Mercurial > libavcodec.hg
annotate cabac.h @ 12230:00a4a6b56c7f libavcodec
Fix stupid bug in VP8 prefetching code
| author | darkshikari |
|---|---|
| date | Thu, 22 Jul 2010 22:15:43 +0000 |
| parents | 7dd2a45249a9 |
| children |
| 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 */ |
| 2967 | 21 |
| 1287 | 22 /** |
|
11644
7dd2a45249a9
Remove explicit filename from Doxygen @file commands.
diego
parents:
10071
diff
changeset
|
23 * @file |
| 1287 | 24 * Context Adaptive Binary Arithmetic Coder. |
| 25 */ | |
| 26 | |
| 7760 | 27 #ifndef AVCODEC_CABAC_H |
| 28 #define AVCODEC_CABAC_H | |
| 4975 | 29 |
|
9411
4cb7c65fc775
Split bitstream.h, put the bitstream writer stuff in the new file
stefano
parents:
8718
diff
changeset
|
30 #include "put_bits.h" |
| 1287 | 31 |
|
3284
a224d9752912
don't force asserts in release builds. 2% faster h264.
lorenm
parents:
3036
diff
changeset
|
32 //#undef NDEBUG |
| 1287 | 33 #include <assert.h> |
| 6763 | 34 #include "libavutil/x86_cpu.h" |
| 1287 | 35 |
|
3948
3edbf131ee44
refill cabac variables in 16bit steps, 3% faster get_cabac()
michael
parents:
3947
diff
changeset
|
36 #define CABAC_BITS 16 |
| 2323 | 37 #define CABAC_MASK ((1<<CABAC_BITS)-1) |
| 3984 | 38 #define BRANCHLESS_CABAC_DECODER 1 |
| 4039 | 39 //#define ARCH_X86_DISABLED 1 |
| 2323 | 40 |
| 1287 | 41 typedef struct CABACContext{ |
| 42 int low; | |
| 43 int range; | |
| 44 int outstanding_count; | |
| 45 #ifdef STRICT_LIMITS | |
| 46 int symCount; | |
| 47 #endif | |
|
2024
f65d87bfdd5a
some of the warning fixes by (Michael Roitzsch <mroi at users dot sourceforge dot net>)
michael
parents:
1787
diff
changeset
|
48 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
|
49 const uint8_t *bytestream; |
| 2116 | 50 const uint8_t *bytestream_end; |
| 1287 | 51 PutBitContext pb; |
| 52 }CABACContext; | |
| 53 | |
|
4014
b2582438effe
dehack *ps_state indexing in the branchless decoder
michael
parents:
4012
diff
changeset
|
54 extern uint8_t ff_h264_mlps_state[4*64]; |
| 4039 | 55 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
|
56 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
|
57 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
|
58 extern const uint8_t ff_h264_norm_shift[512]; |
| 2323 | 59 |
| 1287 | 60 |
| 61 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
|
62 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
|
63 void ff_init_cabac_states(CABACContext *c); |
| 1287 | 64 |
| 65 | |
| 66 static inline void put_cabac_bit(CABACContext *c, int b){ | |
| 2967 | 67 put_bits(&c->pb, 1, b); |
| 68 for(;c->outstanding_count; c->outstanding_count--){ | |
| 1287 | 69 put_bits(&c->pb, 1, 1-b); |
| 70 } | |
| 71 } | |
| 72 | |
| 73 static inline void renorm_cabac_encoder(CABACContext *c){ | |
| 74 while(c->range < 0x100){ | |
| 75 //FIXME optimize | |
| 76 if(c->low<0x100){ | |
| 77 put_cabac_bit(c, 0); | |
| 78 }else if(c->low<0x200){ | |
| 79 c->outstanding_count++; | |
| 80 c->low -= 0x100; | |
| 81 }else{ | |
| 82 put_cabac_bit(c, 1); | |
| 83 c->low -= 0x200; | |
| 84 } | |
| 2967 | 85 |
| 1287 | 86 c->range+= c->range; |
| 87 c->low += c->low; | |
| 88 } | |
| 89 } | |
| 90 | |
|
6170
8cbfc9d75833
Put some disabled functions that are only used in the test program in cabac.c
diego
parents:
5973
diff
changeset
|
91 #ifdef TEST |
|
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
|
92 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
|
93 int RangeLPS= ff_h264_lps_range[2*(c->range&0xC0) + *state]; |
| 2967 | 94 |
| 1287 | 95 if(bit == ((*state)&1)){ |
| 96 c->range -= RangeLPS; | |
|
3993
8b7c59b7af01
make state transition tables global as they are constant and the code is slightly faster that way
michael
parents:
3992
diff
changeset
|
97 *state= ff_h264_mps_state[*state]; |
| 1287 | 98 }else{ |
| 99 c->low += c->range - RangeLPS; | |
| 100 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
|
101 *state= ff_h264_lps_state[*state]; |
| 1287 | 102 } |
| 2967 | 103 |
| 1287 | 104 renorm_cabac_encoder(c); |
| 105 | |
| 106 #ifdef STRICT_LIMITS | |
| 107 c->symCount++; | |
| 108 #endif | |
| 109 } | |
| 110 | |
|
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
|
111 static void put_cabac_static(CABACContext *c, int RangeLPS, int bit){ |
| 1287 | 112 assert(c->range > RangeLPS); |
| 113 | |
| 114 if(!bit){ | |
| 115 c->range -= RangeLPS; | |
| 116 }else{ | |
| 117 c->low += c->range - RangeLPS; | |
| 118 c->range = RangeLPS; | |
| 119 } | |
| 120 | |
| 121 renorm_cabac_encoder(c); | |
| 122 | |
| 123 #ifdef STRICT_LIMITS | |
| 124 c->symCount++; | |
| 125 #endif | |
| 126 } | |
| 127 | |
| 1290 | 128 /** |
| 129 * @param bit 0 -> write zero bit, !=0 write one bit | |
| 130 */ | |
|
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
|
131 static void put_cabac_bypass(CABACContext *c, int bit){ |
| 1287 | 132 c->low += c->low; |
| 133 | |
| 134 if(bit){ | |
| 135 c->low += c->range; | |
| 136 } | |
| 137 //FIXME optimize | |
| 138 if(c->low<0x200){ | |
| 139 put_cabac_bit(c, 0); | |
| 140 }else if(c->low<0x400){ | |
| 141 c->outstanding_count++; | |
| 142 c->low -= 0x200; | |
| 143 }else{ | |
| 144 put_cabac_bit(c, 1); | |
| 145 c->low -= 0x400; | |
| 146 } | |
| 2967 | 147 |
| 1287 | 148 #ifdef STRICT_LIMITS |
| 149 c->symCount++; | |
| 150 #endif | |
| 151 } | |
| 152 | |
|
1300
e18667d1e94d
FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents:
1298
diff
changeset
|
153 /** |
|
e18667d1e94d
FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents:
1298
diff
changeset
|
154 * |
|
e18667d1e94d
FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents:
1298
diff
changeset
|
155 * @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
|
156 */ |
|
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
|
157 static int put_cabac_terminate(CABACContext *c, int bit){ |
| 1287 | 158 c->range -= 2; |
| 159 | |
| 160 if(!bit){ | |
| 161 renorm_cabac_encoder(c); | |
| 162 }else{ | |
| 163 c->low += c->range; | |
| 164 c->range= 2; | |
| 2967 | 165 |
| 1287 | 166 renorm_cabac_encoder(c); |
| 167 | |
| 168 assert(c->low <= 0x1FF); | |
| 169 put_cabac_bit(c, c->low>>9); | |
| 170 put_bits(&c->pb, 2, ((c->low>>7)&3)|1); | |
| 2967 | 171 |
| 1287 | 172 flush_put_bits(&c->pb); //FIXME FIXME FIXME XXX wrong |
| 173 } | |
| 2967 | 174 |
| 1287 | 175 #ifdef STRICT_LIMITS |
| 176 c->symCount++; | |
| 177 #endif | |
|
1300
e18667d1e94d
FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents:
1298
diff
changeset
|
178 |
| 1787 | 179 return (put_bits_count(&c->pb)+7)>>3; |
| 1287 | 180 } |
| 181 | |
| 1290 | 182 /** |
| 183 * put (truncated) unary binarization. | |
| 184 */ | |
|
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
|
185 static void put_cabac_u(CABACContext *c, uint8_t * state, int v, int max, int max_index, int truncated){ |
| 1290 | 186 int i; |
| 2967 | 187 |
| 1290 | 188 assert(v <= max); |
| 2967 | 189 |
| 1290 | 190 #if 1 |
| 191 for(i=0; i<v; i++){ | |
| 192 put_cabac(c, state, 1); | |
| 193 if(i < max_index) state++; | |
| 194 } | |
| 195 if(truncated==0 || v<max) | |
| 196 put_cabac(c, state, 0); | |
| 197 #else | |
| 198 if(v <= max_index){ | |
| 199 for(i=0; i<v; i++){ | |
| 200 put_cabac(c, state+i, 1); | |
| 201 } | |
| 202 if(truncated==0 || v<max) | |
| 203 put_cabac(c, state+i, 0); | |
| 204 }else{ | |
| 205 for(i=0; i<=max_index; i++){ | |
| 206 put_cabac(c, state+i, 1); | |
| 207 } | |
| 208 for(; i<v; i++){ | |
| 209 put_cabac(c, state+max_index, 1); | |
| 210 } | |
| 211 if(truncated==0 || v<max) | |
| 212 put_cabac(c, state+max_index, 0); | |
| 213 } | |
| 214 #endif | |
| 215 } | |
| 216 | |
| 217 /** | |
| 218 * put unary exp golomb k-th order binarization. | |
| 219 */ | |
|
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
|
220 static void put_cabac_ueg(CABACContext *c, uint8_t * state, int v, int max, int is_signed, int k, int max_index){ |
| 1290 | 221 int i; |
| 2967 | 222 |
| 1290 | 223 if(v==0) |
| 224 put_cabac(c, state, 0); | |
| 225 else{ | |
| 1298 | 226 const int sign= v < 0; |
| 2967 | 227 |
| 4001 | 228 if(is_signed) v= FFABS(v); |
| 2967 | 229 |
| 1290 | 230 if(v<max){ |
| 231 for(i=0; i<v; i++){ | |
| 232 put_cabac(c, state, 1); | |
| 233 if(i < max_index) state++; | |
| 234 } | |
| 235 | |
| 236 put_cabac(c, state, 0); | |
| 237 }else{ | |
| 238 int m= 1<<k; | |
| 239 | |
| 240 for(i=0; i<max; i++){ | |
| 241 put_cabac(c, state, 1); | |
| 242 if(i < max_index) state++; | |
| 243 } | |
| 244 | |
| 245 v -= max; | |
| 246 while(v >= m){ //FIXME optimize | |
| 247 put_cabac_bypass(c, 1); | |
| 248 v-= m; | |
| 249 m+= m; | |
| 250 } | |
| 251 put_cabac_bypass(c, 0); | |
| 252 while(m>>=1){ | |
| 253 put_cabac_bypass(c, v&m); | |
| 254 } | |
| 255 } | |
| 256 | |
| 257 if(is_signed) | |
| 258 put_cabac_bypass(c, sign); | |
| 259 } | |
| 260 } | |
|
6170
8cbfc9d75833
Put some disabled functions that are only used in the test program in cabac.c
diego
parents:
5973
diff
changeset
|
261 #endif /* TEST */ |
| 1290 | 262 |
| 2323 | 263 static void refill(CABACContext *c){ |
| 264 #if CABAC_BITS == 16 | |
| 3946 | 265 c->low+= (c->bytestream[0]<<9) + (c->bytestream[1]<<1); |
| 2323 | 266 #else |
| 267 c->low+= c->bytestream[0]<<1; | |
| 268 #endif | |
| 269 c->low -= CABAC_MASK; | |
| 270 c->bytestream+= CABAC_BITS/8; | |
| 271 } | |
| 272 | |
| 8590 | 273 #if ! ( ARCH_X86 && HAVE_7REGS && HAVE_EBX_AVAILABLE && !defined(BROKEN_RELOCATIONS) ) |
| 2323 | 274 static void refill2(CABACContext *c){ |
| 275 int i, x; | |
| 276 | |
| 277 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
|
278 i= 7 - ff_h264_norm_shift[x>>(CABAC_BITS-1)]; |
| 2323 | 279 |
| 280 x= -CABAC_MASK; | |
| 2967 | 281 |
| 2323 | 282 #if CABAC_BITS == 16 |
| 283 x+= (c->bytestream[0]<<9) + (c->bytestream[1]<<1); | |
| 284 #else | |
| 285 x+= c->bytestream[0]<<1; | |
| 286 #endif | |
| 2967 | 287 |
| 2323 | 288 c->low += x<<i; |
| 289 c->bytestream+= CABAC_BITS/8; | |
| 290 } | |
| 5811 | 291 #endif |
| 2323 | 292 |
| 1287 | 293 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
|
294 while(c->range < 0x100){ |
| 1287 | 295 c->range+= c->range; |
| 296 c->low+= c->low; | |
| 2323 | 297 if(!(c->low & CABAC_MASK)) |
| 298 refill(c); | |
| 1287 | 299 } |
| 300 } | |
| 301 | |
| 2323 | 302 static inline void renorm_cabac_decoder_once(CABACContext *c){ |
| 3951 | 303 #ifdef ARCH_X86_DISABLED |
|
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
304 int temp; |
|
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
305 #if 0 |
|
3950
900d21b85dd6
renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents:
3948
diff
changeset
|
306 //P3:683 athlon:475 |
| 8031 | 307 __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
|
308 "lea -0x100(%0), %2 \n\t" |
|
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
309 "shr $31, %2 \n\t" //FIXME 31->63 for x86-64 |
|
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
310 "shl %%cl, %0 \n\t" |
|
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
311 "shl %%cl, %1 \n\t" |
|
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
312 : "+r"(c->range), "+r"(c->low), "+c"(temp) |
|
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
313 ); |
|
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
314 #elif 0 |
|
3950
900d21b85dd6
renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents:
3948
diff
changeset
|
315 //P3:680 athlon:474 |
| 8031 | 316 __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
|
317 "cmp $0x100, %0 \n\t" |
|
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
318 "setb %%cl \n\t" //FIXME 31->63 for x86-64 |
|
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
319 "shl %%cl, %0 \n\t" |
|
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
320 "shl %%cl, %1 \n\t" |
|
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
321 : "+r"(c->range), "+r"(c->low), "+c"(temp) |
|
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
322 ); |
|
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
323 #elif 1 |
|
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
324 int temp2; |
|
3950
900d21b85dd6
renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents:
3948
diff
changeset
|
325 //P3:665 athlon:517 |
| 8031 | 326 __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
|
327 "lea -0x100(%0), %%eax \n\t" |
|
6425
8b119a5a0644
Use cltd instead of cdq asm mnemonic, ICC and gcc support both, but
reimar
parents:
6170
diff
changeset
|
328 "cltd \n\t" |
|
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
329 "mov %0, %%eax \n\t" |
|
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
330 "and %%edx, %0 \n\t" |
|
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
331 "and %1, %%edx \n\t" |
|
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
332 "add %%eax, %0 \n\t" |
|
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
333 "add %%edx, %1 \n\t" |
|
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
334 : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2) |
|
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
335 ); |
|
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
336 #elif 0 |
|
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
337 int temp2; |
|
3950
900d21b85dd6
renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents:
3948
diff
changeset
|
338 //P3:673 athlon:509 |
| 8031 | 339 __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
|
340 "cmp $0x100, %0 \n\t" |
|
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
341 "sbb %%edx, %%edx \n\t" |
|
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
342 "mov %0, %%eax \n\t" |
|
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
343 "and %%edx, %0 \n\t" |
|
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
344 "and %1, %%edx \n\t" |
|
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
345 "add %%eax, %0 \n\t" |
|
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
346 "add %%edx, %1 \n\t" |
|
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
347 : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2) |
|
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
348 ); |
|
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
349 #else |
|
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
350 int temp2; |
|
3950
900d21b85dd6
renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents:
3948
diff
changeset
|
351 //P3:677 athlon:511 |
| 8031 | 352 __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
|
353 "cmp $0x100, %0 \n\t" |
|
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
354 "lea (%0, %0), %%eax \n\t" |
|
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
355 "lea (%1, %1), %%edx \n\t" |
|
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
356 "cmovb %%eax, %0 \n\t" |
|
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
357 "cmovb %%edx, %1 \n\t" |
|
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
358 : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2) |
|
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
359 ); |
|
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
360 #endif |
|
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
361 #else |
|
3950
900d21b85dd6
renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents:
3948
diff
changeset
|
362 //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
|
363 int shift= (uint32_t)(c->range - 0x100)>>31; |
| 3642 | 364 c->range<<= shift; |
| 365 c->low <<= shift; | |
|
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
366 #endif |
| 2323 | 367 if(!(c->low & CABAC_MASK)) |
| 368 refill(c); | |
| 369 } | |
| 370 | |
|
4908
777f250df232
Fix multiple "?inline/static? is not at beginning of declaration" warnings.
diego
parents:
4882
diff
changeset
|
371 static av_always_inline int get_cabac_inline(CABACContext *c, uint8_t * const state){ |
| 3642 | 372 //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
|
373 #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
|
374 #define RANGE "4" |
| 8590 | 375 #if ARCH_X86_64 |
| 4064 | 376 #define BYTESTART "16" |
| 377 #define BYTE "24" | |
| 378 #define BYTEEND "32" | |
| 379 #else | |
|
3993
8b7c59b7af01
make state transition tables global as they are constant and the code is slightly faster that way
michael
parents:
3992
diff
changeset
|
380 #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
|
381 #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
|
382 #define BYTEEND "20" |
| 4064 | 383 #endif |
| 8590 | 384 #if ARCH_X86 && HAVE_7REGS && HAVE_EBX_AVAILABLE && !defined(BROKEN_RELOCATIONS) |
|
4044
5ccdefd60f61
Fix PIC compilation, some defines were under #ifdef !PIC but used
diego
parents:
4043
diff
changeset
|
385 int bit; |
|
5ccdefd60f61
Fix PIC compilation, some defines were under #ifdef !PIC but used
diego
parents:
4043
diff
changeset
|
386 |
| 3984 | 387 #ifndef BRANCHLESS_CABAC_DECODER |
| 8031 | 388 __asm__ volatile( |
| 4035 | 389 "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
|
390 "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
|
391 "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
|
392 "andl $0xC0, %%ebx \n\t" |
| 4035 | 393 "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
|
394 "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
|
395 //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
|
396 "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
|
397 "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
|
398 "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
|
399 "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
|
400 " ja 1f \n\t" |
|
3999
6cbad3675632
slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents:
3996
diff
changeset
|
401 |
|
6cbad3675632
slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents:
3996
diff
changeset
|
402 #if 1 |
|
6cbad3675632
slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents:
3996
diff
changeset
|
403 //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
|
404 "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
|
405 "shr $31, %%ecx \n\t" |
|
6cbad3675632
slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents:
3996
diff
changeset
|
406 "shl %%cl, %%edx \n\t" |
|
6cbad3675632
slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents:
3996
diff
changeset
|
407 "shl %%cl, %%ebx \n\t" |
|
6cbad3675632
slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents:
3996
diff
changeset
|
408 #else |
|
6cbad3675632
slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents:
3996
diff
changeset
|
409 //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
|
410 "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
|
411 "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
|
412 "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
|
413 "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
|
414 #endif |
| 4035 | 415 "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
|
416 "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
|
417 //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
|
418 "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
|
419 " jnz 2f \n\t" |
| 4064 | 420 "mov "BYTE "(%2), %%"REG_S" \n\t" |
|
3969
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
421 "subl $0xFFFF, %%ebx \n\t" |
| 4064 | 422 "movzwl (%%"REG_S"), %%ecx \n\t" |
|
3969
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
423 "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
|
424 "shrl $15, %%ecx \n\t" |
| 4064 | 425 "add $2, %%"REG_S" \n\t" |
|
3969
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
426 "addl %%ecx, %%ebx \n\t" |
| 4064 | 427 "mov %%"REG_S", "BYTE "(%2) \n\t" |
|
3969
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
428 "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
|
429 "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
|
430 //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
|
431 "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
|
432 "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
|
433 "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
|
434 "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
|
435 "shll %%cl, %%edx \n\t" |
| 4035 | 436 "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
|
437 "movb %%cl, (%1) \n\t" |
| 4064 | 438 "add $1, %0 \n\t" |
|
3969
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
439 "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
|
440 " 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
|
441 |
| 4064 | 442 "mov "BYTE "(%2), %%"REG_c" \n\t" |
| 443 "movzwl (%%"REG_c"), %%esi \n\t" | |
|
3969
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
444 "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
|
445 "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
|
446 "subl $0xFFFF, %%esi \n\t" |
| 4064 | 447 "add $2, %%"REG_c" \n\t" |
| 448 "mov %%"REG_c", "BYTE "(%2) \n\t" | |
|
3969
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
449 |
|
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
450 "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
|
451 "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
|
452 "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
|
453 "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
|
454 "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
|
455 "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
|
456 |
|
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
457 "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
|
458 "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
|
459 "2: \n\t" |
|
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
460 "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
|
461 "movl %%ebx, "LOW "(%2) \n\t" |
| 5408 | 462 :"=&a"(bit) //FIXME this is fragile gcc either runs out of registers or miscompiles it (for example if "+a"(bit) or "+m"(*state) is used |
|
3969
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
463 :"r"(state), "r"(c) |
| 4064 | 464 : "%"REG_c, "%ebx", "%edx", "%"REG_S, "memory" |
|
3969
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
465 ); |
|
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
|
466 bit&=1; |
|
4002
ec426fa57dfe
adds some useful comments after some of the #else, #elseif,
gpoirier
parents:
4001
diff
changeset
|
467 #else /* BRANCHLESS_CABAC_DECODER */ |
|
4046
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
468 |
|
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
469 |
| 8590 | 470 #if HAVE_FAST_CMOV |
|
4046
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
471 #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
|
472 "mov "tmp" , %%ecx \n\t"\ |
|
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
473 "shl $17 , "tmp" \n\t"\ |
|
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
474 "cmp "low" , "tmp" \n\t"\ |
|
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
475 "cmova %%ecx , "range" \n\t"\ |
|
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
476 "sbb %%ecx , %%ecx \n\t"\ |
|
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
477 "and %%ecx , "tmp" \n\t"\ |
|
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
478 "sub "tmp" , "low" \n\t"\ |
|
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
479 "xor %%ecx , "ret" \n\t" |
|
4418
4cceb7c877af
rename CMOV_IS_FAST to HAVE_FAST_CMOV and simplify configure
mru
parents:
4345
diff
changeset
|
480 #else /* HAVE_FAST_CMOV */ |
|
4046
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
481 #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
|
482 "mov "tmp" , %%ecx \n\t"\ |
|
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
483 "shl $17 , "tmp" \n\t"\ |
|
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
484 "sub "low" , "tmp" \n\t"\ |
|
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
485 "sar $31 , "tmp" \n\t" /*lps_mask*/\ |
|
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
486 "sub %%ecx , "range" \n\t" /*RangeLPS - range*/\ |
|
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
487 "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
|
488 "add %%ecx , "range" \n\t" /*new range*/\ |
|
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
489 "shl $17 , %%ecx \n\t"\ |
|
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
490 "and "tmp" , %%ecx \n\t"\ |
|
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
491 "sub %%ecx , "low" \n\t"\ |
|
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
492 "xor "tmp" , "ret" \n\t" |
|
4418
4cceb7c877af
rename CMOV_IS_FAST to HAVE_FAST_CMOV and simplify configure
mru
parents:
4345
diff
changeset
|
493 #endif /* HAVE_FAST_CMOV */ |
| 3975 | 494 |
| 495 | |
|
4046
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
496 #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
|
497 "movzbl "statep" , "ret" \n\t"\ |
|
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
498 "mov "range" , "tmp" \n\t"\ |
|
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
499 "and $0xC0 , "range" \n\t"\ |
|
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
500 "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
|
501 "sub "range" , "tmp" \n\t"\ |
|
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
502 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
|
503 "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
|
504 "shl %%cl , "range" \n\t"\ |
|
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
505 "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
|
506 "mov "tmpbyte" , "statep" \n\t"\ |
|
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
507 "shl %%cl , "low" \n\t"\ |
|
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
508 "test "lowword" , "lowword" \n\t"\ |
|
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
509 " jnz 1f \n\t"\ |
| 4064 | 510 "mov "BYTE"("cabac"), %%"REG_c" \n\t"\ |
| 511 "movzwl (%%"REG_c") , "tmp" \n\t"\ | |
|
4046
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
512 "bswap "tmp" \n\t"\ |
|
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
513 "shr $15 , "tmp" \n\t"\ |
|
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
514 "sub $0xFFFF , "tmp" \n\t"\ |
| 4064 | 515 "add $2 , %%"REG_c" \n\t"\ |
| 516 "mov %%"REG_c" , "BYTE "("cabac") \n\t"\ | |
|
4046
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
517 "lea -1("low") , %%ecx \n\t"\ |
|
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
518 "xor "low" , %%ecx \n\t"\ |
|
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
519 "shr $15 , %%ecx \n\t"\ |
|
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
520 "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
|
521 "neg %%ecx \n\t"\ |
|
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
522 "add $7 , %%ecx \n\t"\ |
|
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
523 "shl %%cl , "tmp" \n\t"\ |
|
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
524 "add "tmp" , "low" \n\t"\ |
|
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
525 "1: \n\t" |
| 3975 | 526 |
| 8031 | 527 __asm__ volatile( |
|
4046
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
528 "movl "RANGE "(%2), %%esi \n\t" |
|
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
529 "movl "LOW "(%2), %%ebx \n\t" |
|
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
530 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
|
531 "movl %%esi, "RANGE "(%2) \n\t" |
|
8bbc695c9603
factorize get_cabac asm (0.5% slower but its much cleaner)
michael
parents:
4044
diff
changeset
|
532 "movl %%ebx, "LOW "(%2) \n\t" |
| 3975 | 533 |
| 534 :"=&a"(bit) | |
| 535 :"r"(state), "r"(c) | |
| 4064 | 536 : "%"REG_c, "%ebx", "%edx", "%esi", "memory" |
| 3975 | 537 ); |
|
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
|
538 bit&=1; |
|
4002
ec426fa57dfe
adds some useful comments after some of the #else, #elseif,
gpoirier
parents:
4001
diff
changeset
|
539 #endif /* BRANCHLESS_CABAC_DECODER */ |
| 8590 | 540 #else /* ARCH_X86 && HAVE_7REGS && HAVE_EBX_AVAILABLE && !defined(BROKEN_RELOCATIONS) */ |
| 3642 | 541 int s = *state; |
| 4039 | 542 int RangeLPS= ff_h264_lps_range[2*(c->range&0xC0) + s]; |
|
5083
ce36118abbbb
rename attribute_unused to av_unused and moves its declaration to common.h
benoit
parents:
4975
diff
changeset
|
543 int bit, lps_mask av_unused; |
| 2967 | 544 |
| 1287 | 545 c->range -= RangeLPS; |
| 3984 | 546 #ifndef BRANCHLESS_CABAC_DECODER |
|
4345
88967250d718
replace a few hardcoded numbers with their correct named ones
michael
parents:
4283
diff
changeset
|
547 if(c->low < (c->range<<(CABAC_BITS+1))){ |
| 3642 | 548 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
|
549 *state= ff_h264_mps_state[s]; |
| 2323 | 550 renorm_cabac_decoder_once(c); |
| 1287 | 551 }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
|
552 bit= ff_h264_norm_shift[RangeLPS]; |
|
4345
88967250d718
replace a few hardcoded numbers with their correct named ones
michael
parents:
4283
diff
changeset
|
553 c->low -= (c->range<<(CABAC_BITS+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
|
554 *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
|
555 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
|
556 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
|
557 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
|
558 |
|
4345
88967250d718
replace a few hardcoded numbers with their correct named ones
michael
parents:
4283
diff
changeset
|
559 if(!(c->low & CABAC_MASK)){ |
| 2323 | 560 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
|
561 } |
| 1287 | 562 } |
|
4002
ec426fa57dfe
adds some useful comments after some of the #else, #elseif,
gpoirier
parents:
4001
diff
changeset
|
563 #else /* BRANCHLESS_CABAC_DECODER */ |
|
4345
88967250d718
replace a few hardcoded numbers with their correct named ones
michael
parents:
4283
diff
changeset
|
564 lps_mask= ((c->range<<(CABAC_BITS+1)) - c->low)>>31; |
| 2967 | 565 |
|
4345
88967250d718
replace a few hardcoded numbers with their correct named ones
michael
parents:
4283
diff
changeset
|
566 c->low -= (c->range<<(CABAC_BITS+1)) & lps_mask; |
| 2323 | 567 c->range += (RangeLPS - c->range) & lps_mask; |
| 2967 | 568 |
| 3974 | 569 s^=lps_mask; |
|
4014
b2582438effe
dehack *ps_state indexing in the branchless decoder
michael
parents:
4012
diff
changeset
|
570 *state= (ff_h264_mlps_state+128)[s]; |
| 3974 | 571 bit= s&1; |
| 2967 | 572 |
|
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
|
573 lps_mask= ff_h264_norm_shift[c->range]; |
| 2323 | 574 c->range<<= lps_mask; |
| 575 c->low <<= lps_mask; | |
| 576 if(!(c->low & CABAC_MASK)) | |
| 577 refill2(c); | |
|
4002
ec426fa57dfe
adds some useful comments after some of the #else, #elseif,
gpoirier
parents:
4001
diff
changeset
|
578 #endif /* BRANCHLESS_CABAC_DECODER */ |
| 8590 | 579 #endif /* ARCH_X86 && HAVE_7REGS && HAVE_EBX_AVAILABLE && !defined(BROKEN_RELOCATIONS) */ |
| 2967 | 580 return bit; |
| 1287 | 581 } |
| 582 | |
|
7775
490ee89408c5
Silence a couple of 'defined but not used' warnings by adding an av_unused
diego
parents:
7760
diff
changeset
|
583 static int av_noinline av_unused get_cabac_noinline(CABACContext *c, uint8_t * const state){ |
|
4008
b636f3d59283
prevent "mb level" get_cabac() calls from being inlined (3% faster decode_mb_cabac() on P3)
michael
parents:
4002
diff
changeset
|
584 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
|
585 } |
|
b636f3d59283
prevent "mb level" get_cabac() calls from being inlined (3% faster decode_mb_cabac() on P3)
michael
parents:
4002
diff
changeset
|
586 |
|
7775
490ee89408c5
Silence a couple of 'defined but not used' warnings by adding an av_unused
diego
parents:
7760
diff
changeset
|
587 static int av_unused get_cabac(CABACContext *c, uint8_t * const state){ |
|
4008
b636f3d59283
prevent "mb level" get_cabac() calls from being inlined (3% faster decode_mb_cabac() on P3)
michael
parents:
4002
diff
changeset
|
588 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
|
589 } |
|
b636f3d59283
prevent "mb level" get_cabac() calls from being inlined (3% faster decode_mb_cabac() on P3)
michael
parents:
4002
diff
changeset
|
590 |
|
7775
490ee89408c5
Silence a couple of 'defined but not used' warnings by adding an av_unused
diego
parents:
7760
diff
changeset
|
591 static int av_unused get_cabac_bypass(CABACContext *c){ |
| 4040 | 592 #if 0 //not faster |
| 593 int bit; | |
| 8031 | 594 __asm__ volatile( |
| 4040 | 595 "movl "RANGE "(%1), %%ebx \n\t" |
| 596 "movl "LOW "(%1), %%eax \n\t" | |
| 597 "shl $17, %%ebx \n\t" | |
| 598 "add %%eax, %%eax \n\t" | |
| 599 "sub %%ebx, %%eax \n\t" | |
|
6425
8b119a5a0644
Use cltd instead of cdq asm mnemonic, ICC and gcc support both, but
reimar
parents:
6170
diff
changeset
|
600 "cltd \n\t" |
| 4040 | 601 "and %%edx, %%ebx \n\t" |
| 602 "add %%ebx, %%eax \n\t" | |
| 603 "test %%ax, %%ax \n\t" | |
| 604 " jnz 1f \n\t" | |
| 4064 | 605 "movl "BYTE "(%1), %%"REG_b" \n\t" |
| 4040 | 606 "subl $0xFFFF, %%eax \n\t" |
| 4064 | 607 "movzwl (%%"REG_b"), %%ecx \n\t" |
| 4040 | 608 "bswap %%ecx \n\t" |
| 609 "shrl $15, %%ecx \n\t" | |
| 4064 | 610 "addl $2, %%"REG_b" \n\t" |
| 4040 | 611 "addl %%ecx, %%eax \n\t" |
| 4064 | 612 "movl %%"REG_b", "BYTE "(%1) \n\t" |
| 4040 | 613 "1: \n\t" |
| 614 "movl %%eax, "LOW "(%1) \n\t" | |
| 615 | |
| 616 :"=&d"(bit) | |
| 617 :"r"(c) | |
| 4064 | 618 : "%eax", "%"REG_b, "%ecx", "memory" |
| 4040 | 619 ); |
| 620 return bit+1; | |
| 621 #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
|
622 int range; |
| 1287 | 623 c->low += c->low; |
| 624 | |
| 2323 | 625 if(!(c->low & CABAC_MASK)) |
| 626 refill(c); | |
| 2967 | 627 |
|
4345
88967250d718
replace a few hardcoded numbers with their correct named ones
michael
parents:
4283
diff
changeset
|
628 range= c->range<<(CABAC_BITS+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
|
629 if(c->low < range){ |
| 1287 | 630 return 0; |
| 631 }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
|
632 c->low -= range; |
| 1287 | 633 return 1; |
| 634 } | |
| 4040 | 635 #endif |
| 1287 | 636 } |
| 4040 | 637 |
| 638 | |
|
4283
d6f83e2f8804
rename always_inline to av_always_inline and move to common.h
mru
parents:
4241
diff
changeset
|
639 static av_always_inline int get_cabac_bypass_sign(CABACContext *c, int val){ |
|
10071
6ac0d4957d35
Replace #ifdef PIC checks with the more appropriate HAVE_EBX_AVAILABLE/HAVE_7REGS.
reimar
parents:
9411
diff
changeset
|
640 #if ARCH_X86 && HAVE_EBX_AVAILABLE |
| 8031 | 641 __asm__ volatile( |
| 4040 | 642 "movl "RANGE "(%1), %%ebx \n\t" |
| 643 "movl "LOW "(%1), %%eax \n\t" | |
| 644 "shl $17, %%ebx \n\t" | |
| 645 "add %%eax, %%eax \n\t" | |
| 646 "sub %%ebx, %%eax \n\t" | |
|
6425
8b119a5a0644
Use cltd instead of cdq asm mnemonic, ICC and gcc support both, but
reimar
parents:
6170
diff
changeset
|
647 "cltd \n\t" |
| 4040 | 648 "and %%edx, %%ebx \n\t" |
| 649 "add %%ebx, %%eax \n\t" | |
| 650 "xor %%edx, %%ecx \n\t" | |
| 651 "sub %%edx, %%ecx \n\t" | |
| 652 "test %%ax, %%ax \n\t" | |
| 653 " jnz 1f \n\t" | |
| 4064 | 654 "mov "BYTE "(%1), %%"REG_b" \n\t" |
| 4040 | 655 "subl $0xFFFF, %%eax \n\t" |
| 4064 | 656 "movzwl (%%"REG_b"), %%edx \n\t" |
| 4040 | 657 "bswap %%edx \n\t" |
| 658 "shrl $15, %%edx \n\t" | |
| 4064 | 659 "add $2, %%"REG_b" \n\t" |
| 4040 | 660 "addl %%edx, %%eax \n\t" |
| 4064 | 661 "mov %%"REG_b", "BYTE "(%1) \n\t" |
| 4040 | 662 "1: \n\t" |
| 663 "movl %%eax, "LOW "(%1) \n\t" | |
| 664 | |
| 665 :"+c"(val) | |
| 666 :"r"(c) | |
| 4064 | 667 : "%eax", "%"REG_b, "%edx", "memory" |
| 4040 | 668 ); |
| 669 return val; | |
| 670 #else | |
| 671 int range, mask; | |
| 672 c->low += c->low; | |
| 673 | |
| 674 if(!(c->low & CABAC_MASK)) | |
| 675 refill(c); | |
| 676 | |
|
4345
88967250d718
replace a few hardcoded numbers with their correct named ones
michael
parents:
4283
diff
changeset
|
677 range= c->range<<(CABAC_BITS+1); |
| 4040 | 678 c->low -= range; |
| 679 mask= c->low >> 31; | |
| 680 range &= mask; | |
| 681 c->low += range; | |
| 682 return (val^mask)-mask; | |
| 683 #endif | |
| 684 } | |
| 685 | |
|
1300
e18667d1e94d
FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents:
1298
diff
changeset
|
686 /** |
|
e18667d1e94d
FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents:
1298
diff
changeset
|
687 * |
|
e18667d1e94d
FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents:
1298
diff
changeset
|
688 * @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
|
689 */ |
|
7775
490ee89408c5
Silence a couple of 'defined but not used' warnings by adding an av_unused
diego
parents:
7760
diff
changeset
|
690 static int av_unused 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
|
691 c->range -= 2; |
|
4345
88967250d718
replace a few hardcoded numbers with their correct named ones
michael
parents:
4283
diff
changeset
|
692 if(c->low < c->range<<(CABAC_BITS+1)){ |
| 2323 | 693 renorm_cabac_decoder_once(c); |
| 1287 | 694 return 0; |
| 695 }else{ | |
|
1300
e18667d1e94d
FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents:
1298
diff
changeset
|
696 return c->bytestream - c->bytestream_start; |
| 2967 | 697 } |
| 1287 | 698 } |
| 699 | |
| 5811 | 700 #if 0 |
| 1290 | 701 /** |
| 5686 | 702 * Get (truncated) unary binarization. |
| 1290 | 703 */ |
|
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
|
704 static int get_cabac_u(CABACContext *c, uint8_t * state, int max, int max_index, int truncated){ |
| 1290 | 705 int i; |
| 2967 | 706 |
| 707 for(i=0; i<max; i++){ | |
| 1290 | 708 if(get_cabac(c, state)==0) |
| 709 return i; | |
| 2967 | 710 |
| 1290 | 711 if(i< max_index) state++; |
| 712 } | |
| 713 | |
| 714 return truncated ? max : -1; | |
| 715 } | |
| 716 | |
| 717 /** | |
| 718 * get unary exp golomb k-th order binarization. | |
| 719 */ | |
|
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
|
720 static int get_cabac_ueg(CABACContext *c, uint8_t * state, int max, int is_signed, int k, int max_index){ |
| 1290 | 721 int i, v; |
| 722 int m= 1<<k; | |
| 2967 | 723 |
| 724 if(get_cabac(c, state)==0) | |
| 1290 | 725 return 0; |
| 2967 | 726 |
| 1290 | 727 if(0 < max_index) state++; |
| 2967 | 728 |
| 729 for(i=1; i<max; i++){ | |
| 1290 | 730 if(get_cabac(c, state)==0){ |
| 731 if(is_signed && get_cabac_bypass(c)){ | |
| 732 return -i; | |
| 733 }else | |
| 734 return i; | |
| 735 } | |
| 736 | |
| 737 if(i < max_index) state++; | |
| 738 } | |
| 2967 | 739 |
| 1290 | 740 while(get_cabac_bypass(c)){ |
| 741 i+= m; | |
| 742 m+= m; | |
| 743 } | |
| 2967 | 744 |
| 1290 | 745 v=0; |
| 746 while(m>>=1){ | |
| 747 v+= v + get_cabac_bypass(c); | |
| 748 } | |
| 749 i += v; | |
| 750 | |
| 751 if(is_signed && get_cabac_bypass(c)){ | |
| 752 return -i; | |
| 753 }else | |
| 754 return i; | |
| 755 } | |
| 5811 | 756 #endif /* 0 */ |
| 4975 | 757 |
| 7760 | 758 #endif /* AVCODEC_CABAC_H */ |
