annotate cabac.h @ 4043:87694a28120c libavcodec

unused variable
author michael
date Thu, 19 Oct 2006 18:22:03 +0000
parents 9eaea06c5ba6
children 5ccdefd60f61
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
4039
866a83726985 Kill a warning with MSVC
gpoirier
parents: 4037
diff changeset
36 //#define ARCH_X86_DISABLED 1
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
37
1287
michaelni
parents:
diff changeset
38 typedef struct CABACContext{
michaelni
parents:
diff changeset
39 int low;
michaelni
parents:
diff changeset
40 int range;
michaelni
parents:
diff changeset
41 int outstanding_count;
michaelni
parents:
diff changeset
42 #ifdef STRICT_LIMITS
michaelni
parents:
diff changeset
43 int symCount;
michaelni
parents:
diff changeset
44 #endif
2024
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_start;
f65d87bfdd5a some of the warning fixes by (Michael Roitzsch <mroi at users dot sourceforge dot net>)
michael
parents: 1787
diff changeset
46 const uint8_t *bytestream;
2116
48d9f86fb047 overread fix
michael
parents: 2024
diff changeset
47 const uint8_t *bytestream_end;
1287
michaelni
parents:
diff changeset
48 PutBitContext pb;
michaelni
parents:
diff changeset
49 }CABACContext;
michaelni
parents:
diff changeset
50
4014
b2582438effe dehack *ps_state indexing in the branchless decoder
michael
parents: 4012
diff changeset
51 extern uint8_t ff_h264_mlps_state[4*64];
4039
866a83726985 Kill a warning with MSVC
gpoirier
parents: 4037
diff changeset
52 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
53 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
54 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
55 extern const uint8_t ff_h264_norm_shift[512];
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
56
1287
michaelni
parents:
diff changeset
57
michaelni
parents:
diff changeset
58 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
59 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
60 void ff_init_cabac_states(CABACContext *c);
1287
michaelni
parents:
diff changeset
61
michaelni
parents:
diff changeset
62
michaelni
parents:
diff changeset
63 static inline void put_cabac_bit(CABACContext *c, int b){
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
64 put_bits(&c->pb, 1, b);
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
65 for(;c->outstanding_count; c->outstanding_count--){
1287
michaelni
parents:
diff changeset
66 put_bits(&c->pb, 1, 1-b);
michaelni
parents:
diff changeset
67 }
michaelni
parents:
diff changeset
68 }
michaelni
parents:
diff changeset
69
michaelni
parents:
diff changeset
70 static inline void renorm_cabac_encoder(CABACContext *c){
michaelni
parents:
diff changeset
71 while(c->range < 0x100){
michaelni
parents:
diff changeset
72 //FIXME optimize
michaelni
parents:
diff changeset
73 if(c->low<0x100){
michaelni
parents:
diff changeset
74 put_cabac_bit(c, 0);
michaelni
parents:
diff changeset
75 }else if(c->low<0x200){
michaelni
parents:
diff changeset
76 c->outstanding_count++;
michaelni
parents:
diff changeset
77 c->low -= 0x100;
michaelni
parents:
diff changeset
78 }else{
michaelni
parents:
diff changeset
79 put_cabac_bit(c, 1);
michaelni
parents:
diff changeset
80 c->low -= 0x200;
michaelni
parents:
diff changeset
81 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
82
1287
michaelni
parents:
diff changeset
83 c->range+= c->range;
michaelni
parents:
diff changeset
84 c->low += c->low;
michaelni
parents:
diff changeset
85 }
michaelni
parents:
diff changeset
86 }
michaelni
parents:
diff changeset
87
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
88 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
89 int RangeLPS= ff_h264_lps_range[2*(c->range&0xC0) + *state];
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
90
1287
michaelni
parents:
diff changeset
91 if(bit == ((*state)&1)){
michaelni
parents:
diff changeset
92 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
93 *state= ff_h264_mps_state[*state];
1287
michaelni
parents:
diff changeset
94 }else{
michaelni
parents:
diff changeset
95 c->low += c->range - RangeLPS;
michaelni
parents:
diff changeset
96 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
97 *state= ff_h264_lps_state[*state];
1287
michaelni
parents:
diff changeset
98 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
99
1287
michaelni
parents:
diff changeset
100 renorm_cabac_encoder(c);
michaelni
parents:
diff changeset
101
michaelni
parents:
diff changeset
102 #ifdef STRICT_LIMITS
michaelni
parents:
diff changeset
103 c->symCount++;
michaelni
parents:
diff changeset
104 #endif
michaelni
parents:
diff changeset
105 }
michaelni
parents:
diff changeset
106
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
107 static void put_cabac_static(CABACContext *c, int RangeLPS, int bit){
1287
michaelni
parents:
diff changeset
108 assert(c->range > RangeLPS);
michaelni
parents:
diff changeset
109
michaelni
parents:
diff changeset
110 if(!bit){
michaelni
parents:
diff changeset
111 c->range -= RangeLPS;
michaelni
parents:
diff changeset
112 }else{
michaelni
parents:
diff changeset
113 c->low += c->range - RangeLPS;
michaelni
parents:
diff changeset
114 c->range = RangeLPS;
michaelni
parents:
diff changeset
115 }
michaelni
parents:
diff changeset
116
michaelni
parents:
diff changeset
117 renorm_cabac_encoder(c);
michaelni
parents:
diff changeset
118
michaelni
parents:
diff changeset
119 #ifdef STRICT_LIMITS
michaelni
parents:
diff changeset
120 c->symCount++;
michaelni
parents:
diff changeset
121 #endif
michaelni
parents:
diff changeset
122 }
michaelni
parents:
diff changeset
123
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
124 /**
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
125 * @param bit 0 -> write zero bit, !=0 write one bit
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
126 */
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
127 static void put_cabac_bypass(CABACContext *c, int bit){
1287
michaelni
parents:
diff changeset
128 c->low += c->low;
michaelni
parents:
diff changeset
129
michaelni
parents:
diff changeset
130 if(bit){
michaelni
parents:
diff changeset
131 c->low += c->range;
michaelni
parents:
diff changeset
132 }
michaelni
parents:
diff changeset
133 //FIXME optimize
michaelni
parents:
diff changeset
134 if(c->low<0x200){
michaelni
parents:
diff changeset
135 put_cabac_bit(c, 0);
michaelni
parents:
diff changeset
136 }else if(c->low<0x400){
michaelni
parents:
diff changeset
137 c->outstanding_count++;
michaelni
parents:
diff changeset
138 c->low -= 0x200;
michaelni
parents:
diff changeset
139 }else{
michaelni
parents:
diff changeset
140 put_cabac_bit(c, 1);
michaelni
parents:
diff changeset
141 c->low -= 0x400;
michaelni
parents:
diff changeset
142 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
143
1287
michaelni
parents:
diff changeset
144 #ifdef STRICT_LIMITS
michaelni
parents:
diff changeset
145 c->symCount++;
michaelni
parents:
diff changeset
146 #endif
michaelni
parents:
diff changeset
147 }
michaelni
parents:
diff changeset
148
1300
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 *
e18667d1e94d FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents: 1298
diff changeset
151 * @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
152 */
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
153 static int put_cabac_terminate(CABACContext *c, int bit){
1287
michaelni
parents:
diff changeset
154 c->range -= 2;
michaelni
parents:
diff changeset
155
michaelni
parents:
diff changeset
156 if(!bit){
michaelni
parents:
diff changeset
157 renorm_cabac_encoder(c);
michaelni
parents:
diff changeset
158 }else{
michaelni
parents:
diff changeset
159 c->low += c->range;
michaelni
parents:
diff changeset
160 c->range= 2;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
161
1287
michaelni
parents:
diff changeset
162 renorm_cabac_encoder(c);
michaelni
parents:
diff changeset
163
michaelni
parents:
diff changeset
164 assert(c->low <= 0x1FF);
michaelni
parents:
diff changeset
165 put_cabac_bit(c, c->low>>9);
michaelni
parents:
diff changeset
166 put_bits(&c->pb, 2, ((c->low>>7)&3)|1);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
167
1287
michaelni
parents:
diff changeset
168 flush_put_bits(&c->pb); //FIXME FIXME FIXME XXX wrong
michaelni
parents:
diff changeset
169 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
170
1287
michaelni
parents:
diff changeset
171 #ifdef STRICT_LIMITS
michaelni
parents:
diff changeset
172 c->symCount++;
michaelni
parents:
diff changeset
173 #endif
1300
e18667d1e94d FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents: 1298
diff changeset
174
1787
752b51a3c8ed get_bit_count -> put_bits_count
alex
parents: 1301
diff changeset
175 return (put_bits_count(&c->pb)+7)>>3;
1287
michaelni
parents:
diff changeset
176 }
michaelni
parents:
diff changeset
177
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
178 /**
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
179 * put (truncated) unary binarization.
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
180 */
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
181 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
182 int i;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
183
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
184 assert(v <= max);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
185
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
186 #if 1
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
187 for(i=0; i<v; i++){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
188 put_cabac(c, state, 1);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
189 if(i < max_index) state++;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
190 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
191 if(truncated==0 || v<max)
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
192 put_cabac(c, state, 0);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
193 #else
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
194 if(v <= max_index){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
195 for(i=0; i<v; i++){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
196 put_cabac(c, state+i, 1);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
197 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
198 if(truncated==0 || v<max)
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
199 put_cabac(c, state+i, 0);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
200 }else{
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
201 for(i=0; i<=max_index; i++){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
202 put_cabac(c, state+i, 1);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
203 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
204 for(; i<v; i++){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
205 put_cabac(c, state+max_index, 1);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
206 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
207 if(truncated==0 || v<max)
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
208 put_cabac(c, state+max_index, 0);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
209 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
210 #endif
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 /**
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
214 * put unary exp golomb k-th order binarization.
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
215 */
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
216 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
217 int i;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
218
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
219 if(v==0)
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
220 put_cabac(c, state, 0);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
221 else{
1298
5bc3184810dc cleanup
michaelni
parents: 1290
diff changeset
222 const int sign= v < 0;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
223
4001
34fdffe98bd0 Rename ABS macro to FFABS.
diego
parents: 3999
diff changeset
224 if(is_signed) v= FFABS(v);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
225
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
226 if(v<max){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
227 for(i=0; i<v; i++){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
228 put_cabac(c, state, 1);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
229 if(i < max_index) state++;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
230 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
231
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
232 put_cabac(c, state, 0);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
233 }else{
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
234 int m= 1<<k;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
235
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
236 for(i=0; i<max; i++){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
237 put_cabac(c, state, 1);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
238 if(i < max_index) state++;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
239 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
240
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
241 v -= max;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
242 while(v >= m){ //FIXME optimize
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
243 put_cabac_bypass(c, 1);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
244 v-= m;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
245 m+= m;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
246 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
247 put_cabac_bypass(c, 0);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
248 while(m>>=1){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
249 put_cabac_bypass(c, v&m);
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
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
253 if(is_signed)
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
254 put_cabac_bypass(c, sign);
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
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
258 static void refill(CABACContext *c){
1c39d9786efd optimization
michael
parents: 2116
diff changeset
259 #if CABAC_BITS == 16
3946
michael
parents: 3943
diff changeset
260 c->low+= (c->bytestream[0]<<9) + (c->bytestream[1]<<1);
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
261 #else
1c39d9786efd optimization
michael
parents: 2116
diff changeset
262 c->low+= c->bytestream[0]<<1;
1c39d9786efd optimization
michael
parents: 2116
diff changeset
263 #endif
1c39d9786efd optimization
michael
parents: 2116
diff changeset
264 c->low -= CABAC_MASK;
1c39d9786efd optimization
michael
parents: 2116
diff changeset
265 c->bytestream+= CABAC_BITS/8;
1c39d9786efd optimization
michael
parents: 2116
diff changeset
266 }
1c39d9786efd optimization
michael
parents: 2116
diff changeset
267
1c39d9786efd optimization
michael
parents: 2116
diff changeset
268 static void refill2(CABACContext *c){
1c39d9786efd optimization
michael
parents: 2116
diff changeset
269 int i, x;
1c39d9786efd optimization
michael
parents: 2116
diff changeset
270
1c39d9786efd optimization
michael
parents: 2116
diff changeset
271 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
272 i= 7 - ff_h264_norm_shift[x>>(CABAC_BITS-1)];
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
273
1c39d9786efd optimization
michael
parents: 2116
diff changeset
274 x= -CABAC_MASK;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
275
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
276 #if CABAC_BITS == 16
1c39d9786efd optimization
michael
parents: 2116
diff changeset
277 x+= (c->bytestream[0]<<9) + (c->bytestream[1]<<1);
1c39d9786efd optimization
michael
parents: 2116
diff changeset
278 #else
1c39d9786efd optimization
michael
parents: 2116
diff changeset
279 x+= c->bytestream[0]<<1;
1c39d9786efd optimization
michael
parents: 2116
diff changeset
280 #endif
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
281
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
282 c->low += x<<i;
1c39d9786efd optimization
michael
parents: 2116
diff changeset
283 c->bytestream+= CABAC_BITS/8;
1c39d9786efd optimization
michael
parents: 2116
diff changeset
284 }
1c39d9786efd optimization
michael
parents: 2116
diff changeset
285
1287
michaelni
parents:
diff changeset
286 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
287 while(c->range < 0x100){
1287
michaelni
parents:
diff changeset
288 c->range+= c->range;
michaelni
parents:
diff changeset
289 c->low+= c->low;
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
290 if(!(c->low & CABAC_MASK))
1c39d9786efd optimization
michael
parents: 2116
diff changeset
291 refill(c);
1287
michaelni
parents:
diff changeset
292 }
michaelni
parents:
diff changeset
293 }
michaelni
parents:
diff changeset
294
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
295 static inline void renorm_cabac_decoder_once(CABACContext *c){
3951
751bfc30df72 disable benchmarking code
michael
parents: 3950
diff changeset
296 #ifdef ARCH_X86_DISABLED
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
297 int temp;
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
298 #if 0
3950
900d21b85dd6 renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents: 3948
diff changeset
299 //P3:683 athlon:475
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
300 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
301 "lea -0x100(%0), %2 \n\t"
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
302 "shr $31, %2 \n\t" //FIXME 31->63 for x86-64
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
303 "shl %%cl, %0 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
304 "shl %%cl, %1 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
305 : "+r"(c->range), "+r"(c->low), "+c"(temp)
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
306 );
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
307 #elif 0
3950
900d21b85dd6 renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents: 3948
diff changeset
308 //P3:680 athlon:474
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
309 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
310 "cmp $0x100, %0 \n\t"
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
311 "setb %%cl \n\t" //FIXME 31->63 for x86-64
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
312 "shl %%cl, %0 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
313 "shl %%cl, %1 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
314 : "+r"(c->range), "+r"(c->low), "+c"(temp)
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
315 );
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
316 #elif 1
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
317 int temp2;
3950
900d21b85dd6 renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents: 3948
diff changeset
318 //P3:665 athlon:517
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
319 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
320 "lea -0x100(%0), %%eax \n\t"
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
321 "cdq \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
322 "mov %0, %%eax \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
323 "and %%edx, %0 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
324 "and %1, %%edx \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
325 "add %%eax, %0 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
326 "add %%edx, %1 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
327 : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2)
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
328 );
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
329 #elif 0
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
330 int temp2;
3950
900d21b85dd6 renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents: 3948
diff changeset
331 //P3:673 athlon:509
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
332 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
333 "cmp $0x100, %0 \n\t"
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
334 "sbb %%edx, %%edx \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
335 "mov %0, %%eax \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
336 "and %%edx, %0 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
337 "and %1, %%edx \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
338 "add %%eax, %0 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
339 "add %%edx, %1 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
340 : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2)
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
341 );
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
342 #else
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
343 int temp2;
3950
900d21b85dd6 renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents: 3948
diff changeset
344 //P3:677 athlon:511
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
345 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
346 "cmp $0x100, %0 \n\t"
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
347 "lea (%0, %0), %%eax \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
348 "lea (%1, %1), %%edx \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
349 "cmovb %%eax, %0 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
350 "cmovb %%edx, %1 \n\t"
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
351 : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2)
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
352 );
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
353 #endif
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
354 #else
3950
900d21b85dd6 renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents: 3948
diff changeset
355 //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
356 int shift= (uint32_t)(c->range - 0x100)>>31;
3642
0efda682253c tweak cabac. 0.5% faster h264.
lorenm
parents: 3284
diff changeset
357 c->range<<= shift;
0efda682253c tweak cabac. 0.5% faster h264.
lorenm
parents: 3284
diff changeset
358 c->low <<= shift;
3943
811a9b0d9f32 several x86 renorm_cabac_decoder_once optimizations
michael
parents: 3928
diff changeset
359 #endif
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
360 if(!(c->low & CABAC_MASK))
1c39d9786efd optimization
michael
parents: 2116
diff changeset
361 refill(c);
1c39d9786efd optimization
michael
parents: 2116
diff changeset
362 }
1c39d9786efd optimization
michael
parents: 2116
diff changeset
363
4008
b636f3d59283 prevent "mb level" get_cabac() calls from being inlined (3% faster decode_mb_cabac() on P3)
michael
parents: 4002
diff changeset
364 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
365 //FIXME gcc generates duplicate load/stores for c->low and c->range
4026
michael
parents: 4025
diff changeset
366 #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
367 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
368
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 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
370 #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
371 #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
372 #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
373 #define BYTEEND "20"
3984
bb186452e7da BRANCHLESS_CABAD --> BRANCHLESS_CABAC_DECODER
diego
parents: 3982
diff changeset
374 #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
375 asm volatile(
4035
b7f31a32bb30 cosmetic (%%eax->%0)
michael
parents: 4034
diff changeset
376 "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
377 "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
378 "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
379 "andl $0xC0, %%ebx \n\t"
4035
b7f31a32bb30 cosmetic (%%eax->%0)
michael
parents: 4034
diff changeset
380 "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
381 "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
382 //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
383 "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
384 "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
385 "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
386 "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
387 " ja 1f \n\t"
3999
6cbad3675632 slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents: 3996
diff changeset
388
6cbad3675632 slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents: 3996
diff changeset
389 #if 1
6cbad3675632 slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents: 3996
diff changeset
390 //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
391 "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
392 "shr $31, %%ecx \n\t"
6cbad3675632 slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents: 3996
diff changeset
393 "shl %%cl, %%edx \n\t"
6cbad3675632 slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents: 3996
diff changeset
394 "shl %%cl, %%ebx \n\t"
6cbad3675632 slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents: 3996
diff changeset
395 #else
6cbad3675632 slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents: 3996
diff changeset
396 //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
397 "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
398 "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
399 "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
400 "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
401 #endif
4035
b7f31a32bb30 cosmetic (%%eax->%0)
michael
parents: 4034
diff changeset
402 "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
403 "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
404 //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
405 "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
406 " 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
407 "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
408 "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
409 "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
410 "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
411 "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
412 "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
413 "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
414 "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
415 "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
416 "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
417 //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
418 "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
419 "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
420 "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
421 "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
422 "shll %%cl, %%edx \n\t"
4035
b7f31a32bb30 cosmetic (%%eax->%0)
michael
parents: 4034
diff changeset
423 "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
424 "movb %%cl, (%1) \n\t"
4035
b7f31a32bb30 cosmetic (%%eax->%0)
michael
parents: 4034
diff changeset
425 "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
426 "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
427 " 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
428
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 "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
430 "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
431 "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
432 "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
433 "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
434 "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
435 "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
436
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 "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
438 "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
439 "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
440 "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
441 "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
442 "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
443
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 "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
445 "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
446 "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 %%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
448 "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
449 :"=&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
450 :"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
451 : "%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
452 );
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
453 bit&=1;
4002
ec426fa57dfe adds some useful comments after some of the #else, #elseif,
gpoirier
parents: 4001
diff changeset
454 #else /* BRANCHLESS_CABAC_DECODER */
3975
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
455 asm volatile(
4035
b7f31a32bb30 cosmetic (%%eax->%0)
michael
parents: 4034
diff changeset
456 "movzbl (%1), %0 \n\t"
3975
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
457 "movl "RANGE "(%2), %%ebx \n\t"
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
458 "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
459 "andl $0xC0, %%ebx \n\t"
4035
b7f31a32bb30 cosmetic (%%eax->%0)
michael
parents: 4034
diff changeset
460 "movzbl "MANGLE(ff_h264_lps_range)"(%0, %%ebx, 2), %%esi\n\t"
3975
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
461 "movl "LOW "(%2), %%ebx \n\t"
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
462 //eax:state ebx:low, edx:range, esi:RangeLPS
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
463 "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
464 #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
465 "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
466 "shl $17, %%edx \n\t"
3980
5afe4253a220 replace a few and/sub/... by cmov
michael
parents: 3979
diff changeset
467 "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
468 "cmova %%ecx, %%esi \n\t"
3980
5afe4253a220 replace a few and/sub/... by cmov
michael
parents: 3979
diff changeset
469 "sbbl %%ecx, %%ecx \n\t"
5afe4253a220 replace a few and/sub/... by cmov
michael
parents: 3979
diff changeset
470 "andl %%ecx, %%edx \n\t"
5afe4253a220 replace a few and/sub/... by cmov
michael
parents: 3979
diff changeset
471 "subl %%edx, %%ebx \n\t"
4035
b7f31a32bb30 cosmetic (%%eax->%0)
michael
parents: 4034
diff changeset
472 "xorl %%ecx, %0 \n\t"
4002
ec426fa57dfe adds some useful comments after some of the #else, #elseif,
gpoirier
parents: 4001
diff changeset
473 #else /* CMOV_IS_FAST */
3975
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
474 "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
475 "shl $17, %%edx \n\t"
3975
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
476 "subl %%ebx, %%edx \n\t"
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
477 "sarl $31, %%edx \n\t" //lps_mask
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
478 "subl %%ecx, %%esi \n\t" //RangeLPS - range
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
479 "andl %%edx, %%esi \n\t" //(RangeLPS - range)&lps_mask
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
480 "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
481 "shl $17, %%ecx \n\t"
3975
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
482 "andl %%edx, %%ecx \n\t"
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
483 "subl %%ecx, %%ebx \n\t"
4035
b7f31a32bb30 cosmetic (%%eax->%0)
michael
parents: 4034
diff changeset
484 "xorl %%edx, %0 \n\t"
4002
ec426fa57dfe adds some useful comments after some of the #else, #elseif,
gpoirier
parents: 4001
diff changeset
485 #endif /* CMOV_IS_FAST */
3975
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
486
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
487 //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
488
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
489 //eax:bit ebx:low esi:range
3975
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
490
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
491 "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
492 "shll %%cl, %%esi \n\t"
4035
b7f31a32bb30 cosmetic (%%eax->%0)
michael
parents: 4034
diff changeset
493 "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
494 "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
495 "movl %%esi, "RANGE "(%2) \n\t"
3975
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
496 "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
497 "movl %%ebx, "LOW "(%2) \n\t"
3975
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
498 "test %%bx, %%bx \n\t"
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
499 " jnz 1f \n\t"
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
500
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
501 "movl "BYTE "(%2), %%ecx \n\t"
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
502 "movzwl (%%ecx), %%esi \n\t"
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
503 "bswap %%esi \n\t"
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
504 "shrl $15, %%esi \n\t"
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
505 "subl $0xFFFF, %%esi \n\t"
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
506 "addl $2, %%ecx \n\t"
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
507 "movl %%ecx, "BYTE "(%2) \n\t"
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
508
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
509 "leal -1(%%ebx), %%ecx \n\t"
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
510 "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
511 "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
512 "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
513 "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
514 "add $7, %%ecx \n\t"
3975
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
515
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
516 "shll %%cl , %%esi \n\t"
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
517 "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
518 "movl %%ebx, "LOW "(%2) \n\t"
3975
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
519 "1: \n\t"
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
520 :"=&a"(bit)
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
521 :"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
522 : "%ecx", "%ebx", "%edx", "%esi", "memory"
3975
6cc9eb5ee5e3 x86 branchless cabac decoder
michael
parents: 3974
diff changeset
523 );
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
524 bit&=1;
4002
ec426fa57dfe adds some useful comments after some of the #else, #elseif,
gpoirier
parents: 4001
diff changeset
525 #endif /* BRANCHLESS_CABAC_DECODER */
4033
f7a6b2bb3a2f Expand some #endif comments.
diego
parents: 4027
diff changeset
526 #else /* defined(ARCH_X86) && !(defined(PIC) && defined(__GNUC__)) */
3642
0efda682253c tweak cabac. 0.5% faster h264.
lorenm
parents: 3284
diff changeset
527 int s = *state;
4039
866a83726985 Kill a warning with MSVC
gpoirier
parents: 4037
diff changeset
528 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
529 int bit, lps_mask attribute_unused;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
530
1287
michaelni
parents:
diff changeset
531 c->range -= RangeLPS;
3984
bb186452e7da BRANCHLESS_CABAD --> BRANCHLESS_CABAC_DECODER
diego
parents: 3982
diff changeset
532 #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
533 if(c->low < (c->range<<17)){
3642
0efda682253c tweak cabac. 0.5% faster h264.
lorenm
parents: 3284
diff changeset
534 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
535 *state= ff_h264_mps_state[s];
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
536 renorm_cabac_decoder_once(c);
1287
michaelni
parents:
diff changeset
537 }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
538 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
539 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
540 *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
541 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
542 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
543 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
544
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
545 if(!(c->low & 0xFFFF)){
1c39d9786efd optimization
michael
parents: 2116
diff changeset
546 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
547 }
1287
michaelni
parents:
diff changeset
548 }
4002
ec426fa57dfe adds some useful comments after some of the #else, #elseif,
gpoirier
parents: 4001
diff changeset
549 #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
550 lps_mask= ((c->range<<17) - c->low)>>31;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
551
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
552 c->low -= (c->range<<17) & lps_mask;
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
553 c->range += (RangeLPS - c->range) & lps_mask;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
554
3974
93746612bc78 optimize branchless C CABAC decoder
michael
parents: 3972
diff changeset
555 s^=lps_mask;
4014
b2582438effe dehack *ps_state indexing in the branchless decoder
michael
parents: 4012
diff changeset
556 *state= (ff_h264_mlps_state+128)[s];
3974
93746612bc78 optimize branchless C CABAC decoder
michael
parents: 3972
diff changeset
557 bit= s&1;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
558
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
559 lps_mask= ff_h264_norm_shift[c->range];
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
560 c->range<<= lps_mask;
1c39d9786efd optimization
michael
parents: 2116
diff changeset
561 c->low <<= lps_mask;
1c39d9786efd optimization
michael
parents: 2116
diff changeset
562 if(!(c->low & CABAC_MASK))
1c39d9786efd optimization
michael
parents: 2116
diff changeset
563 refill2(c);
4002
ec426fa57dfe adds some useful comments after some of the #else, #elseif,
gpoirier
parents: 4001
diff changeset
564 #endif /* BRANCHLESS_CABAC_DECODER */
4033
f7a6b2bb3a2f Expand some #endif comments.
diego
parents: 4027
diff changeset
565 #endif /* defined(ARCH_X86) && !(defined(PIC) && defined(__GNUC__)) */
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
566 return bit;
1287
michaelni
parents:
diff changeset
567 }
michaelni
parents:
diff changeset
568
4008
b636f3d59283 prevent "mb level" get_cabac() calls from being inlined (3% faster decode_mb_cabac() on P3)
michael
parents: 4002
diff changeset
569 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
570 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
571 }
b636f3d59283 prevent "mb level" get_cabac() calls from being inlined (3% faster decode_mb_cabac() on P3)
michael
parents: 4002
diff changeset
572
b636f3d59283 prevent "mb level" get_cabac() calls from being inlined (3% faster decode_mb_cabac() on P3)
michael
parents: 4002
diff changeset
573 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
574 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
575 }
b636f3d59283 prevent "mb level" get_cabac() calls from being inlined (3% faster decode_mb_cabac() on P3)
michael
parents: 4002
diff changeset
576
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
577 static int get_cabac_bypass(CABACContext *c){
4040
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
578 #if 0 //not faster
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
579 int bit;
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
580 asm volatile(
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
581 "movl "RANGE "(%1), %%ebx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
582 "movl "LOW "(%1), %%eax \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
583 "shl $17, %%ebx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
584 "add %%eax, %%eax \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
585 "sub %%ebx, %%eax \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
586 "cdq \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
587 "and %%edx, %%ebx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
588 "add %%ebx, %%eax \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
589 "test %%ax, %%ax \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
590 " jnz 1f \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
591 "movl "BYTE "(%1), %%ebx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
592 "subl $0xFFFF, %%eax \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
593 "movzwl (%%ebx), %%ecx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
594 "bswap %%ecx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
595 "shrl $15, %%ecx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
596 "addl $2, %%ebx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
597 "addl %%ecx, %%eax \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
598 "movl %%ebx, "BYTE "(%1) \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
599 "1: \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
600 "movl %%eax, "LOW "(%1) \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
601
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
602 :"=&d"(bit)
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
603 :"r"(c)
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
604 : "%eax", "%ebx", "%ecx", "memory"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
605 );
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
606 return bit+1;
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
607 #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
608 int range;
1287
michaelni
parents:
diff changeset
609 c->low += c->low;
michaelni
parents:
diff changeset
610
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
611 if(!(c->low & CABAC_MASK))
1c39d9786efd optimization
michael
parents: 2116
diff changeset
612 refill(c);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
613
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
614 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
615 if(c->low < range){
1287
michaelni
parents:
diff changeset
616 return 0;
michaelni
parents:
diff changeset
617 }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
618 c->low -= range;
1287
michaelni
parents:
diff changeset
619 return 1;
michaelni
parents:
diff changeset
620 }
4040
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
621 #endif
1287
michaelni
parents:
diff changeset
622 }
4040
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
623
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
624
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
625 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
626 #ifdef ARCH_X86
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
627 asm volatile(
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
628 "movl "RANGE "(%1), %%ebx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
629 "movl "LOW "(%1), %%eax \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
630 "shl $17, %%ebx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
631 "add %%eax, %%eax \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
632 "sub %%ebx, %%eax \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
633 "cdq \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
634 "and %%edx, %%ebx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
635 "add %%ebx, %%eax \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
636 "xor %%edx, %%ecx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
637 "sub %%edx, %%ecx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
638 "test %%ax, %%ax \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
639 " jnz 1f \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
640 "movl "BYTE "(%1), %%ebx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
641 "subl $0xFFFF, %%eax \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
642 "movzwl (%%ebx), %%edx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
643 "bswap %%edx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
644 "shrl $15, %%edx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
645 "addl $2, %%ebx \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
646 "addl %%edx, %%eax \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
647 "movl %%ebx, "BYTE "(%1) \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
648 "1: \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
649 "movl %%eax, "LOW "(%1) \n\t"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
650
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
651 :"+c"(val)
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
652 :"r"(c)
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
653 : "%eax", "%ebx", "%edx", "memory"
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
654 );
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
655 return val;
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
656 #else
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
657 int range, mask;
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
658 c->low += c->low;
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
659
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
660 if(!(c->low & CABAC_MASK))
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
661 refill(c);
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 range= c->range<<17;
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
664 c->low -= range;
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
665 mask= c->low >> 31;
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
666 range &= mask;
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
667 c->low += range;
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
668 return (val^mask)-mask;
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
669 #endif
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
670 }
9eaea06c5ba6 optimize sign decoding code in decode_residual()
michael
parents: 4039
diff changeset
671
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
672 //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
673 //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
674 #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
675 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
676 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
677 int minusstart= -(int)significant_coeff_ctx_base;
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
678 int minusindex= -(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
679 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
680 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
681 "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
682 "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
683
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 "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
685
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 "movzbl (%1), %0 \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
687 "movl %%esi, %%edx \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
688 "andl $0xC0, %%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
689 "movzbl "MANGLE(ff_h264_lps_range)"(%0, %%esi, 2), %%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
690 /*eax:state ebx:low, edx:range, esi:RangeLPS*/
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 "subl %%esi, %%edx \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
692
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 #if (defined CMOV_IS_FAST && __CPU__ >= 686)
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 %%edx, %%ecx \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 "shl $17, %%edx \n\t"
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
696 "cmpl %%ebx, %%edx \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 "cmova %%ecx, %%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
698 "sbbl %%ecx, %%ecx \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
699 "andl %%ecx, %%edx \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
700 "subl %%edx, %%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
701 "xorl %%ecx, %0 \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 #else /* CMOV_IS_FAST */
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
703 "movl %%edx, %%ecx \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
704 "shl $17, %%edx \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
705 "subl %%ebx, %%edx \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
706 "sarl $31, %%edx \n\t" //lps_mask
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
707 "subl %%ecx, %%esi \n\t" //RangeLPS - range
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
708 "andl %%edx, %%esi \n\t" //(RangeLPS - range)&lps_mask
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 "addl %%ecx, %%esi \n\t" //new range
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
710 "shl $17, %%ecx \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
711 "andl %%edx, %%ecx \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 "subl %%ecx, %%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
713 "xorl %%edx, %0 \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
714 #endif /* CMOV_IS_FAST */
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
715
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 "movzbl " MANGLE(ff_h264_norm_shift) "(%%esi), %%ecx \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
717 "shll %%cl, %%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
718 "movzbl "MANGLE(ff_h264_mlps_state)"+128(%0), %%edx \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
719 "movb %%dl, (%1) \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
720 "shll %%cl, %%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
721 "test %%bx, %%bx \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
722 " jnz 1f \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
723
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 "movl "BYTE "(%3), %%ecx \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
725 "movzwl (%%ecx), %%edx \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
726 "bswap %%edx \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 "shrl $15, %%edx \n\t"
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
728 "subl $0xFFFF, %%edx \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 "addl $2, %%ecx \n\t"
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
730 "movl %%ecx, "BYTE "(%3) \n\t"
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
731
53be304c7f54 x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents: 4035
diff changeset
732 "leal -1(%%ebx), %%ecx \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
733 "xorl %%ebx, %%ecx \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
734 "shrl $15, %%ecx \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
735 "movzbl " MANGLE(ff_h264_norm_shift) "(%%ecx), %%ecx \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
736 "neg %%ecx \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
737 "add $7, %%ecx \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
738
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
739 "shll %%cl , %%edx \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
740 "addl %%edx, %%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
741 "1: \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
742
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
743 "test $1, %0 \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
744 " 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
745
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
746 "movl %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
747 "movl %4, %%ecx \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
748 "addl %1, %%ecx \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
749 "movl %%ecx, (%%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
750 "addl $4, %%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
751 "movl %%eax, %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
752
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
753 "movzbl 61(%1), %0 \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
754 "movl %%esi, %%edx \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
755 "andl $0xC0, %%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
756 "movzbl "MANGLE(ff_h264_lps_range)"(%0, %%esi, 2), %%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
757 /*eax:state ebx:low, edx:range, esi:RangeLPS*/
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
758 "subl %%esi, %%edx \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
759
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
760 #if (defined CMOV_IS_FAST && __CPU__ >= 686)
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
761 "movl %%edx, %%ecx \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
762 "shl $17, %%edx \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
763 "cmpl %%ebx, %%edx \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
764 "cmova %%ecx, %%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
765 "sbbl %%ecx, %%ecx \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
766 "andl %%ecx, %%edx \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
767 "subl %%edx, %%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
768 "xorl %%ecx, %0 \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
769 #else /* CMOV_IS_FAST */
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
770 "movl %%edx, %%ecx \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
771 "shl $17, %%edx \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
772 "subl %%ebx, %%edx \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
773 "sarl $31, %%edx \n\t" //lps_mask
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
774 "subl %%ecx, %%esi \n\t" //RangeLPS - range
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
775 "andl %%edx, %%esi \n\t" //(RangeLPS - range)&lps_mask
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
776 "addl %%ecx, %%esi \n\t" //new range
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
777 "shl $17, %%ecx \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
778 "andl %%edx, %%ecx \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
779 "subl %%ecx, %%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
780 "xorl %%edx, %0 \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
781 #endif /* CMOV_IS_FAST */
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
782
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
783 "movzbl " MANGLE(ff_h264_norm_shift) "(%%esi), %%ecx \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
784 "shll %%cl, %%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
785 "movzbl "MANGLE(ff_h264_mlps_state)"+128(%0), %%edx \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
786 "movb %%dl, 61(%1) \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
787 "shll %%cl, %%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
788 "test %%bx, %%bx \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
789 " jnz 1f \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
790
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
791 "movl "BYTE "(%3), %%ecx \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
792 "movzwl (%%ecx), %%edx \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
793 "bswap %%edx \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
794 "shrl $15, %%edx \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
795 "subl $0xFFFF, %%edx \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
796 "addl $2, %%ecx \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
797 "movl %%ecx, "BYTE "(%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
798
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
799 "leal -1(%%ebx), %%ecx \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
800 "xorl %%ebx, %%ecx \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
801 "shrl $15, %%ecx \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
802 "movzbl " MANGLE(ff_h264_norm_shift) "(%%ecx), %%ecx \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
803 "neg %%ecx \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
804 "add $7, %%ecx \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
805
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
806 "shll %%cl , %%edx \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
807 "addl %%edx, %%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
808 "1: \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
809
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
810 "test $1, %%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
811 " 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
812
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
813 "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
814 "addl $1, %1 \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
815 "cmpl %5, %1 \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
816 " jb 2b \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
817 "movl %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
818 "movl %4, %%ecx \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
819 "addl %1, %%ecx \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
820 "movl %%ecx, (%%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
821 "addl $4, %%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
822 "movl %%eax, %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
823 "4: \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
824 "movl %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
825 "addl %6, %%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
826 "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
827
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
828 "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
829 "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
830 :"=&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
831 :"r"(c), "m"(minusstart), "m"(end), "m"(minusindex)\
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
832 : "%ecx", "%ebx", "%edx", "%esi", "memory"\
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
833 );
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
834 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
835 }
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
836 #endif
1287
michaelni
parents:
diff changeset
837
1300
e18667d1e94d FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents: 1298
diff changeset
838 /**
e18667d1e94d FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents: 1298
diff changeset
839 *
e18667d1e94d FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents: 1298
diff changeset
840 * @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
841 */
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
842 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
843 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
844 if(c->low < c->range<<17){
2323
1c39d9786efd optimization
michael
parents: 2116
diff changeset
845 renorm_cabac_decoder_once(c);
1287
michaelni
parents:
diff changeset
846 return 0;
michaelni
parents:
diff changeset
847 }else{
1300
e18667d1e94d FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents: 1298
diff changeset
848 return c->bytestream - c->bytestream_start;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
849 }
1287
michaelni
parents:
diff changeset
850 }
michaelni
parents:
diff changeset
851
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
852 /**
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
853 * get (truncated) unnary binarization.
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
854 */
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
855 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
856 int i;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
857
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
858 for(i=0; i<max; i++){
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
859 if(get_cabac(c, state)==0)
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
860 return i;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
861
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
862 if(i< max_index) state++;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
863 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
864
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
865 return truncated ? max : -1;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
866 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
867
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
868 /**
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
869 * get unary exp golomb k-th order binarization.
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
870 */
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
871 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
872 int i, v;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
873 int m= 1<<k;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
874
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
875 if(get_cabac(c, state)==0)
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
876 return 0;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
877
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
878 if(0 < max_index) state++;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
879
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
880 for(i=1; i<max; i++){
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
881 if(get_cabac(c, state)==0){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
882 if(is_signed && get_cabac_bypass(c)){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
883 return -i;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
884 }else
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
885 return i;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
886 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
887
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
888 if(i < max_index) state++;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
889 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
890
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
891 while(get_cabac_bypass(c)){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
892 i+= m;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
893 m+= m;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
894 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2736
diff changeset
895
1290
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
896 v=0;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
897 while(m>>=1){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
898 v+= v + get_cabac_bypass(c);
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
899 }
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
900 i += v;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
901
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
902 if(is_signed && get_cabac_bypass(c)){
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
903 return -i;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
904 }else
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
905 return i;
dae280b939ca (truncated) unary binerization
michaelni
parents: 1287
diff changeset
906 }