream_start;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
805 }
1287
michaelni
parents:
diff changeset
806 }
michaelni
parents:
diff changeset
807
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
808 /**
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
809 * get (truncated) unnary binarization.
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
810 */
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
811 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
812 int i;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
813
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
814 for(i=0; i<max; i++){
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
815 if(get_cabac(c, state)==0)
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
816 return i;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
817
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
818 if(i< max_index) state++;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
819 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
820
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
821 return truncated ? max : -1;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
822 }
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 * get unary exp golomb k-th order binarization.
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
826 */
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
827 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
828 int i, v;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
829 int m= 1<<k;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
830
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
831 if(get_cabac(c, state)==0)
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
832 return 0;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
833
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
834 if(0 < max_index) state++;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
835
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
836 for(i=1; i<max; i++){
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
837 if(get_cabac(c, state)==0){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
838 if(is_signed && get_cabac_bypass(c)){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
839 return -i;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
840 }else
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
841 return i;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
842 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
843
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
844 if(i < max_index) state++;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
845 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
846
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
847 while(get_cabac_bypass(c)){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
848 i+= m;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
849 m+= m;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
850 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
851
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
852 v=0;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
853 while(m>>=1){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
854 v+= v + get_cabac_bypass(c);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
855 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
856 i += v;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
857
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
858 if(is_signed && get_cabac_bypass(c)){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
859 return -i;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
860 }else
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
861 return i;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
862 }