Mercurial > libavcodec.hg
annotate cabac.h @ 3989:eddcc352c0dc libavcodec
doxy
author | michael |
---|---|
date | Wed, 11 Oct 2006 10:29:00 +0000 |
parents | bb186452e7da |
children | 746a60ba3177 |
rev | line source |
---|---|
1287 | 1 /* |
2 * H.26L/H.264/AVC/JVT/14496-10/... encoder/decoder | |
3 * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at> | |
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 | 8 * modify it under the terms of the GNU Lesser General Public |
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 | 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 | 13 * but WITHOUT ANY WARRANTY; without even the implied warranty of |
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
15 * Lesser General Public License for more details. | |
16 * | |
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 | 20 * |
21 */ | |
2967 | 22 |
1287 | 23 /** |
24 * @file cabac.h | |
25 * Context Adaptive Binary Arithmetic Coder. | |
26 */ | |
27 | |
28 | |
3284
a224d9752912
don't force asserts in release builds. 2% faster h264.
lorenm
parents:
3036
diff
changeset
|
29 //#undef NDEBUG |
1287 | 30 #include <assert.h> |
31 | |
3948
3edbf131ee44
refill cabac variables in 16bit steps, 3% faster get_cabac()
michael
parents:
3947
diff
changeset
|
32 #define CABAC_BITS 16 |
2323 | 33 #define CABAC_MASK ((1<<CABAC_BITS)-1) |
3984 | 34 #define BRANCHLESS_CABAC_DECODER 1 |
2323 | 35 |
1287 | 36 typedef struct CABACContext{ |
37 int low; | |
38 int range; | |
39 int outstanding_count; | |
40 #ifdef STRICT_LIMITS | |
41 int symCount; | |
42 #endif | |
3976
27e90123b346
reverse remainder of the failed attempt to optimize *state=c->mps_state[s]
michael
parents:
3975
diff
changeset
|
43 uint8_t lps_range[2*65][4]; ///< rangeTabLPS |
27e90123b346
reverse remainder of the failed attempt to optimize *state=c->mps_state[s]
michael
parents:
3975
diff
changeset
|
44 uint8_t lps_state[2*64]; ///< transIdxLPS |
27e90123b346
reverse remainder of the failed attempt to optimize *state=c->mps_state[s]
michael
parents:
3975
diff
changeset
|
45 uint8_t mps_state[2*64]; ///< transIdxMPS |
2024
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_start; |
f65d87bfdd5a
some of the warning fixes by (Michael Roitzsch <mroi at users dot sourceforge dot net>)
michael
parents:
1787
diff
changeset
|
47 const uint8_t *bytestream; |
2116 | 48 const uint8_t *bytestream_end; |
1287 | 49 PutBitContext pb; |
50 }CABACContext; | |
51 | |
1301 | 52 extern const uint8_t ff_h264_lps_range[64][4]; |
53 extern const uint8_t ff_h264_mps_state[64]; | |
54 extern const uint8_t ff_h264_lps_state[64]; | |
3964 | 55 extern const uint8_t ff_h264_norm_shift[128]; |
2323 | 56 |
1287 | 57 |
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); |
2967 | 60 void ff_init_cabac_states(CABACContext *c, uint8_t const (*lps_range)[4], |
1287 | 61 uint8_t const *mps_state, uint8_t const *lps_state, int state_count); |
62 | |
63 | |
64 static inline void put_cabac_bit(CABACContext *c, int b){ | |
2967 | 65 put_bits(&c->pb, 1, b); |
66 for(;c->outstanding_count; c->outstanding_count--){ | |
1287 | 67 put_bits(&c->pb, 1, 1-b); |
68 } | |
69 } | |
70 | |
71 static inline void renorm_cabac_encoder(CABACContext *c){ | |
72 while(c->range < 0x100){ | |
73 //FIXME optimize | |
74 if(c->low<0x100){ | |
75 put_cabac_bit(c, 0); | |
76 }else if(c->low<0x200){ | |
77 c->outstanding_count++; | |
78 c->low -= 0x100; | |
79 }else{ | |
80 put_cabac_bit(c, 1); | |
81 c->low -= 0x200; | |
82 } | |
2967 | 83 |
1287 | 84 c->range+= c->range; |
85 c->low += c->low; | |
86 } | |
87 } | |
88 | |
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
|
89 static void put_cabac(CABACContext *c, uint8_t * const state, int bit){ |
2323 | 90 int RangeLPS= c->lps_range[*state][c->range>>6]; |
2967 | 91 |
1287 | 92 if(bit == ((*state)&1)){ |
93 c->range -= RangeLPS; | |
94 *state= c->mps_state[*state]; | |
95 }else{ | |
96 c->low += c->range - RangeLPS; | |
97 c->range = RangeLPS; | |
98 *state= c->lps_state[*state]; | |
99 } | |
2967 | 100 |
1287 | 101 renorm_cabac_encoder(c); |
102 | |
103 #ifdef STRICT_LIMITS | |
104 c->symCount++; | |
105 #endif | |
106 } | |
107 | |
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
|
108 static void put_cabac_static(CABACContext *c, int RangeLPS, int bit){ |
1287 | 109 assert(c->range > RangeLPS); |
110 | |
111 if(!bit){ | |
112 c->range -= RangeLPS; | |
113 }else{ | |
114 c->low += c->range - RangeLPS; | |
115 c->range = RangeLPS; | |
116 } | |
117 | |
118 renorm_cabac_encoder(c); | |
119 | |
120 #ifdef STRICT_LIMITS | |
121 c->symCount++; | |
122 #endif | |
123 } | |
124 | |
1290 | 125 /** |
126 * @param bit 0 -> write zero bit, !=0 write one bit | |
127 */ | |
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
|
128 static void put_cabac_bypass(CABACContext *c, int bit){ |
1287 | 129 c->low += c->low; |
130 | |
131 if(bit){ | |
132 c->low += c->range; | |
133 } | |
134 //FIXME optimize | |
135 if(c->low<0x200){ | |
136 put_cabac_bit(c, 0); | |
137 }else if(c->low<0x400){ | |
138 c->outstanding_count++; | |
139 c->low -= 0x200; | |
140 }else{ | |
141 put_cabac_bit(c, 1); | |
142 c->low -= 0x400; | |
143 } | |
2967 | 144 |
1287 | 145 #ifdef STRICT_LIMITS |
146 c->symCount++; | |
147 #endif | |
148 } | |
149 | |
1300
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 * |
e18667d1e94d
FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents:
1298
diff
changeset
|
152 * @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
|
153 */ |
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
|
154 static int put_cabac_terminate(CABACContext *c, int bit){ |
1287 | 155 c->range -= 2; |
156 | |
157 if(!bit){ | |
158 renorm_cabac_encoder(c); | |
159 }else{ | |
160 c->low += c->range; | |
161 c->range= 2; | |
2967 | 162 |
1287 | 163 renorm_cabac_encoder(c); |
164 | |
165 assert(c->low <= 0x1FF); | |
166 put_cabac_bit(c, c->low>>9); | |
167 put_bits(&c->pb, 2, ((c->low>>7)&3)|1); | |
2967 | 168 |
1287 | 169 flush_put_bits(&c->pb); //FIXME FIXME FIXME XXX wrong |
170 } | |
2967 | 171 |
1287 | 172 #ifdef STRICT_LIMITS |
173 c->symCount++; | |
174 #endif | |
1300
e18667d1e94d
FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents:
1298
diff
changeset
|
175 |
1787 | 176 return (put_bits_count(&c->pb)+7)>>3; |
1287 | 177 } |
178 | |
1290 | 179 /** |
180 * put (truncated) unary binarization. | |
181 */ | |
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
|
182 static void put_cabac_u(CABACContext *c, uint8_t * state, int v, int max, int max_index, int truncated){ |
1290 | 183 int i; |
2967 | 184 |
1290 | 185 assert(v <= max); |
2967 | 186 |
1290 | 187 #if 1 |
188 for(i=0; i<v; i++){ | |
189 put_cabac(c, state, 1); | |
190 if(i < max_index) state++; | |
191 } | |
192 if(truncated==0 || v<max) | |
193 put_cabac(c, state, 0); | |
194 #else | |
195 if(v <= max_index){ | |
196 for(i=0; i<v; i++){ | |
197 put_cabac(c, state+i, 1); | |
198 } | |
199 if(truncated==0 || v<max) | |
200 put_cabac(c, state+i, 0); | |
201 }else{ | |
202 for(i=0; i<=max_index; i++){ | |
203 put_cabac(c, state+i, 1); | |
204 } | |
205 for(; i<v; i++){ | |
206 put_cabac(c, state+max_index, 1); | |
207 } | |
208 if(truncated==0 || v<max) | |
209 put_cabac(c, state+max_index, 0); | |
210 } | |
211 #endif | |
212 } | |
213 | |
214 /** | |
215 * put unary exp golomb k-th order binarization. | |
216 */ | |
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
|
217 static void put_cabac_ueg(CABACContext *c, uint8_t * state, int v, int max, int is_signed, int k, int max_index){ |
1290 | 218 int i; |
2967 | 219 |
1290 | 220 if(v==0) |
221 put_cabac(c, state, 0); | |
222 else{ | |
1298 | 223 const int sign= v < 0; |
2967 | 224 |
1298 | 225 if(is_signed) v= ABS(v); |
2967 | 226 |
1290 | 227 if(v<max){ |
228 for(i=0; i<v; i++){ | |
229 put_cabac(c, state, 1); | |
230 if(i < max_index) state++; | |
231 } | |
232 | |
233 put_cabac(c, state, 0); | |
234 }else{ | |
235 int m= 1<<k; | |
236 | |
237 for(i=0; i<max; i++){ | |
238 put_cabac(c, state, 1); | |
239 if(i < max_index) state++; | |
240 } | |
241 | |
242 v -= max; | |
243 while(v >= m){ //FIXME optimize | |
244 put_cabac_bypass(c, 1); | |
245 v-= m; | |
246 m+= m; | |
247 } | |
248 put_cabac_bypass(c, 0); | |
249 while(m>>=1){ | |
250 put_cabac_bypass(c, v&m); | |
251 } | |
252 } | |
253 | |
254 if(is_signed) | |
255 put_cabac_bypass(c, sign); | |
256 } | |
257 } | |
258 | |
2323 | 259 static void refill(CABACContext *c){ |
260 #if CABAC_BITS == 16 | |
3946 | 261 c->low+= (c->bytestream[0]<<9) + (c->bytestream[1]<<1); |
2323 | 262 #else |
263 c->low+= c->bytestream[0]<<1; | |
264 #endif | |
265 c->low -= CABAC_MASK; | |
266 c->bytestream+= CABAC_BITS/8; | |
267 } | |
268 | |
269 static void refill2(CABACContext *c){ | |
270 int i, x; | |
271 | |
272 x= c->low ^ (c->low-1); | |
3964 | 273 i= 7 - ff_h264_norm_shift[x>>(CABAC_BITS+1)]; |
2323 | 274 |
275 x= -CABAC_MASK; | |
2967 | 276 |
2323 | 277 #if CABAC_BITS == 16 |
278 x+= (c->bytestream[0]<<9) + (c->bytestream[1]<<1); | |
279 #else | |
280 x+= c->bytestream[0]<<1; | |
281 #endif | |
2967 | 282 |
2323 | 283 c->low += x<<i; |
284 c->bytestream+= CABAC_BITS/8; | |
285 } | |
286 | |
1287 | 287 static inline void renorm_cabac_decoder(CABACContext *c){ |
2323 | 288 while(c->range < (0x200 << CABAC_BITS)){ |
1287 | 289 c->range+= c->range; |
290 c->low+= c->low; | |
2323 | 291 if(!(c->low & CABAC_MASK)) |
292 refill(c); | |
1287 | 293 } |
294 } | |
295 | |
2323 | 296 static inline void renorm_cabac_decoder_once(CABACContext *c){ |
3951 | 297 #ifdef ARCH_X86_DISABLED |
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
298 int temp; |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
299 #if 0 |
3950
900d21b85dd6
renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents:
3948
diff
changeset
|
300 //P3:683 athlon:475 |
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
301 asm( |
3948
3edbf131ee44
refill cabac variables in 16bit steps, 3% faster get_cabac()
michael
parents:
3947
diff
changeset
|
302 "lea -0x2000000(%0), %2 \n\t" |
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
303 "shr $31, %2 \n\t" //FIXME 31->63 for x86-64 |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
304 "shl %%cl, %0 \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
305 "shl %%cl, %1 \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
306 : "+r"(c->range), "+r"(c->low), "+c"(temp) |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
307 ); |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
308 #elif 0 |
3950
900d21b85dd6
renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents:
3948
diff
changeset
|
309 //P3:680 athlon:474 |
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
310 asm( |
3948
3edbf131ee44
refill cabac variables in 16bit steps, 3% faster get_cabac()
michael
parents:
3947
diff
changeset
|
311 "cmp $0x2000000, %0 \n\t" |
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
312 "setb %%cl \n\t" //FIXME 31->63 for x86-64 |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
313 "shl %%cl, %0 \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
314 "shl %%cl, %1 \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
315 : "+r"(c->range), "+r"(c->low), "+c"(temp) |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
316 ); |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
317 #elif 1 |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
318 int temp2; |
3950
900d21b85dd6
renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents:
3948
diff
changeset
|
319 //P3:665 athlon:517 |
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
320 asm( |
3948
3edbf131ee44
refill cabac variables in 16bit steps, 3% faster get_cabac()
michael
parents:
3947
diff
changeset
|
321 "lea -0x2000000(%0), %%eax \n\t" |
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
322 "cdq \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
323 "mov %0, %%eax \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
324 "and %%edx, %0 \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
325 "and %1, %%edx \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
326 "add %%eax, %0 \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
327 "add %%edx, %1 \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
328 : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2) |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
329 ); |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
330 #elif 0 |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
331 int temp2; |
3950
900d21b85dd6
renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents:
3948
diff
changeset
|
332 //P3:673 athlon:509 |
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
333 asm( |
3948
3edbf131ee44
refill cabac variables in 16bit steps, 3% faster get_cabac()
michael
parents:
3947
diff
changeset
|
334 "cmp $0x2000000, %0 \n\t" |
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
335 "sbb %%edx, %%edx \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
336 "mov %0, %%eax \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
337 "and %%edx, %0 \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
338 "and %1, %%edx \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
339 "add %%eax, %0 \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
340 "add %%edx, %1 \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
341 : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2) |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
342 ); |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
343 #else |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
344 int temp2; |
3950
900d21b85dd6
renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents:
3948
diff
changeset
|
345 //P3:677 athlon:511 |
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
346 asm( |
3948
3edbf131ee44
refill cabac variables in 16bit steps, 3% faster get_cabac()
michael
parents:
3947
diff
changeset
|
347 "cmp $0x2000000, %0 \n\t" |
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
348 "lea (%0, %0), %%eax \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
349 "lea (%1, %1), %%edx \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
350 "cmovb %%eax, %0 \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
351 "cmovb %%edx, %1 \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
352 : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2) |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
353 ); |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
354 #endif |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
355 #else |
3950
900d21b85dd6
renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents:
3948
diff
changeset
|
356 //P3:675 athlon:476 |
3642 | 357 int shift= (uint32_t)(c->range - (0x200 << CABAC_BITS))>>31; |
358 c->range<<= shift; | |
359 c->low <<= shift; | |
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
360 #endif |
2323 | 361 if(!(c->low & CABAC_MASK)) |
362 refill(c); | |
363 } | |
364 | |
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
|
365 static int get_cabac(CABACContext *c, uint8_t * const state){ |
3642 | 366 //FIXME gcc generates duplicate load/stores for c->low and c->range |
3969
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
367 #ifdef ARCH_X86 |
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 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
|
369 |
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 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
|
371 #define RANGE "4" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
372 #define LPS_RANGE "12" |
3976
27e90123b346
reverse remainder of the failed attempt to optimize *state=c->mps_state[s]
michael
parents:
3975
diff
changeset
|
373 #define LPS_STATE "12+2*65*4" |
27e90123b346
reverse remainder of the failed attempt to optimize *state=c->mps_state[s]
michael
parents:
3975
diff
changeset
|
374 #define MPS_STATE "12+2*65*4+2*64" |
27e90123b346
reverse remainder of the failed attempt to optimize *state=c->mps_state[s]
michael
parents:
3975
diff
changeset
|
375 #define BYTESTART "12+2*65*4+4*64" |
27e90123b346
reverse remainder of the failed attempt to optimize *state=c->mps_state[s]
michael
parents:
3975
diff
changeset
|
376 #define BYTE "16+2*65*4+4*64" |
27e90123b346
reverse remainder of the failed attempt to optimize *state=c->mps_state[s]
michael
parents:
3975
diff
changeset
|
377 #define BYTEEND "20+2*65*4+4*64" |
3984 | 378 #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
|
379 asm volatile( |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
380 "movzbl (%1), %%eax \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
381 "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
|
382 "movl "RANGE "(%2), %%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
|
383 "shrl $23, %%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
|
384 "leal "LPS_RANGE"(%2, %%eax, 4), %%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
|
385 "movzbl (%%ebx, %%esi), %%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
|
386 "shll $17, %%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
|
387 "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
|
388 //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
|
389 "subl %%esi, %%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
|
390 "cmpl %%edx, %%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
|
391 " ja 1f \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
|
392 "cmp $0x2000000, %%edx \n\t" //FIXME avoidable |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
393 "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
|
394 "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
|
395 "shl %%cl, %%ebx \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
|
396 "movzbl "MPS_STATE"(%2, %%eax), %%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
|
397 "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
|
398 //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
|
399 "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
|
400 " 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
|
401 "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
|
402 "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
|
403 "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
|
404 "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
|
405 "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
|
406 "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
|
407 "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
|
408 "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
|
409 "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
|
410 "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
|
411 //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
|
412 "subl %%edx, %%ebx \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
413 "movl %%esi, %%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
|
414 "shr $19, %%esi \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
|
415 "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
|
416 "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
|
417 "shll %%cl, %%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
|
418 "movzbl "LPS_STATE"(%2, %%eax), %%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
|
419 "movb %%cl, (%1) \n\t" |
3978 | 420 "addl $1, %%eax \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 "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
|
422 " 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
|
423 |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
424 "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
|
425 "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
|
426 "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
|
427 "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
|
428 "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
|
429 "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
|
430 "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
|
431 |
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 "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
|
433 "xorl %%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
|
434 "shrl $17, %%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
|
435 "movzbl " MANGLE(ff_h264_norm_shift) "(%%ecx), %%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
|
436 "neg %%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
|
437 "add $7, %%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
|
438 |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
439 "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
|
440 "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
|
441 "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
|
442 "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
|
443 "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
|
444 :"=&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
|
445 :"r"(state), "r"(c) |
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 : "%ecx", "%ebx", "%edx", "%esi" |
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 ); |
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
|
448 bit&=1; |
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
|
449 #else |
3975 | 450 asm volatile( |
451 "movzbl (%1), %%eax \n\t" | |
452 "movl "RANGE "(%2), %%ebx \n\t" | |
453 "movl "RANGE "(%2), %%edx \n\t" | |
454 "shrl $23, %%ebx \n\t" | |
455 "leal "LPS_RANGE"(%2, %%eax, 4), %%esi \n\t" | |
456 "movzbl (%%ebx, %%esi), %%esi \n\t" | |
457 "shll $17, %%esi \n\t" | |
458 "movl "LOW "(%2), %%ebx \n\t" | |
459 //eax:state ebx:low, edx:range, esi:RangeLPS | |
460 "subl %%esi, %%edx \n\t" | |
3980 | 461 #ifdef CMOV_IS_FAST //FIXME actually define this somewhere |
462 "cmpl %%ebx, %%edx \n\t" | |
463 "cmova %%edx, %%esi \n\t" | |
464 "sbbl %%ecx, %%ecx \n\t" | |
465 "andl %%ecx, %%edx \n\t" | |
466 "subl %%edx, %%ebx \n\t" | |
467 "xorl %%ecx, %%eax \n\t" | |
468 #else | |
3975 | 469 "movl %%edx, %%ecx \n\t" |
470 "subl %%ebx, %%edx \n\t" | |
471 "sarl $31, %%edx \n\t" //lps_mask | |
472 "subl %%ecx, %%esi \n\t" //RangeLPS - range | |
473 "andl %%edx, %%esi \n\t" //(RangeLPS - range)&lps_mask | |
474 "addl %%ecx, %%esi \n\t" //new range | |
475 "andl %%edx, %%ecx \n\t" | |
476 "subl %%ecx, %%ebx \n\t" | |
3980 | 477 "xorl %%edx, %%eax \n\t" |
478 #endif | |
3975 | 479 |
480 //eax:state ebx:low edx:mask esi:range | |
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
|
481 "movzbl "MPS_STATE"(%2, %%eax), %%ecx \n\t" |
3975 | 482 "movb %%cl, (%1) \n\t" |
483 | |
484 "movl %%esi, %%edx \n\t" | |
485 //eax:bit ebx:low edx:range esi:range | |
486 | |
487 "shr $19, %%esi \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
|
488 "movzbl " MANGLE(ff_h264_norm_shift) "(%%esi), %%ecx \n\t" |
3975 | 489 "shll %%cl, %%ebx \n\t" |
490 "shll %%cl, %%edx \n\t" | |
491 "test %%bx, %%bx \n\t" | |
492 " jnz 1f \n\t" | |
493 | |
494 "movl "BYTE "(%2), %%ecx \n\t" | |
495 "movzwl (%%ecx), %%esi \n\t" | |
496 "bswap %%esi \n\t" | |
497 "shrl $15, %%esi \n\t" | |
498 "subl $0xFFFF, %%esi \n\t" | |
499 "addl $2, %%ecx \n\t" | |
500 "movl %%ecx, "BYTE "(%2) \n\t" | |
501 | |
502 "leal -1(%%ebx), %%ecx \n\t" | |
503 "xorl %%ebx, %%ecx \n\t" | |
504 "shrl $17, %%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
|
505 "movzbl " MANGLE(ff_h264_norm_shift) "(%%ecx), %%ecx \n\t" |
3975 | 506 "neg %%cl \n\t" |
507 "add $7, %%cl \n\t" | |
508 | |
509 "shll %%cl , %%esi \n\t" | |
510 "addl %%esi, %%ebx \n\t" | |
511 "1: \n\t" | |
512 "movl %%edx, "RANGE "(%2) \n\t" | |
513 "movl %%ebx, "LOW "(%2) \n\t" | |
514 :"=&a"(bit) | |
515 :"r"(state), "r"(c) | |
516 : "%ecx", "%ebx", "%edx", "%esi" | |
517 ); | |
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
|
518 bit&=1; |
3975 | 519 #endif |
520 #else | |
3642 | 521 int s = *state; |
522 int RangeLPS= c->lps_range[s][c->range>>(CABAC_BITS+7)]<<(CABAC_BITS+1); | |
2522
e25782262d7d
kill warnings patch by (M«©ns Rullg«©rd <mru inprovide com>)
michael
parents:
2323
diff
changeset
|
523 int bit, lps_mask attribute_unused; |
2967 | 524 |
1287 | 525 c->range -= RangeLPS; |
3984 | 526 #ifndef BRANCHLESS_CABAC_DECODER |
1287 | 527 if(c->low < c->range){ |
3642 | 528 bit= s&1; |
3971
e8a6d5c1ab0b
drop failed attempt to optimize *state= c->mps_state[s];
michael
parents:
3970
diff
changeset
|
529 *state= c->mps_state[s]; |
2323 | 530 renorm_cabac_decoder_once(c); |
1287 | 531 }else{ |
3964 | 532 bit= ff_h264_norm_shift[RangeLPS>>19]; |
1287 | 533 c->low -= c->range; |
3642 | 534 *state= c->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
|
535 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
|
536 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
|
537 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
|
538 |
2323 | 539 if(!(c->low & 0xFFFF)){ |
540 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
|
541 } |
1287 | 542 } |
2323 | 543 #else |
544 lps_mask= (c->range - c->low)>>31; | |
2967 | 545 |
2323 | 546 c->low -= c->range & lps_mask; |
547 c->range += (RangeLPS - c->range) & lps_mask; | |
2967 | 548 |
3974 | 549 s^=lps_mask; |
550 *state= c->mps_state[s]; | |
551 bit= s&1; | |
2967 | 552 |
3970 | 553 lps_mask= ff_h264_norm_shift[c->range>>(CABAC_BITS+3)]; |
2323 | 554 c->range<<= lps_mask; |
555 c->low <<= lps_mask; | |
556 if(!(c->low & CABAC_MASK)) | |
557 refill2(c); | |
558 #endif | |
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
|
559 #endif |
2967 | 560 return bit; |
1287 | 561 } |
562 | |
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
|
563 static int get_cabac_bypass(CABACContext *c){ |
1287 | 564 c->low += c->low; |
565 | |
2323 | 566 if(!(c->low & CABAC_MASK)) |
567 refill(c); | |
2967 | 568 |
1287 | 569 if(c->low < c->range){ |
570 return 0; | |
571 }else{ | |
572 c->low -= c->range; | |
573 return 1; | |
574 } | |
575 } | |
576 | |
1300
e18667d1e94d
FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents:
1298
diff
changeset
|
577 /** |
e18667d1e94d
FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents:
1298
diff
changeset
|
578 * |
e18667d1e94d
FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents:
1298
diff
changeset
|
579 * @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
|
580 */ |
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
|
581 static int get_cabac_terminate(CABACContext *c){ |
2323 | 582 c->range -= 4<<CABAC_BITS; |
1287 | 583 if(c->low < c->range){ |
2323 | 584 renorm_cabac_decoder_once(c); |
1287 | 585 return 0; |
586 }else{ | |
1300
e18667d1e94d
FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents:
1298
diff
changeset
|
587 return c->bytestream - c->bytestream_start; |
2967 | 588 } |
1287 | 589 } |
590 | |
1290 | 591 /** |
592 * get (truncated) unnary binarization. | |
593 */ | |
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
|
594 static int get_cabac_u(CABACContext *c, uint8_t * state, int max, int max_index, int truncated){ |
1290 | 595 int i; |
2967 | 596 |
597 for(i=0; i<max; i++){ | |
1290 | 598 if(get_cabac(c, state)==0) |
599 return i; | |
2967 | 600 |
1290 | 601 if(i< max_index) state++; |
602 } | |
603 | |
604 return truncated ? max : -1; | |
605 } | |
606 | |
607 /** | |
608 * get unary exp golomb k-th order binarization. | |
609 */ | |
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
|
610 static int get_cabac_ueg(CABACContext *c, uint8_t * state, int max, int is_signed, int k, int max_index){ |
1290 | 611 int i, v; |
612 int m= 1<<k; | |
2967 | 613 |
614 if(get_cabac(c, state)==0) | |
1290 | 615 return 0; |
2967 | 616 |
1290 | 617 if(0 < max_index) state++; |
2967 | 618 |
619 for(i=1; i<max; i++){ | |
1290 | 620 if(get_cabac(c, state)==0){ |
621 if(is_signed && get_cabac_bypass(c)){ | |
622 return -i; | |
623 }else | |
624 return i; | |
625 } | |
626 | |
627 if(i < max_index) state++; | |
628 } | |
2967 | 629 |
1290 | 630 while(get_cabac_bypass(c)){ |
631 i+= m; | |
632 m+= m; | |
633 } | |
2967 | 634 |
1290 | 635 v=0; |
636 while(m>>=1){ | |
637 v+= v + get_cabac_bypass(c); | |
638 } | |
639 i += v; | |
640 | |
641 if(is_signed && get_cabac_bypass(c)){ | |
642 return -i; | |
643 }else | |
644 return i; | |
645 } |