Mercurial > libavcodec.hg
annotate cabac.h @ 4050:2c79a8281cb6 libavcodec
Protect code that uses CMOV instructions with HAVE_CMOV,
Make configure set CMOV_IS_FAST on arches on which cmov has a low latency
(typically non-Netburst based processor)
author | gpoirier |
---|---|
date | Fri, 20 Oct 2006 17:53:19 +0000 |
parents | 8c1a5ed03a00 |
children | 19f07b651d79 |
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 |
4039 | 35 //#define ARCH_X86_DISABLED 1 |
2323 | 36 |
1287 | 37 typedef struct CABACContext{ |
38 int low; | |
39 int range; | |
40 int outstanding_count; | |
41 #ifdef STRICT_LIMITS | |
42 int symCount; | |
43 #endif | |
2024
f65d87bfdd5a
some of the warning fixes by (Michael Roitzsch <mroi at users dot sourceforge dot net>)
michael
parents:
1787
diff
changeset
|
44 const uint8_t *bytestream_start; |
f65d87bfdd5a
some of the warning fixes by (Michael Roitzsch <mroi at users dot sourceforge dot net>)
michael
parents:
1787
diff
changeset
|
45 const uint8_t *bytestream; |
2116 | 46 const uint8_t *bytestream_end; |
1287 | 47 PutBitContext pb; |
48 }CABACContext; | |
49 | |
4014
b2582438effe
dehack *ps_state indexing in the branchless decoder
michael
parents:
4012
diff
changeset
|
50 extern uint8_t ff_h264_mlps_state[4*64]; |
4039 | 51 extern uint8_t ff_h264_lps_range[4*2*64]; ///< rangeTabLPS |
3993
8b7c59b7af01
make state transition tables global as they are constant and the code is slightly faster that way
michael
parents:
3992
diff
changeset
|
52 extern uint8_t ff_h264_mps_state[2*64]; ///< transIdxMPS |
8b7c59b7af01
make state transition tables global as they are constant and the code is slightly faster that way
michael
parents:
3992
diff
changeset
|
53 extern uint8_t ff_h264_lps_state[2*64]; ///< transIdxLPS |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
54 extern const uint8_t ff_h264_norm_shift[512]; |
2323 | 55 |
1287 | 56 |
57 void ff_init_cabac_encoder(CABACContext *c, uint8_t *buf, int buf_size); | |
2024
f65d87bfdd5a
some of the warning fixes by (Michael Roitzsch <mroi at users dot sourceforge dot net>)
michael
parents:
1787
diff
changeset
|
58 void ff_init_cabac_decoder(CABACContext *c, const uint8_t *buf, int buf_size); |
3993
8b7c59b7af01
make state transition tables global as they are constant and the code is slightly faster that way
michael
parents:
3992
diff
changeset
|
59 void ff_init_cabac_states(CABACContext *c); |
1287 | 60 |
61 | |
62 static inline void put_cabac_bit(CABACContext *c, int b){ | |
2967 | 63 put_bits(&c->pb, 1, b); |
64 for(;c->outstanding_count; c->outstanding_count--){ | |
1287 | 65 put_bits(&c->pb, 1, 1-b); |
66 } | |
67 } | |
68 | |
69 static inline void renorm_cabac_encoder(CABACContext *c){ | |
70 while(c->range < 0x100){ | |
71 //FIXME optimize | |
72 if(c->low<0x100){ | |
73 put_cabac_bit(c, 0); | |
74 }else if(c->low<0x200){ | |
75 c->outstanding_count++; | |
76 c->low -= 0x100; | |
77 }else{ | |
78 put_cabac_bit(c, 1); | |
79 c->low -= 0x200; | |
80 } | |
2967 | 81 |
1287 | 82 c->range+= c->range; |
83 c->low += c->low; | |
84 } | |
85 } | |
86 | |
3928
987fffdf6ae7
don't try to inline cabac functions. gcc ignored the hint anyway, and forcing it would make h264 slower.
lorenm
parents:
3642
diff
changeset
|
87 static void put_cabac(CABACContext *c, uint8_t * const state, int bit){ |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
88 int RangeLPS= ff_h264_lps_range[2*(c->range&0xC0) + *state]; |
2967 | 89 |
1287 | 90 if(bit == ((*state)&1)){ |
91 c->range -= RangeLPS; | |
3993
8b7c59b7af01
make state transition tables global as they are constant and the code is slightly faster that way
michael
parents:
3992
diff
changeset
|
92 *state= ff_h264_mps_state[*state]; |
1287 | 93 }else{ |
94 c->low += c->range - RangeLPS; | |
95 c->range = RangeLPS; | |
3993
8b7c59b7af01
make state transition tables global as they are constant and the code is slightly faster that way
michael
parents:
3992
diff
changeset
|
96 *state= ff_h264_lps_state[*state]; |
1287 | 97 } |
2967 | 98 |
1287 | 99 renorm_cabac_encoder(c); |
100 | |
101 #ifdef STRICT_LIMITS | |
102 c->symCount++; | |
103 #endif | |
104 } | |
105 | |
3928
987fffdf6ae7
don't try to inline cabac functions. gcc ignored the hint anyway, and forcing it would make h264 slower.
lorenm
parents:
3642
diff
changeset
|
106 static void put_cabac_static(CABACContext *c, int RangeLPS, int bit){ |
1287 | 107 assert(c->range > RangeLPS); |
108 | |
109 if(!bit){ | |
110 c->range -= RangeLPS; | |
111 }else{ | |
112 c->low += c->range - RangeLPS; | |
113 c->range = RangeLPS; | |
114 } | |
115 | |
116 renorm_cabac_encoder(c); | |
117 | |
118 #ifdef STRICT_LIMITS | |
119 c->symCount++; | |
120 #endif | |
121 } | |
122 | |
1290 | 123 /** |
124 * @param bit 0 -> write zero bit, !=0 write one bit | |
125 */ | |
3928
987fffdf6ae7
don't try to inline cabac functions. gcc ignored the hint anyway, and forcing it would make h264 slower.
lorenm
parents:
3642
diff
changeset
|
126 static void put_cabac_bypass(CABACContext *c, int bit){ |
1287 | 127 c->low += c->low; |
128 | |
129 if(bit){ | |
130 c->low += c->range; | |
131 } | |
132 //FIXME optimize | |
133 if(c->low<0x200){ | |
134 put_cabac_bit(c, 0); | |
135 }else if(c->low<0x400){ | |
136 c->outstanding_count++; | |
137 c->low -= 0x200; | |
138 }else{ | |
139 put_cabac_bit(c, 1); | |
140 c->low -= 0x400; | |
141 } | |
2967 | 142 |
1287 | 143 #ifdef STRICT_LIMITS |
144 c->symCount++; | |
145 #endif | |
146 } | |
147 | |
1300
e18667d1e94d
FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents:
1298
diff
changeset
|
148 /** |
e18667d1e94d
FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents:
1298
diff
changeset
|
149 * |
e18667d1e94d
FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents:
1298
diff
changeset
|
150 * @return the number of bytes written |
e18667d1e94d
FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents:
1298
diff
changeset
|
151 */ |
3928
987fffdf6ae7
don't try to inline cabac functions. gcc ignored the hint anyway, and forcing it would make h264 slower.
lorenm
parents:
3642
diff
changeset
|
152 static int put_cabac_terminate(CABACContext *c, int bit){ |
1287 | 153 c->range -= 2; |
154 | |
155 if(!bit){ | |
156 renorm_cabac_encoder(c); | |
157 }else{ | |
158 c->low += c->range; | |
159 c->range= 2; | |
2967 | 160 |
1287 | 161 renorm_cabac_encoder(c); |
162 | |
163 assert(c->low <= 0x1FF); | |
164 put_cabac_bit(c, c->low>>9); | |
165 put_bits(&c->pb, 2, ((c->low>>7)&3)|1); | |
2967 | 166 |
1287 | 167 flush_put_bits(&c->pb); //FIXME FIXME FIXME XXX wrong |
168 } | |
2967 | 169 |
1287 | 170 #ifdef STRICT_LIMITS |
171 c->symCount++; | |
172 #endif | |
1300
e18667d1e94d
FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents:
1298
diff
changeset
|
173 |
1787 | 174 return (put_bits_count(&c->pb)+7)>>3; |
1287 | 175 } |
176 | |
1290 | 177 /** |
178 * put (truncated) unary binarization. | |
179 */ | |
3928
987fffdf6ae7
don't try to inline cabac functions. gcc ignored the hint anyway, and forcing it would make h264 slower.
lorenm
parents:
3642
diff
changeset
|
180 static void put_cabac_u(CABACContext *c, uint8_t * state, int v, int max, int max_index, int truncated){ |
1290 | 181 int i; |
2967 | 182 |
1290 | 183 assert(v <= max); |
2967 | 184 |
1290 | 185 #if 1 |
186 for(i=0; i<v; i++){ | |
187 put_cabac(c, state, 1); | |
188 if(i < max_index) state++; | |
189 } | |
190 if(truncated==0 || v<max) | |
191 put_cabac(c, state, 0); | |
192 #else | |
193 if(v <= max_index){ | |
194 for(i=0; i<v; i++){ | |
195 put_cabac(c, state+i, 1); | |
196 } | |
197 if(truncated==0 || v<max) | |
198 put_cabac(c, state+i, 0); | |
199 }else{ | |
200 for(i=0; i<=max_index; i++){ | |
201 put_cabac(c, state+i, 1); | |
202 } | |
203 for(; i<v; i++){ | |
204 put_cabac(c, state+max_index, 1); | |
205 } | |
206 if(truncated==0 || v<max) | |
207 put_cabac(c, state+max_index, 0); | |
208 } | |
209 #endif | |
210 } | |
211 | |
212 /** | |
213 * put unary exp golomb k-th order binarization. | |
214 */ | |
3928
987fffdf6ae7
don't try to inline cabac functions. gcc ignored the hint anyway, and forcing it would make h264 slower.
lorenm
parents:
3642
diff
changeset
|
215 static void put_cabac_ueg(CABACContext *c, uint8_t * state, int v, int max, int is_signed, int k, int max_index){ |
1290 | 216 int i; |
2967 | 217 |
1290 | 218 if(v==0) |
219 put_cabac(c, state, 0); | |
220 else{ | |
1298 | 221 const int sign= v < 0; |
2967 | 222 |
4001 | 223 if(is_signed) v= FFABS(v); |
2967 | 224 |
1290 | 225 if(v<max){ |
226 for(i=0; i<v; i++){ | |
227 put_cabac(c, state, 1); | |
228 if(i < max_index) state++; | |
229 } | |
230 | |
231 put_cabac(c, state, 0); | |
232 }else{ | |
233 int m= 1<<k; | |
234 | |
235 for(i=0; i<max; i++){ | |
236 put_cabac(c, state, 1); | |
237 if(i < max_index) state++; | |
238 } | |
239 | |
240 v -= max; | |
241 while(v >= m){ //FIXME optimize | |
242 put_cabac_bypass(c, 1); | |
243 v-= m; | |
244 m+= m; | |
245 } | |
246 put_cabac_bypass(c, 0); | |
247 while(m>>=1){ | |
248 put_cabac_bypass(c, v&m); | |
249 } | |
250 } | |
251 | |
252 if(is_signed) | |
253 put_cabac_bypass(c, sign); | |
254 } | |
255 } | |
256 | |
2323 | 257 static void refill(CABACContext *c){ |
258 #if CABAC_BITS == 16 | |
3946 | 259 c->low+= (c->bytestream[0]<<9) + (c->bytestream[1]<<1); |
2323 | 260 #else |
261 c->low+= c->bytestream[0]<<1; | |
262 #endif | |
263 c->low -= CABAC_MASK; | |
264 c->bytestream+= CABAC_BITS/8; | |
265 } | |
266 | |
267 static void refill2(CABACContext *c){ | |
268 int i, x; | |
269 | |
270 x= c->low ^ (c->low-1); | |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
271 i= 7 - ff_h264_norm_shift[x>>(CABAC_BITS-1)]; |
2323 | 272 |
273 x= -CABAC_MASK; | |
2967 | 274 |
2323 | 275 #if CABAC_BITS == 16 |
276 x+= (c->bytestream[0]<<9) + (c->bytestream[1]<<1); | |
277 #else | |
278 x+= c->bytestream[0]<<1; | |
279 #endif | |
2967 | 280 |
2323 | 281 c->low += x<<i; |
282 c->bytestream+= CABAC_BITS/8; | |
283 } | |
284 | |
1287 | 285 static inline void renorm_cabac_decoder(CABACContext *c){ |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
286 while(c->range < 0x100){ |
1287 | 287 c->range+= c->range; |
288 c->low+= c->low; | |
2323 | 289 if(!(c->low & CABAC_MASK)) |
290 refill(c); | |
1287 | 291 } |
292 } | |
293 | |
2323 | 294 static inline void renorm_cabac_decoder_once(CABACContext *c){ |
3951 | 295 #ifdef ARCH_X86_DISABLED |
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
296 int temp; |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
297 #if 0 |
3950
900d21b85dd6
renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents:
3948
diff
changeset
|
298 //P3:683 athlon:475 |
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
299 asm( |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
300 "lea -0x100(%0), %2 \n\t" |
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
301 "shr $31, %2 \n\t" //FIXME 31->63 for x86-64 |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
302 "shl %%cl, %0 \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
303 "shl %%cl, %1 \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
304 : "+r"(c->range), "+r"(c->low), "+c"(temp) |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
305 ); |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
306 #elif 0 |
3950
900d21b85dd6
renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents:
3948
diff
changeset
|
307 //P3:680 athlon:474 |
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
308 asm( |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
309 "cmp $0x100, %0 \n\t" |
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
310 "setb %%cl \n\t" //FIXME 31->63 for x86-64 |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
311 "shl %%cl, %0 \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
312 "shl %%cl, %1 \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
313 : "+r"(c->range), "+r"(c->low), "+c"(temp) |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
314 ); |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
315 #elif 1 |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
316 int temp2; |
3950
900d21b85dd6
renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents:
3948
diff
changeset
|
317 //P3:665 athlon:517 |
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
318 asm( |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
319 "lea -0x100(%0), %%eax \n\t" |
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
320 "cdq \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
321 "mov %0, %%eax \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
322 "and %%edx, %0 \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
323 "and %1, %%edx \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
324 "add %%eax, %0 \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
325 "add %%edx, %1 \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
326 : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2) |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
327 ); |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
328 #elif 0 |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
329 int temp2; |
3950
900d21b85dd6
renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents:
3948
diff
changeset
|
330 //P3:673 athlon:509 |
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
331 asm( |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
332 "cmp $0x100, %0 \n\t" |
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
333 "sbb %%edx, %%edx \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
334 "mov %0, %%eax \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
335 "and %%edx, %0 \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
336 "and %1, %%edx \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
337 "add %%eax, %0 \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
338 "add %%edx, %1 \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
339 : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2) |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
340 ); |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
341 #else |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
342 int temp2; |
3950
900d21b85dd6
renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents:
3948
diff
changeset
|
343 //P3:677 athlon:511 |
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
344 asm( |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
345 "cmp $0x100, %0 \n\t" |
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
346 "lea (%0, %0), %%eax \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
347 "lea (%1, %1), %%edx \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
348 "cmovb %%eax, %0 \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
349 "cmovb %%edx, %1 \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
350 : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2) |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
351 ); |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
352 #endif |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
353 #else |
3950
900d21b85dd6
renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents:
3948
diff
changeset
|
354 //P3:675 athlon:476 |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
355 int shift= (uint32_t)(c->range - 0x100)>>31; |
3642 | 356 c->range<<= shift; |
357 c->low <<= shift; | |
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
358 #endif |
2323 | 359 if(!(c->low & CABAC_MASK)) |
360 refill(c); | |
361 } | |
362 | |
4008
b636f3d59283
prevent "mb level" get_cabac() calls from being inlined (3% faster decode_mb_cabac() on P3)
michael
parents:
4002
diff
changeset
|
363 static int always_inline get_cabac_inline(CABACContext *c, uint8_t * const state){ |
3642 | 364 //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
|
365 #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
|
366 #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
|
367 #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
|
368 #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
|
369 #define BYTEEND "20" |
4044
5ccdefd60f61
Fix PIC compilation, some defines were under #ifdef !PIC but used
diego
parents:
4043
diff
changeset
|
370 #if defined(ARCH_X86) && !(defined(PIC) && defined(__GNUC__)) |
5ccdefd60f61
Fix PIC compilation, some defines were under #ifdef !PIC but used
diego
parents:
4043
diff
changeset
|
371 int bit; |
5ccdefd60f61
Fix PIC compilation, some defines were under #ifdef !PIC but used
diego
parents:
4043
diff
changeset
|
372 |
3984 | 373 #ifndef BRANCHLESS_CABAC_DECODER |
3969
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
374 asm volatile( |
4035 | 375 "movzbl (%1), %0 \n\t" |
3969
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
376 "movl "RANGE "(%2), %%ebx \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
377 "movl "RANGE "(%2), %%edx \n\t" |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
378 "andl $0xC0, %%ebx \n\t" |
4035 | 379 "movzbl "MANGLE(ff_h264_lps_range)"(%0, %%ebx, 2), %%esi\n\t" |
3969
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
380 "movl "LOW "(%2), %%ebx \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
381 //eax:state ebx:low, edx:range, esi:RangeLPS |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
382 "subl %%esi, %%edx \n\t" |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
383 "movl %%edx, %%ecx \n\t" |
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
384 "shll $17, %%ecx \n\t" |
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
385 "cmpl %%ecx, %%ebx \n\t" |
3969
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
386 " ja 1f \n\t" |
3999
6cbad3675632
slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents:
3996
diff
changeset
|
387 |
6cbad3675632
slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents:
3996
diff
changeset
|
388 #if 1 |
6cbad3675632
slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents:
3996
diff
changeset
|
389 //athlon:4067 P3:4110 |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
390 "lea -0x100(%%edx), %%ecx \n\t" |
3999
6cbad3675632
slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents:
3996
diff
changeset
|
391 "shr $31, %%ecx \n\t" |
6cbad3675632
slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents:
3996
diff
changeset
|
392 "shl %%cl, %%edx \n\t" |
6cbad3675632
slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents:
3996
diff
changeset
|
393 "shl %%cl, %%ebx \n\t" |
6cbad3675632
slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents:
3996
diff
changeset
|
394 #else |
6cbad3675632
slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents:
3996
diff
changeset
|
395 //athlon:4057 P3:4130 |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
396 "cmp $0x100, %%edx \n\t" //FIXME avoidable |
3969
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
397 "setb %%cl \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
398 "shl %%cl, %%edx \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
399 "shl %%cl, %%ebx \n\t" |
3999
6cbad3675632
slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents:
3996
diff
changeset
|
400 #endif |
4035 | 401 "movzbl "MANGLE(ff_h264_mps_state)"(%0), %%ecx \n\t" |
3969
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
402 "movb %%cl, (%1) \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
403 //eax:state ebx:low, edx:range, esi:RangeLPS |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
404 "test %%bx, %%bx \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
405 " jnz 2f \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
406 "movl "BYTE "(%2), %%esi \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
407 "subl $0xFFFF, %%ebx \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
408 "movzwl (%%esi), %%ecx \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
409 "bswap %%ecx \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
410 "shrl $15, %%ecx \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
411 "addl $2, %%esi \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
412 "addl %%ecx, %%ebx \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
413 "movl %%esi, "BYTE "(%2) \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
414 "jmp 2f \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
415 "1: \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
416 //eax:state ebx:low, edx:range, esi:RangeLPS |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
417 "subl %%ecx, %%ebx \n\t" |
3969
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
418 "movl %%esi, %%edx \n\t" |
3979
ce16f66a48ad
reading 8bit mem into a 8bit register needs 2 uops on P4, 8bit->32bit with zero extension needs just 1
michael
parents:
3978
diff
changeset
|
419 "movzbl " MANGLE(ff_h264_norm_shift) "(%%esi), %%ecx \n\t" |
3969
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
420 "shll %%cl, %%ebx \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
421 "shll %%cl, %%edx \n\t" |
4035 | 422 "movzbl "MANGLE(ff_h264_lps_state)"(%0), %%ecx \n\t" |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
423 "movb %%cl, (%1) \n\t" |
4035 | 424 "addl $1, %0 \n\t" |
3969
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
425 "test %%bx, %%bx \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
426 " jnz 2f \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
427 |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
428 "movl "BYTE "(%2), %%ecx \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
429 "movzwl (%%ecx), %%esi \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
430 "bswap %%esi \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
431 "shrl $15, %%esi \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
432 "subl $0xFFFF, %%esi \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
433 "addl $2, %%ecx \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
434 "movl %%ecx, "BYTE "(%2) \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
435 |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
436 "leal -1(%%ebx), %%ecx \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
437 "xorl %%ebx, %%ecx \n\t" |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
438 "shrl $15, %%ecx \n\t" |
3979
ce16f66a48ad
reading 8bit mem into a 8bit register needs 2 uops on P4, 8bit->32bit with zero extension needs just 1
michael
parents:
3978
diff
changeset
|
439 "movzbl " MANGLE(ff_h264_norm_shift) "(%%ecx), %%ecx \n\t" |
3994
2734b228fc87
use ecx instead of cl (no speed change on P3 but might avoid partial register stalls on some cpus)
michael
parents:
3993
diff
changeset
|
440 "neg %%ecx \n\t" |
2734b228fc87
use ecx instead of cl (no speed change on P3 but might avoid partial register stalls on some cpus)
michael
parents:
3993
diff
changeset
|
441 "add $7, %%ecx \n\t" |
3969
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
442 |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
443 "shll %%cl , %%esi \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
444 "addl %%esi, %%ebx \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
445 "2: \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
446 "movl %%edx, "RANGE "(%2) \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
447 "movl %%ebx, "LOW "(%2) \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
448 :"=&a"(bit) //FIXME this is fragile gcc either runs out of registers or misscompiles it (for example if "+a"(bit) or "+m"(*state) is used |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
449 :"r"(state), "r"(c) |
4012
f8c649ac09dd
add "memory" to the clobber list we change memory so we need it, this also fixes some problems with gcc svn
michael
parents:
4008
diff
changeset
|
450 : "%ecx", "%ebx", "%edx", "%esi", "memory" |
3969
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
451 ); |
3982
af16271634c2
moving another bit&1 out, this is as fast as with it in there, but it makes more sense with it outside of the loop
michael
parents:
3981
diff
changeset
|
452 bit&=1; |
4002
ec426fa57dfe
adds some useful comments after some of the #else, #elseif,
gpoirier
parents:
4001
diff
changeset
|
453 #else /* BRANCHLESS_CABAC_DECODER */ |
4046
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
454 |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
455 |
4050
2c79a8281cb6
Protect code that uses CMOV instructions with HAVE_CMOV,
gpoirier
parents:
4049
diff
changeset
|
456 #if defined CMOV_IS_FAST |
4046
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
457 #define BRANCHLESS_GET_CABAC_UPDATE(ret, cabac, statep, low, lowword, range, tmp, tmpbyte)\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
458 "mov "tmp" , %%ecx \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
459 "shl $17 , "tmp" \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
460 "cmp "low" , "tmp" \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
461 "cmova %%ecx , "range" \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
462 "sbb %%ecx , %%ecx \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
463 "and %%ecx , "tmp" \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
464 "sub "tmp" , "low" \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
465 "xor %%ecx , "ret" \n\t" |
4002
ec426fa57dfe
adds some useful comments after some of the #else, #elseif,
gpoirier
parents:
4001
diff
changeset
|
466 #else /* CMOV_IS_FAST */ |
4046
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
467 #define BRANCHLESS_GET_CABAC_UPDATE(ret, cabac, statep, low, lowword, range, tmp, tmpbyte)\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
468 "mov "tmp" , %%ecx \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
469 "shl $17 , "tmp" \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
470 "sub "low" , "tmp" \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
471 "sar $31 , "tmp" \n\t" /*lps_mask*/\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
472 "sub %%ecx , "range" \n\t" /*RangeLPS - range*/\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
473 "and "tmp" , "range" \n\t" /*(RangeLPS - range)&lps_mask*/\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
474 "add %%ecx , "range" \n\t" /*new range*/\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
475 "shl $17 , %%ecx \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
476 "and "tmp" , %%ecx \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
477 "sub %%ecx , "low" \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
478 "xor "tmp" , "ret" \n\t" |
4002
ec426fa57dfe
adds some useful comments after some of the #else, #elseif,
gpoirier
parents:
4001
diff
changeset
|
479 #endif /* CMOV_IS_FAST */ |
3975 | 480 |
481 | |
4046
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
482 #define BRANCHLESS_GET_CABAC(ret, cabac, statep, low, lowword, range, tmp, tmpbyte)\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
483 "movzbl "statep" , "ret" \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
484 "mov "range" , "tmp" \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
485 "and $0xC0 , "range" \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
486 "movzbl "MANGLE(ff_h264_lps_range)"("ret", "range", 2), "range" \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
487 "sub "range" , "tmp" \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
488 BRANCHLESS_GET_CABAC_UPDATE(ret, cabac, statep, low, lowword, range, tmp, tmpbyte)\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
489 "movzbl " MANGLE(ff_h264_norm_shift) "("range"), %%ecx \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
490 "shl %%cl , "range" \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
491 "movzbl "MANGLE(ff_h264_mlps_state)"+128("ret"), "tmp" \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
492 "mov "tmpbyte" , "statep" \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
493 "shl %%cl , "low" \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
494 "test "lowword" , "lowword" \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
495 " jnz 1f \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
496 "mov "BYTE"("cabac"), %%ecx \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
497 "movzwl (%%ecx) , "tmp" \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
498 "bswap "tmp" \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
499 "shr $15 , "tmp" \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
500 "sub $0xFFFF , "tmp" \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
501 "add $2 , %%ecx \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
502 "mov %%ecx , "BYTE "("cabac") \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
503 "lea -1("low") , %%ecx \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
504 "xor "low" , %%ecx \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
505 "shr $15 , %%ecx \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
506 "movzbl " MANGLE(ff_h264_norm_shift) "(%%ecx), %%ecx \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
507 "neg %%ecx \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
508 "add $7 , %%ecx \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
509 "shl %%cl , "tmp" \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
510 "add "tmp" , "low" \n\t"\ |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
511 "1: \n\t" |
3975 | 512 |
4046
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
513 asm volatile( |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
514 "movl "RANGE "(%2), %%esi \n\t" |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
515 "movl "LOW "(%2), %%ebx \n\t" |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
516 BRANCHLESS_GET_CABAC("%0", "%2", "(%1)", "%%ebx", "%%bx", "%%esi", "%%edx", "%%dl") |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
517 "movl %%esi, "RANGE "(%2) \n\t" |
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
518 "movl %%ebx, "LOW "(%2) \n\t" |
3975 | 519 |
520 :"=&a"(bit) | |
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 | 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 | 526 #else /* defined(ARCH_X86) && !(defined(PIC) && defined(__GNUC__)) */ |
3642 | 527 int s = *state; |
4039 | 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 | 530 |
1287 | 531 c->range -= RangeLPS; |
3984 | 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 | 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 | 536 renorm_cabac_decoder_once(c); |
1287 | 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 | 545 if(!(c->low & 0xFFFF)){ |
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 | 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 | 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 | 553 c->range += (RangeLPS - c->range) & lps_mask; |
2967 | 554 |
3974 | 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 | 557 bit= s&1; |
2967 | 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 | 560 c->range<<= lps_mask; |
561 c->low <<= lps_mask; | |
562 if(!(c->low & CABAC_MASK)) | |
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 | 565 #endif /* defined(ARCH_X86) && !(defined(PIC) && defined(__GNUC__)) */ |
2967 | 566 return bit; |
1287 | 567 } |
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 | 578 #if 0 //not faster |
579 int bit; | |
580 asm volatile( | |
581 "movl "RANGE "(%1), %%ebx \n\t" | |
582 "movl "LOW "(%1), %%eax \n\t" | |
583 "shl $17, %%ebx \n\t" | |
584 "add %%eax, %%eax \n\t" | |
585 "sub %%ebx, %%eax \n\t" | |
586 "cdq \n\t" | |
587 "and %%edx, %%ebx \n\t" | |
588 "add %%ebx, %%eax \n\t" | |
589 "test %%ax, %%ax \n\t" | |
590 " jnz 1f \n\t" | |
591 "movl "BYTE "(%1), %%ebx \n\t" | |
592 "subl $0xFFFF, %%eax \n\t" | |
593 "movzwl (%%ebx), %%ecx \n\t" | |
594 "bswap %%ecx \n\t" | |
595 "shrl $15, %%ecx \n\t" | |
596 "addl $2, %%ebx \n\t" | |
597 "addl %%ecx, %%eax \n\t" | |
598 "movl %%ebx, "BYTE "(%1) \n\t" | |
599 "1: \n\t" | |
600 "movl %%eax, "LOW "(%1) \n\t" | |
601 | |
602 :"=&d"(bit) | |
603 :"r"(c) | |
604 : "%eax", "%ebx", "%ecx", "memory" | |
605 ); | |
606 return bit+1; | |
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 | 609 c->low += c->low; |
610 | |
2323 | 611 if(!(c->low & CABAC_MASK)) |
612 refill(c); | |
2967 | 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 | 616 return 0; |
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 | 619 return 1; |
620 } | |
4040 | 621 #endif |
1287 | 622 } |
4040 | 623 |
624 | |
625 static always_inline int get_cabac_bypass_sign(CABACContext *c, int val){ | |
626 #ifdef ARCH_X86 | |
627 asm volatile( | |
628 "movl "RANGE "(%1), %%ebx \n\t" | |
629 "movl "LOW "(%1), %%eax \n\t" | |
630 "shl $17, %%ebx \n\t" | |
631 "add %%eax, %%eax \n\t" | |
632 "sub %%ebx, %%eax \n\t" | |
633 "cdq \n\t" | |
634 "and %%edx, %%ebx \n\t" | |
635 "add %%ebx, %%eax \n\t" | |
636 "xor %%edx, %%ecx \n\t" | |
637 "sub %%edx, %%ecx \n\t" | |
638 "test %%ax, %%ax \n\t" | |
639 " jnz 1f \n\t" | |
640 "movl "BYTE "(%1), %%ebx \n\t" | |
641 "subl $0xFFFF, %%eax \n\t" | |
642 "movzwl (%%ebx), %%edx \n\t" | |
643 "bswap %%edx \n\t" | |
644 "shrl $15, %%edx \n\t" | |
645 "addl $2, %%ebx \n\t" | |
646 "addl %%edx, %%eax \n\t" | |
647 "movl %%ebx, "BYTE "(%1) \n\t" | |
648 "1: \n\t" | |
649 "movl %%eax, "LOW "(%1) \n\t" | |
650 | |
651 :"+c"(val) | |
652 :"r"(c) | |
653 : "%eax", "%ebx", "%edx", "memory" | |
654 ); | |
655 return val; | |
656 #else | |
657 int range, mask; | |
658 c->low += c->low; | |
659 | |
660 if(!(c->low & CABAC_MASK)) | |
661 refill(c); | |
662 | |
663 range= c->range<<17; | |
664 c->low -= range; | |
665 mask= c->low >> 31; | |
666 range &= mask; | |
667 c->low += range; | |
668 return (val^mask)-mask; | |
669 #endif | |
670 } | |
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; |
4049
8c1a5ed03a00
another instruction less in decode_significance_x86() -> 1% faster ion P3
michael
parents:
4048
diff
changeset
|
678 int minusindex= 4-(int)index; |
4037
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
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 |
4047
61a4e7218a45
reordering instructions a little in decode_significance_x86() -> 2 instructions less / 1% faster decode_residual on P3
michael
parents:
4046
diff
changeset
|
686 BRANCHLESS_GET_CABAC("%%edx", "%3", "(%1)", "%%ebx", "%%bx", "%%esi", "%%eax", "%%al") |
4037
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
687 |
4047
61a4e7218a45
reordering instructions a little in decode_significance_x86() -> 2 instructions less / 1% faster decode_residual on P3
michael
parents:
4046
diff
changeset
|
688 "test $1, %%edx \n\t" |
4037
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
689 " 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
|
690 |
4047
61a4e7218a45
reordering instructions a little in decode_significance_x86() -> 2 instructions less / 1% faster decode_residual on P3
michael
parents:
4046
diff
changeset
|
691 BRANCHLESS_GET_CABAC("%%edx", "%3", "61(%1)", "%%ebx", "%%bx", "%%esi", "%%eax", "%%al") |
61a4e7218a45
reordering instructions a little in decode_significance_x86() -> 2 instructions less / 1% faster decode_residual on P3
michael
parents:
4046
diff
changeset
|
692 |
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
|
693 "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
|
694 "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
|
695 "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
|
696 "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
|
697 |
4047
61a4e7218a45
reordering instructions a little in decode_significance_x86() -> 2 instructions less / 1% faster decode_residual on P3
michael
parents:
4046
diff
changeset
|
698 "test $1, %%edx \n\t" |
4037
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
699 " 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
|
700 |
4049
8c1a5ed03a00
another instruction less in decode_significance_x86() -> 1% faster ion P3
michael
parents:
4048
diff
changeset
|
701 "addl $4, %%eax \n\t" |
4048 | 702 "movl %%eax, %2 \n\t" |
703 | |
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
|
704 "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
|
705 "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
|
706 "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
|
707 " 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
|
708 "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
|
709 "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
|
710 "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
|
711 "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
|
712 "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
|
713 "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
|
714 "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
|
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 "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
|
717 "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
|
718 :"=&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
|
719 :"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
|
720 : "%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
|
721 ); |
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 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
|
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 #endif |
1287 | 725 |
1300
e18667d1e94d
FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents:
1298
diff
changeset
|
726 /** |
e18667d1e94d
FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents:
1298
diff
changeset
|
727 * |
e18667d1e94d
FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents:
1298
diff
changeset
|
728 * @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
|
729 */ |
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
|
730 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
|
731 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
|
732 if(c->low < c->range<<17){ |
2323 | 733 renorm_cabac_decoder_once(c); |
1287 | 734 return 0; |
735 }else{ | |
1300
e18667d1e94d
FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents:
1298
diff
changeset
|
736 return c->bytestream - c->bytestream_start; |
2967 | 737 } |
1287 | 738 } |
739 | |
1290 | 740 /** |
741 * get (truncated) unnary binarization. | |
742 */ | |
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
|
743 static int get_cabac_u(CABACContext *c, uint8_t * state, int max, int max_index, int truncated){ |
1290 | 744 int i; |
2967 | 745 |
746 for(i=0; i<max; i++){ | |
1290 | 747 if(get_cabac(c, state)==0) |
748 return i; | |
2967 | 749 |
1290 | 750 if(i< max_index) state++; |
751 } | |
752 | |
753 return truncated ? max : -1; | |
754 } | |
755 | |
756 /** | |
757 * get unary exp golomb k-th order binarization. | |
758 */ | |
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
|
759 static int get_cabac_ueg(CABACContext *c, uint8_t * state, int max, int is_signed, int k, int max_index){ |
1290 | 760 int i, v; |
761 int m= 1<<k; | |
2967 | 762 |
763 if(get_cabac(c, state)==0) | |
1290 | 764 return 0; |
2967 | 765 |
1290 | 766 if(0 < max_index) state++; |
2967 | 767 |
768 for(i=1; i<max; i++){ | |
1290 | 769 if(get_cabac(c, state)==0){ |
770 if(is_signed && get_cabac_bypass(c)){ | |
771 return -i; | |
772 }else | |
773 return i; | |
774 } | |
775 | |
776 if(i < max_index) state++; | |
777 } | |
2967 | 778 |
1290 | 779 while(get_cabac_bypass(c)){ |
780 i+= m; | |
781 m+= m; | |
782 } | |
2967 | 783 |
1290 | 784 v=0; |
785 while(m>>=1){ | |
786 v+= v + get_cabac_bypass(c); | |
787 } | |
788 i += v; | |
789 | |
790 if(is_signed && get_cabac_bypass(c)){ | |
791 return -i; | |
792 }else | |
793 return i; | |
794 } |