annotate cabac.h @ 4035:b7f31a32bb30 libavcodec

cosmetic (%%eax->%0)
author michael
date Tue, 17 Oct 2006 09:38:37 +0000
parents fbe263601152
children 53be304c7f54
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>
michaelni
parents:
diff changeset
31
3948
3edbf131ee44 refill cabac variables in 16bit steps, 3% faster get_cabac()
michael
parents: 3947
diff changeset
32 #define CABAC_BITS 16
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
33 #define CABAC_MASK ((1<<CABAC_BITS)-1)
3984
bb186452e7da BRANCHLESS_CABAD --> BRANCHLESS_CABAC_DECODER
diego
parents: 3982
diff changeset
34 #define BRANCHLESS_CABAC_DECODER 1
3990
746a60ba3177 enable CMOV_IS_FAST as its faster or equal speed on every cpu (duron, athlon, PM, P3) from which ive seen benchmarks, it might be slower on P4 but noone has posted benchmarks ...
michael
parents: 3984
diff changeset
35 #define CMOV_IS_FAST 1
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
36
1287
michaelni
parents:
diff changeset
37 typedef struct CABACContext{
michaelni
parents:
diff changeset
38 int low;
michaelni
parents:
diff changeset
39 int range;
michaelni
parents:
diff changeset
40 int outstanding_count;
michaelni
parents:
diff changeset
41 #ifdef STRICT_LIMITS
michaelni
parents:
diff changeset
42 int symCount;
michaelni
parents:
diff changeset
43 #endif
2024
f65d87bfdd5a some of the warning fixes by (Michael Roitzsch <mroi at users dot sourceforge dot net>)
michael
parents: 1787
diff changeset
44 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
45 const uint8_t *bytestream;
2116
48d9f86fb047 overread fix
michael
parents: 2024
diff changeset
46 const uint8_t *bytestream_end;
1287
michaelni
parents:
diff changeset
47 PutBitContext pb;
michaelni
parents:
diff changeset
48 }CABACContext;
michaelni
parents:
diff changeset
49
4014
b2582438effe dehack *ps_state indexing in the branchless decoder
michael
parents: 4012
diff changeset
50 extern uint8_t ff_h264_mlps_state[4*64];
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
51 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
52 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
53 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
54 extern const uint8_t ff_h264_norm_shift[512];
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
55
1287
michaelni
parents:
diff changeset
56
michaelni
parents:
diff changeset
57 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
58 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
59 void ff_init_cabac_states(CABACContext *c);
1287
michaelni
parents:
diff changeset
60
michaelni
parents:
diff changeset
61
michaelni
parents:
diff changeset
62 static inline void put_cabac_bit(CABACContext *c, int b){
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
63 put_bits(&c->pb, 1, b);
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
64 for(;c->outstanding_count; c->outstanding_count--){
1287
michaelni
parents:
diff changeset
65 put_bits(&c->pb, 1, 1-b);
michaelni
parents:
diff changeset
66 }
michaelni
parents:
diff changeset
67 }
michaelni
parents:
diff changeset
68
michaelni
parents:
diff changeset
69 static inline void renorm_cabac_encoder(CABACContext *c){
michaelni
parents:
diff changeset
70 while(c->range < 0x100){
michaelni
parents:
diff changeset
71 //FIXME optimize
michaelni
parents:
diff changeset
72 if(c->low<0x100){
michaelni
parents:
diff changeset
73 put_cabac_bit(c, 0);
michaelni
parents:
diff changeset
74 }else if(c->low<0x200){
michaelni
parents:
diff changeset
75 c->outstanding_count++;
michaelni
parents:
diff changeset
76 c->low -= 0x100;
michaelni
parents:
diff changeset
77 }else{
michaelni
parents:
diff changeset
78 put_cabac_bit(c, 1);
michaelni
parents:
diff changeset
79 c->low -= 0x200;
michaelni
parents:
diff changeset
80 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
81
1287
michaelni
parents:
diff changeset
82 c->range+= c->range;
michaelni
parents:
diff changeset
83 c->low += c->low;
michaelni
parents:
diff changeset
84 }
michaelni
parents:
diff changeset
85 }
michaelni
parents:
diff changeset
86
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
87 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
88 int RangeLPS= ff_h264_lps_range[2*(c->range&0xC0) + *state];
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
89
1287
michaelni
parents:
diff changeset
90 if(bit == ((*state)&1)){
michaelni
parents:
diff changeset
91 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
92 *state= ff_h264_mps_state[*state];
1287
michaelni
parents:
diff changeset
93 }else{
michaelni
parents:
diff changeset
94 c->low += c->range - RangeLPS;
michaelni
parents:
diff changeset
95 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
96 *state= ff_h264_lps_state[*state];
1287
michaelni
parents:
diff changeset
97 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
98
1287
michaelni
parents:
diff changeset
99 renorm_cabac_encoder(c);
michaelni
parents:
diff changeset
100
michaelni
parents:
diff changeset
101 #ifdef STRICT_LIMITS
michaelni
parents:
diff changeset
102 c->symCount++;
michaelni
parents:
diff changeset
103 #endif
michaelni
parents:
diff changeset
104 }
michaelni
parents:
diff changeset
105
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
106 static void put_cabac_static(CABACContext *c, int RangeLPS, int bit){
1287
michaelni
parents:
diff changeset
107 assert(c->range > RangeLPS);
michaelni
parents:
diff changeset
108
michaelni
parents:
diff changeset
109 if(!bit){
michaelni
parents:
diff changeset
110 c->range -= RangeLPS;
michaelni
parents:
diff changeset
111 }else{
michaelni
parents:
diff changeset
112 c->low += c->range - RangeLPS;
michaelni
parents:
diff changeset
113 c->range = RangeLPS;
michaelni
parents:
diff changeset
114 }
michaelni
parents:
diff changeset
115
michaelni
parents:
diff changeset
116 renorm_cabac_encoder(c);
michaelni
parents:
diff changeset
117
michaelni
parents:
diff changeset
118 #ifdef STRICT_LIMITS
michaelni
parents:
diff changeset
119 c->symCount++;
michaelni
parents:
diff changeset
120 #endif
michaelni
parents:
diff changeset
121 }
michaelni
parents:
diff changeset
122
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
123 /**
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
124 * @param bit 0 -> write zero bit, !=0 write one bit
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
125 */
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
126 static void put_cabac_bypass(CABACContext *c, int bit){
1287
michaelni
parents:
diff changeset
127 c->low += c->low;
michaelni
parents:
diff changeset
128
michaelni
parents:
diff changeset
129 if(bit){
michaelni
parents:
diff changeset
130 c->low += c->range;
michaelni
parents:
diff changeset
131 }
michaelni
parents:
diff changeset
132 //FIXME optimize
michaelni
parents:
diff changeset
133 if(c->low<0x200){
michaelni
parents:
diff changeset
134 put_cabac_bit(c, 0);
michaelni
parents:
diff changeset
135 }else if(c->low<0x400){
michaelni
parents:
diff changeset
136 c->outstanding_count++;
michaelni
parents:
diff changeset
137 c->low -= 0x200;
michaelni
parents:
diff changeset
138 }else{
michaelni
parents:
diff changeset
139 put_cabac_bit(c, 1);
michaelni
parents:
diff changeset
140 c->low -= 0x400;
michaelni
parents:
diff changeset
141 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
142
1287
michaelni
parents:
diff changeset
143 #ifdef STRICT_LIMITS
michaelni
parents:
diff changeset
144 c->symCount++;
michaelni
parents:
diff changeset
145 #endif
michaelni
parents:
diff changeset
146 }
michaelni
parents:
diff changeset
147
1300
e18667d1e94d FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents: 1298
diff changeset
148 /**
e18667d1e94d FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents: 1298
diff changeset
149 *
e18667d1e94d FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents: 1298
diff changeset
150 * @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
151 */
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
152 static int put_cabac_terminate(CABACContext *c, int bit){
1287
michaelni
parents:
diff changeset
153 c->range -= 2;
michaelni
parents:
diff changeset
154
michaelni
parents:
diff changeset
155 if(!bit){
michaelni
parents:
diff changeset
156 renorm_cabac_encoder(c);
michaelni
parents:
diff changeset
157 }else{
michaelni
parents:
diff changeset
158 c->low += c->range;
michaelni
parents:
diff changeset
159 c->range= 2;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
160
1287
michaelni
parents:
diff changeset
161 renorm_cabac_encoder(c);
michaelni
parents:
diff changeset
162
michaelni
parents:
diff changeset
163 assert(c->low <= 0x1FF);
michaelni
parents:
diff changeset
164 put_cabac_bit(c, c->low>>9);
michaelni
parents:
diff changeset
165 put_bits(&c->pb, 2, ((c->low>>7)&3)|1);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
166
1287
michaelni
parents:
diff changeset
167 flush_put_bits(&c->pb); //FIXME FIXME FIXME XXX wrong
michaelni
parents:
diff changeset
168 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
169
1287
michaelni
parents:
diff changeset
170 #ifdef STRICT_LIMITS
michaelni
parents:
diff changeset
171 c->symCount++;
michaelni
parents:
diff changeset
172 #endif
1300
e18667d1e94d FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents: 1298
diff changeset
173
1787
752b51a3c8ed get_bit_count -> put_bits_count
alex
parents: 1301
diff changeset
174 return (put_bits_count(&c->pb)+7)>>3;
1287
michaelni
parents:
diff changeset
175 }
michaelni
parents:
diff changeset
176
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
177 /**
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
178 * put (truncated) unary binarization.
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
179 */
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
180 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
181 int i;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
182
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
183 assert(v <= max);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
184
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
185 #if 1
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
186 for(i=0; i<v; i++){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
187 put_cabac(c, state, 1);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
188 if(i < max_index) state++;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
189 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
190 if(truncated==0 || v<max)
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
191 put_cabac(c, state, 0);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
192 #else
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
193 if(v <= max_index){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
194 for(i=0; i<v; i++){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
195 put_cabac(c, state+i, 1);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
196 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
197 if(truncated==0 || v<max)
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
198 put_cabac(c, state+i, 0);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
199 }else{
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
200 for(i=0; i<=max_index; 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 for(; i<v; i++){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
204 put_cabac(c, state+max_index, 1);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
205 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
206 if(truncated==0 || v<max)
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
207 put_cabac(c, state+max_index, 0);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
208 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
209 #endif
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
210 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
211
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
212 /**
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
213 * put unary exp golomb k-th order binarization.
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
214 */
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
215 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
216 int i;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
217
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
218 if(v==0)
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
219 put_cabac(c, state, 0);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
220 else{
1298
5bc3184810dc cleanup
michaelni
parents: 1290
diff changeset
221 const int sign= v < 0;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
222
4001
34fdffe98bd0 Rename ABS macro to FFABS.
diego
parents: 3999
diff changeset
223 if(is_signed) v= FFABS(v);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
224
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
225 if(v<max){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
226 for(i=0; i<v; i++){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
227 put_cabac(c, state, 1);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
228 if(i < max_index) state++;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
229 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
230
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
231 put_cabac(c, state, 0);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
232 }else{
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
233 int m= 1<<k;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
234
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
235 for(i=0; i<max; i++){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
236 put_cabac(c, state, 1);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
237 if(i < max_index) state++;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
238 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
239
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
240 v -= max;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
241 while(v >= m){ //FIXME optimize
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
242 put_cabac_bypass(c, 1);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
243 v-= m;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
244 m+= m;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
245 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
246 put_cabac_bypass(c, 0);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
247 while(m>>=1){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
248 put_cabac_bypass(c, v&m);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
249 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
250 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
251
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
252 if(is_signed)
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
253 put_cabac_bypass(c, sign);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
254 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
255 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
256
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
257 static void refill(CABACContext *c){
1c39d9786efd optimization
michael
parents: 2116
diff changeset
258 #if CABAC_BITS == 16
3946
michael
parents: 3943
diff changeset
259 c->low+= (c->bytestream[0]<<9) + (c->bytestream[1]<<1);
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
260 #else
1c39d9786efd optimization
michael
parents: 2116
diff changeset
261 c->low+= c->bytestream[0]<<1;
1c39d9786efd optimization
michael
parents: 2116
diff changeset
262 #endif
1c39d9786efd optimization
michael
parents: 2116
diff changeset
263 c->low -= CABAC_MASK;
1c39d9786efd optimization
michael
parents: 2116
diff changeset
264 c->bytestream+= CABAC_BITS/8;
1c39d9786efd optimization
michael
parents: 2116
diff changeset
265 }
1c39d9786efd optimization
michael
parents: 2116
diff changeset
266
1c39d9786efd optimization
michael
parents: 2116
diff changeset
267 static void refill2(CABACContext *c){
1c39d9786efd optimization
michael
parents: 2116
diff changeset
268 int i, x;
1c39d9786efd optimization
michael
parents: 2116
diff changeset
269
1c39d9786efd optimization
michael
parents: 2116
diff changeset
270 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
271 i= 7 - ff_h264_norm_shift[x>>(CABAC_BITS-1)];
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
272
1c39d9786efd optimization
michael
parents: 2116
diff changeset
273 x= -CABAC_MASK;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
274
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
275 #if CABAC_BITS == 16
1c39d9786efd optimization
michael
parents: 2116
diff changeset
276 x+= (c->bytestream[0]<<9) + (c->bytestream[1]<<1);
1c39d9786efd optimization
michael
parents: 2116
diff changeset
277 #else
1c39d9786efd optimization
michael
parents: 2116
diff changeset
278 x+= c->bytestream[0]<<1;
1c39d9786efd optimization
michael
parents: 2116
diff changeset
279 #endif
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
280
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
281 c->low += x<<i;
1c39d9786efd optimization
michael
parents: 2116
diff changeset
282 c->bytestream+= CABAC_BITS/8;
1c39d9786efd optimization
michael
parents: 2116
diff changeset
283 }
1c39d9786efd optimization
michael
parents: 2116
diff changeset
284
1287
michaelni
parents:
diff changeset
285 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
286 while(c->range < 0x100){
1287
michaelni
parents:
diff changeset
287 c->range+= c->range;
michaelni
parents:
diff changeset
288 c->low+= c->low;
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
289 if(!(c->low & CABAC_MASK))
1c39d9786efd optimization
michael
parents: 2116
diff changeset
290 refill(c);
1287
michaelni
parents:
diff changeset
291 }
michaelni
parents:
diff changeset
292 }
michaelni
parents:
diff changeset
293
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
294 static inline void renorm_cabac_decoder_once(CABACContext *c){
3951
751bfc30df72 disable benchmarking code
michael
parents: 3950
diff changeset
295 #ifdef ARCH_X86_DISABLED
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
296 int temp;
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
297 #if 0
3950
900d21b85dd6 renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents: 3948
diff changeset
298 //P3:683 athlon:475
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
299 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
300 "lea -0x100(%0), %2 \n\t"
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
301 "shr $31, %2 \n\t" //FIXME 31->63 for x86-64
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
302 "shl %%cl, %0 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
303 "shl %%cl, %1 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
304 : "+r"(c->range), "+r"(c->low), "+c"(temp)
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
305 );
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
306 #elif 0
3950
900d21b85dd6 renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents: 3948
diff changeset
307 //P3:680 athlon:474
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
308 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
309 "cmp $0x100, %0 \n\t"
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
310 "setb %%cl \n\t" //FIXME 31->63 for x86-64
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
311 "shl %%cl, %0 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
312 "shl %%cl, %1 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
313 : "+r"(c->range), "+r"(c->low), "+c"(temp)
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
314 );
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
315 #elif 1
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
316 int temp2;
3950
900d21b85dd6 renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents: 3948
diff changeset
317 //P3:665 athlon:517
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
318 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
319 "lea -0x100(%0), %%eax \n\t"
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
320 "cdq \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
321 "mov %0, %%eax \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
322 "and %%edx, %0 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
323 "and %1, %%edx \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
324 "add %%eax, %0 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
325 "add %%edx, %1 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
326 : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2)
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
327 );
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
328 #elif 0
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
329 int temp2;
3950
900d21b85dd6 renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents: 3948
diff changeset
330 //P3:673 athlon:509
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
331 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
332 "cmp $0x100, %0 \n\t"
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
333 "sbb %%edx, %%edx \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
334 "mov %0, %%eax \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
335 "and %%edx, %0 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
336 "and %1, %%edx \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
337 "add %%eax, %0 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
338 "add %%edx, %1 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
339 : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2)
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
340 );
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
341 #else
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
342 int temp2;
3950
900d21b85dd6 renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents: 3948
diff changeset
343 //P3:677 athlon:511
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
344 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
345 "cmp $0x100, %0 \n\t"
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
346 "lea (%0, %0), %%eax \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
347 "lea (%1, %1), %%edx \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
348 "cmovb %%eax, %0 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
349 "cmovb %%edx, %1 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
350 : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2)
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
351 );
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
352 #endif
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
353 #else
3950
900d21b85dd6 renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents: 3948
diff changeset
354 //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
355 int shift= (uint32_t)(c->range - 0x100)>>31;
3642
0efda682253c tweak cabac. 0.5% faster h264.
lorenm
parents: 3284
diff changeset
356 c->range<<= shift;
0efda682253c tweak cabac. 0.5% faster h264.
lorenm
parents: 3284
diff changeset
357 c->low <<= shift;
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
358 #endif
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
359 if(!(c->low & CABAC_MASK))
1c39d9786efd optimization
michael
parents: 2116
diff changeset
360 refill(c);
1c39d9786efd optimization
michael
parents: 2116
diff changeset
361 }
1c39d9786efd optimization
michael
parents: 2116
diff changeset
362
4008
b636f3d59283 prevent "mb level" get_cabac() calls from being inlined (3% faster decode_mb_cabac() on P3)
michael
parents: 4002
diff changeset
363 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
364 //FIXME gcc generates duplicate load/stores for c->low and c->range
4026
michael
parents: 4025
diff changeset
365 #if defined(ARCH_X86) && !(defined(PIC) && defined(__GNUC__))
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
366 int bit;
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
367
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
368 #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
369 #define RANGE "4"
3993
8b7c59b7af01 make state transition tables global as they are constant and the code is slightly faster that way
michael
parents: 3992
diff changeset
370 #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
371 #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
372 #define BYTEEND "20"
3984
bb186452e7da BRANCHLESS_CABAD --> BRANCHLESS_CABAC_DECODER
diego
parents: 3982
diff changeset
373 #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
374 asm volatile(
4035
b7f31a32bb30 cosmetic (%%eax->%0)
michael
parents: 4034
diff changeset
375 "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
376 "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
377 "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
378 "andl $0xC0, %%ebx \n\t"
4035
b7f31a32bb30 cosmetic (%%eax->%0)
michael
parents: 4034
diff changeset
379 "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
380 "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
381 //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
382 "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
383 "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
384 "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
385 "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
386 " ja 1f \n\t"
3999
6cbad3675632 slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents: 3996
diff changeset
387
6cbad3675632 slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents: 3996
diff changeset
388 #if 1
6cbad3675632 slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents: 3996
diff changeset
389 //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
390 "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
391 "shr $31, %%ecx \n\t"
6cbad3675632 slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents: 3996
diff changeset
392 "shl %%cl, %%edx \n\t"
6cbad3675632 slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents: 3996
diff changeset
393 "shl %%cl, %%ebx \n\t"
6cbad3675632 slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents: 3996
diff changeset
394 #else
6cbad3675632 slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents: 3996
diff changeset
395 //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
396 "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
397 "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
398 "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
399 "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
400 #endif
4035
b7f31a32bb30 cosmetic (%%eax->%0)
michael
parents: 4034
diff changeset
401 "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
402 "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
403 //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
404 "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
405 " 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
406 "movl "BYTE "(%2), %%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
407 "subl $0xFFFF, %%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
408 "movzwl (%%esi), %%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
409 "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
410 "shrl $15, %%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
411 "addl $2, %%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
412 "addl %%ecx, %%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
413 "movl %%esi, "BYTE "(%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
414 "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
415 "1: \n\t"
fc6e0942353b first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents: 3967
diff changeset
416 //eax:state ebx:low, edx:range, esi:RangeLPS
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
417 "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
418 "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
419 "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
420 "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
421 "shll %%cl, %%edx \n\t"
4035
b7f31a32bb30 cosmetic (%%eax->%0)
michael
parents: 4034
diff changeset
422 "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
423 "movb %%cl, (%1) \n\t"
4035
b7f31a32bb30 cosmetic (%%eax->%0)
michael
parents: 4034
diff changeset
424 "addl $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
425 "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
426 " 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
427
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 "movl "BYTE "(%2), %%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
429 "movzwl (%%ecx), %%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
430 "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
431 "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
432 "subl $0xFFFF, %%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
433 "addl $2, %%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
434 "movl %%ecx, "BYTE "(%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
435
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
436 "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
437 "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
438 "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
439 "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
440 "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
441 "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
442
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 "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
444 "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
445 "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
446 "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
447 "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
448 :"=&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
449 :"r"(state), "r"(c)
4012
f8c649ac09dd add "memory" to the clobber list we change memory so we need it, this also fixes some problems with gcc svn
michael
parents: 4008
diff changeset
450 : "%ecx", "%ebx", "%edx", "%esi", "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
451 );
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
452 bit&=1;
4002
ec426fa57dfe adds some useful comments after some of the #else, #elseif,
gpoirier
parents: 4001
diff changeset
453 #else /* BRANCHLESS_CABAC_DECODER */
3975
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
454 asm volatile(
4035
b7f31a32bb30 cosmetic (%%eax->%0)
michael
parents: 4034
diff changeset
455 "movzbl (%1), %0 \n\t"
3975
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
456 "movl "RANGE "(%2), %%ebx \n\t"
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
457 "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
458 "andl $0xC0, %%ebx \n\t"
4035
b7f31a32bb30 cosmetic (%%eax->%0)
michael
parents: 4034
diff changeset
459 "movzbl "MANGLE(ff_h264_lps_range)"(%0, %%ebx, 2), %%esi\n\t"
3975
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
460 "movl "LOW "(%2), %%ebx \n\t"
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
461 //eax:state ebx:low, edx:range, esi:RangeLPS
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
462 "subl %%esi, %%edx \n\t"
4034
fbe263601152 Fix crash with illegal instruction, cmov is available on 686 and later only.
diego
parents: 4033
diff changeset
463 #if (defined CMOV_IS_FAST && __CPU__ >= 686)
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
464 "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
465 "shl $17, %%edx \n\t"
3980
5afe4253a220 replace a few and/sub/... by cmov
michael
parents: 3979
diff changeset
466 "cmpl %%ebx, %%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
467 "cmova %%ecx, %%esi \n\t"
3980
5afe4253a220 replace a few and/sub/... by cmov
michael
parents: 3979
diff changeset
468 "sbbl %%ecx, %%ecx \n\t"
5afe4253a220 replace a few and/sub/... by cmov
michael
parents: 3979
diff changeset
469 "andl %%ecx, %%edx \n\t"
5afe4253a220 replace a few and/sub/... by cmov
michael
parents: 3979
diff changeset
470 "subl %%edx, %%ebx \n\t"
4035
b7f31a32bb30 cosmetic (%%eax->%0)
michael
parents: 4034
diff changeset
471 "xorl %%ecx, %0 \n\t"
4002
ec426fa57dfe adds some useful comments after some of the #else, #elseif,
gpoirier
parents: 4001
diff changeset
472 #else /* CMOV_IS_FAST */
3975
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
473 "movl %%edx, %%ecx \n\t"
4027
ca4a837b896d fix !CMOV_IS_FAST case (iam not really happy with the fix but i didnt come up with a better one quickly)
michael
parents: 4026
diff changeset
474 "shl $17, %%edx \n\t"
3975
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
475 "subl %%ebx, %%edx \n\t"
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
476 "sarl $31, %%edx \n\t" //lps_mask
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
477 "subl %%ecx, %%esi \n\t" //RangeLPS - range
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
478 "andl %%edx, %%esi \n\t" //(RangeLPS - range)&lps_mask
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
479 "addl %%ecx, %%esi \n\t" //new range
4027
ca4a837b896d fix !CMOV_IS_FAST case (iam not really happy with the fix but i didnt come up with a better one quickly)
michael
parents: 4026
diff changeset
480 "shl $17, %%ecx \n\t"
3975
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
481 "andl %%edx, %%ecx \n\t"
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
482 "subl %%ecx, %%ebx \n\t"
4035
b7f31a32bb30 cosmetic (%%eax->%0)
michael
parents: 4034
diff changeset
483 "xorl %%edx, %0 \n\t"
4002
ec426fa57dfe adds some useful comments after some of the #else, #elseif,
gpoirier
parents: 4001
diff changeset
484 #endif /* CMOV_IS_FAST */
3975
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
485
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
486 //eax:state ebx:low edx:mask esi:range
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
487
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
488 //eax:bit ebx:low esi:range
3975
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
489
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
490 "movzbl " MANGLE(ff_h264_norm_shift) "(%%esi), %%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
491 "shll %%cl, %%esi \n\t"
4035
b7f31a32bb30 cosmetic (%%eax->%0)
michael
parents: 4034
diff changeset
492 "movzbl "MANGLE(ff_h264_mlps_state)"+128(%0), %%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
493 "movb %%dl, (%1) \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
494 "movl %%esi, "RANGE "(%2) \n\t"
3975
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
495 "shll %%cl, %%ebx \n\t"
3995
b00c06477dff write cabac low and range variables as early as possible to prevent stalls from reading them before they where written, the P4 is said to disslike that alot, on P3 its 2% faster (START/STOP_TIMER over decode_residual)
michael
parents: 3994
diff changeset
496 "movl %%ebx, "LOW "(%2) \n\t"
3975
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
497 "test %%bx, %%bx \n\t"
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
498 " jnz 1f \n\t"
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
499
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
500 "movl "BYTE "(%2), %%ecx \n\t"
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
501 "movzwl (%%ecx), %%esi \n\t"
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
502 "bswap %%esi \n\t"
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
503 "shrl $15, %%esi \n\t"
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
504 "subl $0xFFFF, %%esi \n\t"
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
505 "addl $2, %%ecx \n\t"
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
506 "movl %%ecx, "BYTE "(%2) \n\t"
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
507
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
508 "leal -1(%%ebx), %%ecx \n\t"
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
509 "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
510 "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
511 "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
512 "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
513 "add $7, %%ecx \n\t"
3975
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
514
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
515 "shll %%cl , %%esi \n\t"
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
516 "addl %%esi, %%ebx \n\t"
3995
b00c06477dff write cabac low and range variables as early as possible to prevent stalls from reading them before they where written, the P4 is said to disslike that alot, on P3 its 2% faster (START/STOP_TIMER over decode_residual)
michael
parents: 3994
diff changeset
517 "movl %%ebx, "LOW "(%2) \n\t"
3975
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
518 "1: \n\t"
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
519 :"=&a"(bit)
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
520 :"r"(state), "r"(c)
4012
f8c649ac09dd add "memory" to the clobber list we change memory so we need it, this also fixes some problems with gcc svn
michael
parents: 4008
diff changeset
521 : "%ecx", "%ebx", "%edx", "%esi", "memory"
3975
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
522 );
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
523 bit&=1;
4002
ec426fa57dfe adds some useful comments after some of the #else, #elseif,
gpoirier
parents: 4001
diff changeset
524 #endif /* BRANCHLESS_CABAC_DECODER */
4033
f7a6b2bb3a2f Expand some #endif comments.
diego
parents: 4027
diff changeset
525 #else /* defined(ARCH_X86) && !(defined(PIC) && defined(__GNUC__)) */
3642
0efda682253c tweak cabac. 0.5% faster h264.
lorenm
parents: 3284
diff changeset
526 int s = *state;
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
527 int RangeLPS= ff_h264_lps_range[0][2*(c->range&0xC0) + s];
2522
e25782262d7d kill warnings patch by (M«©ns Rullg«©rd <mru inprovide com>)
michael
parents: 2323
diff changeset
528 int bit, lps_mask attribute_unused;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
529
1287
michaelni
parents:
diff changeset
530 c->range -= RangeLPS;
3984
bb186452e7da BRANCHLESS_CABAD --> BRANCHLESS_CABAC_DECODER
diego
parents: 3982
diff changeset
531 #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
532 if(c->low < (c->range<<17)){
3642
0efda682253c tweak cabac. 0.5% faster h264.
lorenm
parents: 3284
diff changeset
533 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
534 *state= ff_h264_mps_state[s];
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
535 renorm_cabac_decoder_once(c);
1287
michaelni
parents:
diff changeset
536 }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
537 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
538 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
539 *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
540 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
541 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
542 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
543
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
544 if(!(c->low & 0xFFFF)){
1c39d9786efd optimization
michael
parents: 2116
diff changeset
545 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
546 }
1287
michaelni
parents:
diff changeset
547 }
4002
ec426fa57dfe adds some useful comments after some of the #else, #elseif,
gpoirier
parents: 4001
diff changeset
548 #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
549 lps_mask= ((c->range<<17) - c->low)>>31;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
550
4024
d550343b5dac shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents: 4014
diff changeset
551 c->low -= (c->range<<17) & lps_mask;
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
552 c->range += (RangeLPS - c->range) & lps_mask;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
553
3974
93746612bc78 optimize branchless C CABAC decoder
michael
parents: 3972
diff changeset
554 s^=lps_mask;
4014
b2582438effe dehack *ps_state indexing in the branchless decoder
michael
parents: 4012
diff changeset
555 *state= (ff_h264_mlps_state+128)[s];
3974
93746612bc78 optimize branchless C CABAC decoder
michael
parents: 3972
diff changeset
556 bit= s&1;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
557
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
558 lps_mask= ff_h264_norm_shift[c->range];
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
559 c->range<<= lps_mask;
1c39d9786efd optimization
michael
parents: 2116
diff changeset
560 c->low <<= lps_mask;
1c39d9786efd optimization
michael
parents: 2116
diff changeset
561 if(!(c->low & CABAC_MASK))
1c39d9786efd optimization
michael
parents: 2116
diff changeset
562 refill2(c);
4002
ec426fa57dfe adds some useful comments after some of the #else, #elseif,
gpoirier
parents: 4001
diff changeset
563 #endif /* BRANCHLESS_CABAC_DECODER */
4033
f7a6b2bb3a2f Expand some #endif comments.
diego
parents: 4027
diff changeset
564 #endif /* defined(ARCH_X86) && !(defined(PIC) && defined(__GNUC__)) */
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
565 return bit;
1287
michaelni
parents:
diff changeset
566 }
michaelni
parents:
diff changeset
567
4008
b636f3d59283 prevent "mb level" get_cabac() calls from being inlined (3% faster decode_mb_cabac() on P3)
michael
parents: 4002
diff changeset
568 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
569 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
570 }
b636f3d59283 prevent "mb level" get_cabac() calls from being inlined (3% faster decode_mb_cabac() on P3)
michael
parents: 4002
diff changeset
571
b636f3d59283 prevent "mb level" get_cabac() calls from being inlined (3% faster decode_mb_cabac() on P3)
michael
parents: 4002
diff changeset
572 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
573 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
574 }
b636f3d59283 prevent "mb level" get_cabac() calls from being inlined (3% faster decode_mb_cabac() on P3)
michael
parents: 4002
diff changeset
575
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
576 static int get_cabac_bypass(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
577 int range;
1287
michaelni
parents:
diff changeset
578 c->low += c->low;
michaelni
parents:
diff changeset
579
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
580 if(!(c->low & CABAC_MASK))
1c39d9786efd optimization
michael
parents: 2116
diff changeset
581 refill(c);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
582
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
583 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
584 if(c->low < range){
1287
michaelni
parents:
diff changeset
585 return 0;
michaelni
parents:
diff changeset
586 }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
587 c->low -= range;
1287
michaelni
parents:
diff changeset
588 return 1;
michaelni
parents:
diff changeset
589 }
michaelni
parents:
diff changeset
590 }
michaelni
parents:
diff changeset
591
1300
e18667d1e94d FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents: 1298
diff changeset
592 /**
e18667d1e94d FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents: 1298
diff changeset
593 *
e18667d1e94d FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents: 1298
diff changeset
594 * @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
595 */
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
596 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
597 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
598 if(c->low < c->range<<17){
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
599 renorm_cabac_decoder_once(c);
1287
michaelni
parents:
diff changeset
600 return 0;
michaelni
parents:
diff changeset
601 }else{
1300
e18667d1e94d FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents: 1298
diff changeset
602 return c->bytestream - c->bytestream_start;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
603 }
1287
michaelni
parents:
diff changeset
604 }
michaelni
parents:
diff changeset
605
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
606 /**
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
607 * get (truncated) unnary binarization.
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
608 */
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
609 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
610 int i;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
611
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
612 for(i=0; i<max; i++){
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
613 if(get_cabac(c, state)==0)
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
614 return i;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
615
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
616 if(i< max_index) state++;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
617 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
618
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
619 return truncated ? max : -1;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
620 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
621
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
622 /**
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
623 * get unary exp golomb k-th order binarization.
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
624 */
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
625 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
626 int i, v;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
627 int m= 1<<k;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
628
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
629 if(get_cabac(c, state)==0)
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
630 return 0;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
631
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
632 if(0 < max_index) state++;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
633
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
634 for(i=1; i<max; i++){
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
635 if(get_cabac(c, state)==0){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
636 if(is_signed && get_cabac_bypass(c)){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
637 return -i;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
638 }else
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
639 return i;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
640 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
641
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
642 if(i < max_index) state++;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
643 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
644
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
645 while(get_cabac_bypass(c)){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
646 i+= m;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
647 m+= m;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
648 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
649
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
650 v=0;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
651 while(m>>=1){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
652 v+= v + get_cabac_bypass(c);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
653 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
654 i += v;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
655
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
656 if(is_signed && get_cabac_bypass(c)){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
657 return -i;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
658 }else
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
659 return i;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
660 }