Mercurial > libavcodec.hg
annotate cabac.h @ 4040:9eaea06c5ba6 libavcodec
optimize sign decoding code in decode_residual()
x86 is 4% faster on P3
C sign stuff + x86 code for everything else is also faster then before (sorry forgot to test pure C)
... and if i replace the second occurance of the sign decoding in decode_residual by the asm too then everything gets slower iam starting to think that it might be best to write the whole function in asm, playing this avoid random deoptimizations game with gcc is not fun at all
author | michael |
---|---|
date | Thu, 19 Oct 2006 01:19:03 +0000 |
parents | 866a83726985 |
children | 87694a28120c |
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 |
3990
746a60ba3177
enable CMOV_IS_FAST as its faster or equal speed on every cpu (duron, athlon, PM, P3) from which ive seen benchmarks, it might be slower on P4 but noone has posted benchmarks ...
michael
parents:
3984
diff
changeset
|
35 #define CMOV_IS_FAST 1 |
4039 | 36 //#define ARCH_X86_DISABLED 1 |
2323 | 37 |
1287 | 38 typedef struct CABACContext{ |
39 int low; | |
40 int range; | |
41 int outstanding_count; | |
42 #ifdef STRICT_LIMITS | |
43 int symCount; | |
44 #endif | |
2024
f65d87bfdd5a
some of the warning fixes by (Michael Roitzsch <mroi at users dot sourceforge dot net>)
michael
parents:
1787
diff
changeset
|
45 const uint8_t *bytestream_start; |
f65d87bfdd5a
some of the warning fixes by (Michael Roitzsch <mroi at users dot sourceforge dot net>)
michael
parents:
1787
diff
changeset
|
46 const uint8_t *bytestream; |
2116 | 47 const uint8_t *bytestream_end; |
1287 | 48 PutBitContext pb; |
49 }CABACContext; | |
50 | |
4014
b2582438effe
dehack *ps_state indexing in the branchless decoder
michael
parents:
4012
diff
changeset
|
51 extern uint8_t ff_h264_mlps_state[4*64]; |
4039 | 52 extern uint8_t ff_h264_lps_range[4*2*64]; ///< rangeTabLPS |
3993
8b7c59b7af01
make state transition tables global as they are constant and the code is slightly faster that way
michael
parents:
3992
diff
changeset
|
53 extern uint8_t ff_h264_mps_state[2*64]; ///< transIdxMPS |
8b7c59b7af01
make state transition tables global as they are constant and the code is slightly faster that way
michael
parents:
3992
diff
changeset
|
54 extern uint8_t ff_h264_lps_state[2*64]; ///< transIdxLPS |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
55 extern const uint8_t ff_h264_norm_shift[512]; |
2323 | 56 |
1287 | 57 |
58 void ff_init_cabac_encoder(CABACContext *c, uint8_t *buf, int buf_size); | |
2024
f65d87bfdd5a
some of the warning fixes by (Michael Roitzsch <mroi at users dot sourceforge dot net>)
michael
parents:
1787
diff
changeset
|
59 void ff_init_cabac_decoder(CABACContext *c, const uint8_t *buf, int buf_size); |
3993
8b7c59b7af01
make state transition tables global as they are constant and the code is slightly faster that way
michael
parents:
3992
diff
changeset
|
60 void ff_init_cabac_states(CABACContext *c); |
1287 | 61 |
62 | |
63 static inline void put_cabac_bit(CABACContext *c, int b){ | |
2967 | 64 put_bits(&c->pb, 1, b); |
65 for(;c->outstanding_count; c->outstanding_count--){ | |
1287 | 66 put_bits(&c->pb, 1, 1-b); |
67 } | |
68 } | |
69 | |
70 static inline void renorm_cabac_encoder(CABACContext *c){ | |
71 while(c->range < 0x100){ | |
72 //FIXME optimize | |
73 if(c->low<0x100){ | |
74 put_cabac_bit(c, 0); | |
75 }else if(c->low<0x200){ | |
76 c->outstanding_count++; | |
77 c->low -= 0x100; | |
78 }else{ | |
79 put_cabac_bit(c, 1); | |
80 c->low -= 0x200; | |
81 } | |
2967 | 82 |
1287 | 83 c->range+= c->range; |
84 c->low += c->low; | |
85 } | |
86 } | |
87 | |
3928
987fffdf6ae7
don't try to inline cabac functions. gcc ignored the hint anyway, and forcing it would make h264 slower.
lorenm
parents:
3642
diff
changeset
|
88 static void put_cabac(CABACContext *c, uint8_t * const state, int bit){ |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
89 int RangeLPS= ff_h264_lps_range[2*(c->range&0xC0) + *state]; |
2967 | 90 |
1287 | 91 if(bit == ((*state)&1)){ |
92 c->range -= RangeLPS; | |
3993
8b7c59b7af01
make state transition tables global as they are constant and the code is slightly faster that way
michael
parents:
3992
diff
changeset
|
93 *state= ff_h264_mps_state[*state]; |
1287 | 94 }else{ |
95 c->low += c->range - RangeLPS; | |
96 c->range = RangeLPS; | |
3993
8b7c59b7af01
make state transition tables global as they are constant and the code is slightly faster that way
michael
parents:
3992
diff
changeset
|
97 *state= ff_h264_lps_state[*state]; |
1287 | 98 } |
2967 | 99 |
1287 | 100 renorm_cabac_encoder(c); |
101 | |
102 #ifdef STRICT_LIMITS | |
103 c->symCount++; | |
104 #endif | |
105 } | |
106 | |
3928
987fffdf6ae7
don't try to inline cabac functions. gcc ignored the hint anyway, and forcing it would make h264 slower.
lorenm
parents:
3642
diff
changeset
|
107 static void put_cabac_static(CABACContext *c, int RangeLPS, int bit){ |
1287 | 108 assert(c->range > RangeLPS); |
109 | |
110 if(!bit){ | |
111 c->range -= RangeLPS; | |
112 }else{ | |
113 c->low += c->range - RangeLPS; | |
114 c->range = RangeLPS; | |
115 } | |
116 | |
117 renorm_cabac_encoder(c); | |
118 | |
119 #ifdef STRICT_LIMITS | |
120 c->symCount++; | |
121 #endif | |
122 } | |
123 | |
1290 | 124 /** |
125 * @param bit 0 -> write zero bit, !=0 write one bit | |
126 */ | |
3928
987fffdf6ae7
don't try to inline cabac functions. gcc ignored the hint anyway, and forcing it would make h264 slower.
lorenm
parents:
3642
diff
changeset
|
127 static void put_cabac_bypass(CABACContext *c, int bit){ |
1287 | 128 c->low += c->low; |
129 | |
130 if(bit){ | |
131 c->low += c->range; | |
132 } | |
133 //FIXME optimize | |
134 if(c->low<0x200){ | |
135 put_cabac_bit(c, 0); | |
136 }else if(c->low<0x400){ | |
137 c->outstanding_count++; | |
138 c->low -= 0x200; | |
139 }else{ | |
140 put_cabac_bit(c, 1); | |
141 c->low -= 0x400; | |
142 } | |
2967 | 143 |
1287 | 144 #ifdef STRICT_LIMITS |
145 c->symCount++; | |
146 #endif | |
147 } | |
148 | |
1300
e18667d1e94d
FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents:
1298
diff
changeset
|
149 /** |
e18667d1e94d
FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents:
1298
diff
changeset
|
150 * |
e18667d1e94d
FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents:
1298
diff
changeset
|
151 * @return the number of bytes written |
e18667d1e94d
FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents:
1298
diff
changeset
|
152 */ |
3928
987fffdf6ae7
don't try to inline cabac functions. gcc ignored the hint anyway, and forcing it would make h264 slower.
lorenm
parents:
3642
diff
changeset
|
153 static int put_cabac_terminate(CABACContext *c, int bit){ |
1287 | 154 c->range -= 2; |
155 | |
156 if(!bit){ | |
157 renorm_cabac_encoder(c); | |
158 }else{ | |
159 c->low += c->range; | |
160 c->range= 2; | |
2967 | 161 |
1287 | 162 renorm_cabac_encoder(c); |
163 | |
164 assert(c->low <= 0x1FF); | |
165 put_cabac_bit(c, c->low>>9); | |
166 put_bits(&c->pb, 2, ((c->low>>7)&3)|1); | |
2967 | 167 |
1287 | 168 flush_put_bits(&c->pb); //FIXME FIXME FIXME XXX wrong |
169 } | |
2967 | 170 |
1287 | 171 #ifdef STRICT_LIMITS |
172 c->symCount++; | |
173 #endif | |
1300
e18667d1e94d
FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents:
1298
diff
changeset
|
174 |
1787 | 175 return (put_bits_count(&c->pb)+7)>>3; |
1287 | 176 } |
177 | |
1290 | 178 /** |
179 * put (truncated) unary binarization. | |
180 */ | |
3928
987fffdf6ae7
don't try to inline cabac functions. gcc ignored the hint anyway, and forcing it would make h264 slower.
lorenm
parents:
3642
diff
changeset
|
181 static void put_cabac_u(CABACContext *c, uint8_t * state, int v, int max, int max_index, int truncated){ |
1290 | 182 int i; |
2967 | 183 |
1290 | 184 assert(v <= max); |
2967 | 185 |
1290 | 186 #if 1 |
187 for(i=0; i<v; i++){ | |
188 put_cabac(c, state, 1); | |
189 if(i < max_index) state++; | |
190 } | |
191 if(truncated==0 || v<max) | |
192 put_cabac(c, state, 0); | |
193 #else | |
194 if(v <= max_index){ | |
195 for(i=0; i<v; i++){ | |
196 put_cabac(c, state+i, 1); | |
197 } | |
198 if(truncated==0 || v<max) | |
199 put_cabac(c, state+i, 0); | |
200 }else{ | |
201 for(i=0; i<=max_index; i++){ | |
202 put_cabac(c, state+i, 1); | |
203 } | |
204 for(; i<v; i++){ | |
205 put_cabac(c, state+max_index, 1); | |
206 } | |
207 if(truncated==0 || v<max) | |
208 put_cabac(c, state+max_index, 0); | |
209 } | |
210 #endif | |
211 } | |
212 | |
213 /** | |
214 * put unary exp golomb k-th order binarization. | |
215 */ | |
3928
987fffdf6ae7
don't try to inline cabac functions. gcc ignored the hint anyway, and forcing it would make h264 slower.
lorenm
parents:
3642
diff
changeset
|
216 static void put_cabac_ueg(CABACContext *c, uint8_t * state, int v, int max, int is_signed, int k, int max_index){ |
1290 | 217 int i; |
2967 | 218 |
1290 | 219 if(v==0) |
220 put_cabac(c, state, 0); | |
221 else{ | |
1298 | 222 const int sign= v < 0; |
2967 | 223 |
4001 | 224 if(is_signed) v= FFABS(v); |
2967 | 225 |
1290 | 226 if(v<max){ |
227 for(i=0; i<v; i++){ | |
228 put_cabac(c, state, 1); | |
229 if(i < max_index) state++; | |
230 } | |
231 | |
232 put_cabac(c, state, 0); | |
233 }else{ | |
234 int m= 1<<k; | |
235 | |
236 for(i=0; i<max; i++){ | |
237 put_cabac(c, state, 1); | |
238 if(i < max_index) state++; | |
239 } | |
240 | |
241 v -= max; | |
242 while(v >= m){ //FIXME optimize | |
243 put_cabac_bypass(c, 1); | |
244 v-= m; | |
245 m+= m; | |
246 } | |
247 put_cabac_bypass(c, 0); | |
248 while(m>>=1){ | |
249 put_cabac_bypass(c, v&m); | |
250 } | |
251 } | |
252 | |
253 if(is_signed) | |
254 put_cabac_bypass(c, sign); | |
255 } | |
256 } | |
257 | |
2323 | 258 static void refill(CABACContext *c){ |
259 #if CABAC_BITS == 16 | |
3946 | 260 c->low+= (c->bytestream[0]<<9) + (c->bytestream[1]<<1); |
2323 | 261 #else |
262 c->low+= c->bytestream[0]<<1; | |
263 #endif | |
264 c->low -= CABAC_MASK; | |
265 c->bytestream+= CABAC_BITS/8; | |
266 } | |
267 | |
268 static void refill2(CABACContext *c){ | |
269 int i, x; | |
270 | |
271 x= c->low ^ (c->low-1); | |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
272 i= 7 - ff_h264_norm_shift[x>>(CABAC_BITS-1)]; |
2323 | 273 |
274 x= -CABAC_MASK; | |
2967 | 275 |
2323 | 276 #if CABAC_BITS == 16 |
277 x+= (c->bytestream[0]<<9) + (c->bytestream[1]<<1); | |
278 #else | |
279 x+= c->bytestream[0]<<1; | |
280 #endif | |
2967 | 281 |
2323 | 282 c->low += x<<i; |
283 c->bytestream+= CABAC_BITS/8; | |
284 } | |
285 | |
1287 | 286 static inline void renorm_cabac_decoder(CABACContext *c){ |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
287 while(c->range < 0x100){ |
1287 | 288 c->range+= c->range; |
289 c->low+= c->low; | |
2323 | 290 if(!(c->low & CABAC_MASK)) |
291 refill(c); | |
1287 | 292 } |
293 } | |
294 | |
2323 | 295 static inline void renorm_cabac_decoder_once(CABACContext *c){ |
3951 | 296 #ifdef ARCH_X86_DISABLED |
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
297 int temp; |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
298 #if 0 |
3950
900d21b85dd6
renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents:
3948
diff
changeset
|
299 //P3:683 athlon:475 |
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
300 asm( |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
301 "lea -0x100(%0), %2 \n\t" |
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
302 "shr $31, %2 \n\t" //FIXME 31->63 for x86-64 |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
303 "shl %%cl, %0 \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
304 "shl %%cl, %1 \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
305 : "+r"(c->range), "+r"(c->low), "+c"(temp) |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
306 ); |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
307 #elif 0 |
3950
900d21b85dd6
renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents:
3948
diff
changeset
|
308 //P3:680 athlon:474 |
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
309 asm( |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
310 "cmp $0x100, %0 \n\t" |
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
311 "setb %%cl \n\t" //FIXME 31->63 for x86-64 |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
312 "shl %%cl, %0 \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
313 "shl %%cl, %1 \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
314 : "+r"(c->range), "+r"(c->low), "+c"(temp) |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
315 ); |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
316 #elif 1 |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
317 int temp2; |
3950
900d21b85dd6
renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents:
3948
diff
changeset
|
318 //P3:665 athlon:517 |
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
319 asm( |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
320 "lea -0x100(%0), %%eax \n\t" |
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
321 "cdq \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
322 "mov %0, %%eax \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
323 "and %%edx, %0 \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
324 "and %1, %%edx \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
325 "add %%eax, %0 \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
326 "add %%edx, %1 \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
327 : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2) |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
328 ); |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
329 #elif 0 |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
330 int temp2; |
3950
900d21b85dd6
renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents:
3948
diff
changeset
|
331 //P3:673 athlon:509 |
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
332 asm( |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
333 "cmp $0x100, %0 \n\t" |
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
334 "sbb %%edx, %%edx \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
335 "mov %0, %%eax \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
336 "and %%edx, %0 \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
337 "and %1, %%edx \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
338 "add %%eax, %0 \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
339 "add %%edx, %1 \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
340 : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2) |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
341 ); |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
342 #else |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
343 int temp2; |
3950
900d21b85dd6
renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents:
3948
diff
changeset
|
344 //P3:677 athlon:511 |
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
345 asm( |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
346 "cmp $0x100, %0 \n\t" |
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
347 "lea (%0, %0), %%eax \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
348 "lea (%1, %1), %%edx \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
349 "cmovb %%eax, %0 \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
350 "cmovb %%edx, %1 \n\t" |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
351 : "+r"(c->range), "+r"(c->low), "+a"(temp), "+d"(temp2) |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
352 ); |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
353 #endif |
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
354 #else |
3950
900d21b85dd6
renorm_cabac_decoder_once START/STOP_TIMER scores for athlon
michael
parents:
3948
diff
changeset
|
355 //P3:675 athlon:476 |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
356 int shift= (uint32_t)(c->range - 0x100)>>31; |
3642 | 357 c->range<<= shift; |
358 c->low <<= shift; | |
3943
811a9b0d9f32
several x86 renorm_cabac_decoder_once optimizations
michael
parents:
3928
diff
changeset
|
359 #endif |
2323 | 360 if(!(c->low & CABAC_MASK)) |
361 refill(c); | |
362 } | |
363 | |
4008
b636f3d59283
prevent "mb level" get_cabac() calls from being inlined (3% faster decode_mb_cabac() on P3)
michael
parents:
4002
diff
changeset
|
364 static int always_inline get_cabac_inline(CABACContext *c, uint8_t * const state){ |
3642 | 365 //FIXME gcc generates duplicate load/stores for c->low and c->range |
4026 | 366 #if defined(ARCH_X86) && !(defined(PIC) && defined(__GNUC__)) |
3969
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
367 int bit; |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
368 |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
369 #define LOW "0" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
370 #define RANGE "4" |
3993
8b7c59b7af01
make state transition tables global as they are constant and the code is slightly faster that way
michael
parents:
3992
diff
changeset
|
371 #define BYTESTART "12" |
8b7c59b7af01
make state transition tables global as they are constant and the code is slightly faster that way
michael
parents:
3992
diff
changeset
|
372 #define BYTE "16" |
8b7c59b7af01
make state transition tables global as they are constant and the code is slightly faster that way
michael
parents:
3992
diff
changeset
|
373 #define BYTEEND "20" |
3984 | 374 #ifndef BRANCHLESS_CABAC_DECODER |
3969
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
375 asm volatile( |
4035 | 376 "movzbl (%1), %0 \n\t" |
3969
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
377 "movl "RANGE "(%2), %%ebx \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
378 "movl "RANGE "(%2), %%edx \n\t" |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
379 "andl $0xC0, %%ebx \n\t" |
4035 | 380 "movzbl "MANGLE(ff_h264_lps_range)"(%0, %%ebx, 2), %%esi\n\t" |
3969
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
381 "movl "LOW "(%2), %%ebx \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
382 //eax:state ebx:low, edx:range, esi:RangeLPS |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
383 "subl %%esi, %%edx \n\t" |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
384 "movl %%edx, %%ecx \n\t" |
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
385 "shll $17, %%ecx \n\t" |
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
386 "cmpl %%ecx, %%ebx \n\t" |
3969
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
387 " ja 1f \n\t" |
3999
6cbad3675632
slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents:
3996
diff
changeset
|
388 |
6cbad3675632
slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents:
3996
diff
changeset
|
389 #if 1 |
6cbad3675632
slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents:
3996
diff
changeset
|
390 //athlon:4067 P3:4110 |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
391 "lea -0x100(%%edx), %%ecx \n\t" |
3999
6cbad3675632
slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents:
3996
diff
changeset
|
392 "shr $31, %%ecx \n\t" |
6cbad3675632
slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents:
3996
diff
changeset
|
393 "shl %%cl, %%edx \n\t" |
6cbad3675632
slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents:
3996
diff
changeset
|
394 "shl %%cl, %%ebx \n\t" |
6cbad3675632
slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents:
3996
diff
changeset
|
395 #else |
6cbad3675632
slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents:
3996
diff
changeset
|
396 //athlon:4057 P3:4130 |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
397 "cmp $0x100, %%edx \n\t" //FIXME avoidable |
3969
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
398 "setb %%cl \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
399 "shl %%cl, %%edx \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
400 "shl %%cl, %%ebx \n\t" |
3999
6cbad3675632
slightly faster on P3 slightly slower on athlon and probably faster on P4
michael
parents:
3996
diff
changeset
|
401 #endif |
4035 | 402 "movzbl "MANGLE(ff_h264_mps_state)"(%0), %%ecx \n\t" |
3969
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
403 "movb %%cl, (%1) \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
404 //eax:state ebx:low, edx:range, esi:RangeLPS |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
405 "test %%bx, %%bx \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
406 " jnz 2f \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
407 "movl "BYTE "(%2), %%esi \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
408 "subl $0xFFFF, %%ebx \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
409 "movzwl (%%esi), %%ecx \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
410 "bswap %%ecx \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
411 "shrl $15, %%ecx \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
412 "addl $2, %%esi \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
413 "addl %%ecx, %%ebx \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
414 "movl %%esi, "BYTE "(%2) \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
415 "jmp 2f \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
416 "1: \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
417 //eax:state ebx:low, edx:range, esi:RangeLPS |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
418 "subl %%ecx, %%ebx \n\t" |
3969
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
419 "movl %%esi, %%edx \n\t" |
3979
ce16f66a48ad
reading 8bit mem into a 8bit register needs 2 uops on P4, 8bit->32bit with zero extension needs just 1
michael
parents:
3978
diff
changeset
|
420 "movzbl " MANGLE(ff_h264_norm_shift) "(%%esi), %%ecx \n\t" |
3969
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
421 "shll %%cl, %%ebx \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
422 "shll %%cl, %%edx \n\t" |
4035 | 423 "movzbl "MANGLE(ff_h264_lps_state)"(%0), %%ecx \n\t" |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
424 "movb %%cl, (%1) \n\t" |
4035 | 425 "addl $1, %0 \n\t" |
3969
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
426 "test %%bx, %%bx \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
427 " jnz 2f \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
428 |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
429 "movl "BYTE "(%2), %%ecx \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
430 "movzwl (%%ecx), %%esi \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
431 "bswap %%esi \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
432 "shrl $15, %%esi \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
433 "subl $0xFFFF, %%esi \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
434 "addl $2, %%ecx \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
435 "movl %%ecx, "BYTE "(%2) \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
436 |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
437 "leal -1(%%ebx), %%ecx \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
438 "xorl %%ebx, %%ecx \n\t" |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
439 "shrl $15, %%ecx \n\t" |
3979
ce16f66a48ad
reading 8bit mem into a 8bit register needs 2 uops on P4, 8bit->32bit with zero extension needs just 1
michael
parents:
3978
diff
changeset
|
440 "movzbl " MANGLE(ff_h264_norm_shift) "(%%ecx), %%ecx \n\t" |
3994
2734b228fc87
use ecx instead of cl (no speed change on P3 but might avoid partial register stalls on some cpus)
michael
parents:
3993
diff
changeset
|
441 "neg %%ecx \n\t" |
2734b228fc87
use ecx instead of cl (no speed change on P3 but might avoid partial register stalls on some cpus)
michael
parents:
3993
diff
changeset
|
442 "add $7, %%ecx \n\t" |
3969
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
443 |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
444 "shll %%cl , %%esi \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
445 "addl %%esi, %%ebx \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
446 "2: \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
447 "movl %%edx, "RANGE "(%2) \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
448 "movl %%ebx, "LOW "(%2) \n\t" |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
449 :"=&a"(bit) //FIXME this is fragile gcc either runs out of registers or misscompiles it (for example if "+a"(bit) or "+m"(*state) is used |
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
450 :"r"(state), "r"(c) |
4012
f8c649ac09dd
add "memory" to the clobber list we change memory so we need it, this also fixes some problems with gcc svn
michael
parents:
4008
diff
changeset
|
451 : "%ecx", "%ebx", "%edx", "%esi", "memory" |
3969
fc6e0942353b
first try of a handwritten get_cabac() for x86, this is 10-20% faster on P3 depening on if you try to subtract the START/STOP_TIMER overhead
michael
parents:
3967
diff
changeset
|
452 ); |
3982
af16271634c2
moving another bit&1 out, this is as fast as with it in there, but it makes more sense with it outside of the loop
michael
parents:
3981
diff
changeset
|
453 bit&=1; |
4002
ec426fa57dfe
adds some useful comments after some of the #else, #elseif,
gpoirier
parents:
4001
diff
changeset
|
454 #else /* BRANCHLESS_CABAC_DECODER */ |
3975 | 455 asm volatile( |
4035 | 456 "movzbl (%1), %0 \n\t" |
3975 | 457 "movl "RANGE "(%2), %%ebx \n\t" |
458 "movl "RANGE "(%2), %%edx \n\t" | |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
459 "andl $0xC0, %%ebx \n\t" |
4035 | 460 "movzbl "MANGLE(ff_h264_lps_range)"(%0, %%ebx, 2), %%esi\n\t" |
3975 | 461 "movl "LOW "(%2), %%ebx \n\t" |
462 //eax:state ebx:low, edx:range, esi:RangeLPS | |
463 "subl %%esi, %%edx \n\t" | |
4034
fbe263601152
Fix crash with illegal instruction, cmov is available on 686 and later only.
diego
parents:
4033
diff
changeset
|
464 #if (defined CMOV_IS_FAST && __CPU__ >= 686) |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
465 "movl %%edx, %%ecx \n\t" |
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
466 "shl $17, %%edx \n\t" |
3980 | 467 "cmpl %%ebx, %%edx \n\t" |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
468 "cmova %%ecx, %%esi \n\t" |
3980 | 469 "sbbl %%ecx, %%ecx \n\t" |
470 "andl %%ecx, %%edx \n\t" | |
471 "subl %%edx, %%ebx \n\t" | |
4035 | 472 "xorl %%ecx, %0 \n\t" |
4002
ec426fa57dfe
adds some useful comments after some of the #else, #elseif,
gpoirier
parents:
4001
diff
changeset
|
473 #else /* CMOV_IS_FAST */ |
3975 | 474 "movl %%edx, %%ecx \n\t" |
4027
ca4a837b896d
fix !CMOV_IS_FAST case (iam not really happy with the fix but i didnt come up with a better one quickly)
michael
parents:
4026
diff
changeset
|
475 "shl $17, %%edx \n\t" |
3975 | 476 "subl %%ebx, %%edx \n\t" |
477 "sarl $31, %%edx \n\t" //lps_mask | |
478 "subl %%ecx, %%esi \n\t" //RangeLPS - range | |
479 "andl %%edx, %%esi \n\t" //(RangeLPS - range)&lps_mask | |
480 "addl %%ecx, %%esi \n\t" //new range | |
4027
ca4a837b896d
fix !CMOV_IS_FAST case (iam not really happy with the fix but i didnt come up with a better one quickly)
michael
parents:
4026
diff
changeset
|
481 "shl $17, %%ecx \n\t" |
3975 | 482 "andl %%edx, %%ecx \n\t" |
483 "subl %%ecx, %%ebx \n\t" | |
4035 | 484 "xorl %%edx, %0 \n\t" |
4002
ec426fa57dfe
adds some useful comments after some of the #else, #elseif,
gpoirier
parents:
4001
diff
changeset
|
485 #endif /* CMOV_IS_FAST */ |
3975 | 486 |
487 //eax:state ebx:low edx:mask esi:range | |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
488 |
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
489 //eax:bit ebx:low esi:range |
3975 | 490 |
3979
ce16f66a48ad
reading 8bit mem into a 8bit register needs 2 uops on P4, 8bit->32bit with zero extension needs just 1
michael
parents:
3978
diff
changeset
|
491 "movzbl " MANGLE(ff_h264_norm_shift) "(%%esi), %%ecx \n\t" |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
492 "shll %%cl, %%esi \n\t" |
4035 | 493 "movzbl "MANGLE(ff_h264_mlps_state)"+128(%0), %%edx \n\t" |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
494 "movb %%dl, (%1) \n\t" |
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
495 "movl %%esi, "RANGE "(%2) \n\t" |
3975 | 496 "shll %%cl, %%ebx \n\t" |
3995
b00c06477dff
write cabac low and range variables as early as possible to prevent stalls from reading them before they where written, the P4 is said to disslike that alot, on P3 its 2% faster (START/STOP_TIMER over decode_residual)
michael
parents:
3994
diff
changeset
|
497 "movl %%ebx, "LOW "(%2) \n\t" |
3975 | 498 "test %%bx, %%bx \n\t" |
499 " jnz 1f \n\t" | |
500 | |
501 "movl "BYTE "(%2), %%ecx \n\t" | |
502 "movzwl (%%ecx), %%esi \n\t" | |
503 "bswap %%esi \n\t" | |
504 "shrl $15, %%esi \n\t" | |
505 "subl $0xFFFF, %%esi \n\t" | |
506 "addl $2, %%ecx \n\t" | |
507 "movl %%ecx, "BYTE "(%2) \n\t" | |
508 | |
509 "leal -1(%%ebx), %%ecx \n\t" | |
510 "xorl %%ebx, %%ecx \n\t" | |
4024
d550343b5dac
shift CABACContext.range right, this reduces the number of shifts needed in get_cabac() and is slightly faster on P3 (and should be much faster on P4 as the P4 except the more recent variants lacks an integer shifter and so shifts have ~10 times longer latency then simple operations like adds)
michael
parents:
4014
diff
changeset
|
511 "shrl $15, %%ecx \n\t" |
3979
ce16f66a48ad
reading 8bit mem into a 8bit register needs 2 uops on P4, 8bit->32bit with zero extension needs just 1
michael
parents:
3978
diff
changeset
|
512 "movzbl " MANGLE(ff_h264_norm_shift) "(%%ecx), %%ecx \n\t" |
3994
2734b228fc87
use ecx instead of cl (no speed change on P3 but might avoid partial register stalls on some cpus)
michael
parents:
3993
diff
changeset
|
513 "neg %%ecx \n\t" |
2734b228fc87
use ecx instead of cl (no speed change on P3 but might avoid partial register stalls on some cpus)
michael
parents:
3993
diff
changeset
|
514 "add $7, %%ecx \n\t" |
3975 | 515 |
516 "shll %%cl , %%esi \n\t" | |
517 "addl %%esi, %%ebx \n\t" | |
3995
b00c06477dff
write cabac low and range variables as early as possible to prevent stalls from reading them before they where written, the P4 is said to disslike that alot, on P3 its 2% faster (START/STOP_TIMER over decode_residual)
michael
parents:
3994
diff
changeset
|
518 "movl %%ebx, "LOW "(%2) \n\t" |
3975 | 519 "1: \n\t" |
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 int bit; | |
628 asm volatile( | |
629 "movl "RANGE "(%1), %%ebx \n\t" | |
630 "movl "LOW "(%1), %%eax \n\t" | |
631 "shl $17, %%ebx \n\t" | |
632 "add %%eax, %%eax \n\t" | |
633 "sub %%ebx, %%eax \n\t" | |
634 "cdq \n\t" | |
635 "and %%edx, %%ebx \n\t" | |
636 "add %%ebx, %%eax \n\t" | |
637 "xor %%edx, %%ecx \n\t" | |
638 "sub %%edx, %%ecx \n\t" | |
639 "test %%ax, %%ax \n\t" | |
640 " jnz 1f \n\t" | |
641 "movl "BYTE "(%1), %%ebx \n\t" | |
642 "subl $0xFFFF, %%eax \n\t" | |
643 "movzwl (%%ebx), %%edx \n\t" | |
644 "bswap %%edx \n\t" | |
645 "shrl $15, %%edx \n\t" | |
646 "addl $2, %%ebx \n\t" | |
647 "addl %%edx, %%eax \n\t" | |
648 "movl %%ebx, "BYTE "(%1) \n\t" | |
649 "1: \n\t" | |
650 "movl %%eax, "LOW "(%1) \n\t" | |
651 | |
652 :"+c"(val) | |
653 :"r"(c) | |
654 : "%eax", "%ebx", "%edx", "memory" | |
655 ); | |
656 return val; | |
657 #else | |
658 int range, mask; | |
659 c->low += c->low; | |
660 | |
661 if(!(c->low & CABAC_MASK)) | |
662 refill(c); | |
663 | |
664 range= c->range<<17; | |
665 c->low -= range; | |
666 mask= c->low >> 31; | |
667 range &= mask; | |
668 c->low += range; | |
669 return (val^mask)-mask; | |
670 #endif | |
671 } | |
672 | |
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
|
673 //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
|
674 //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
|
675 #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
|
676 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
|
677 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
|
678 int minusstart= -(int)significant_coeff_ctx_base; |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
679 int minusindex= -(int)index; |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
680 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
|
681 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
|
682 "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
|
683 "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
|
684 |
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 "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
|
686 |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
687 "movzbl (%1), %0 \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
688 "movl %%esi, %%edx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
689 "andl $0xC0, %%esi \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
690 "movzbl "MANGLE(ff_h264_lps_range)"(%0, %%esi, 2), %%esi\n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
691 /*eax:state ebx:low, edx:range, esi:RangeLPS*/ |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
692 "subl %%esi, %%edx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
693 |
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 #if (defined CMOV_IS_FAST && __CPU__ >= 686) |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
695 "movl %%edx, %%ecx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
696 "shl $17, %%edx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
697 "cmpl %%ebx, %%edx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
698 "cmova %%ecx, %%esi \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
699 "sbbl %%ecx, %%ecx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
700 "andl %%ecx, %%edx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
701 "subl %%edx, %%ebx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
702 "xorl %%ecx, %0 \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
703 #else /* CMOV_IS_FAST */ |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
704 "movl %%edx, %%ecx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
705 "shl $17, %%edx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
706 "subl %%ebx, %%edx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
707 "sarl $31, %%edx \n\t" //lps_mask |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
708 "subl %%ecx, %%esi \n\t" //RangeLPS - range |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
709 "andl %%edx, %%esi \n\t" //(RangeLPS - range)&lps_mask |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
710 "addl %%ecx, %%esi \n\t" //new range |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
711 "shl $17, %%ecx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
712 "andl %%edx, %%ecx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
713 "subl %%ecx, %%ebx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
714 "xorl %%edx, %0 \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
715 #endif /* CMOV_IS_FAST */ |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
716 |
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 "movzbl " MANGLE(ff_h264_norm_shift) "(%%esi), %%ecx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
718 "shll %%cl, %%esi \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
719 "movzbl "MANGLE(ff_h264_mlps_state)"+128(%0), %%edx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
720 "movb %%dl, (%1) \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
721 "shll %%cl, %%ebx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
722 "test %%bx, %%bx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
723 " jnz 1f \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
724 |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
725 "movl "BYTE "(%3), %%ecx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
726 "movzwl (%%ecx), %%edx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
727 "bswap %%edx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
728 "shrl $15, %%edx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
729 "subl $0xFFFF, %%edx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
730 "addl $2, %%ecx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
731 "movl %%ecx, "BYTE "(%3) \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
732 |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
733 "leal -1(%%ebx), %%ecx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
734 "xorl %%ebx, %%ecx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
735 "shrl $15, %%ecx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
736 "movzbl " MANGLE(ff_h264_norm_shift) "(%%ecx), %%ecx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
737 "neg %%ecx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
738 "add $7, %%ecx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
739 |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
740 "shll %%cl , %%edx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
741 "addl %%edx, %%ebx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
742 "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
|
743 |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
744 "test $1, %0 \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
745 " 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
|
746 |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
747 "movl %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
|
748 "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
|
749 "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
|
750 "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
|
751 "addl $4, %%eax \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
752 "movl %%eax, %2 \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
753 |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
754 "movzbl 61(%1), %0 \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
755 "movl %%esi, %%edx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
756 "andl $0xC0, %%esi \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
757 "movzbl "MANGLE(ff_h264_lps_range)"(%0, %%esi, 2), %%esi\n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
758 /*eax:state ebx:low, edx:range, esi:RangeLPS*/ |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
759 "subl %%esi, %%edx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
760 |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
761 #if (defined CMOV_IS_FAST && __CPU__ >= 686) |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
762 "movl %%edx, %%ecx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
763 "shl $17, %%edx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
764 "cmpl %%ebx, %%edx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
765 "cmova %%ecx, %%esi \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
766 "sbbl %%ecx, %%ecx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
767 "andl %%ecx, %%edx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
768 "subl %%edx, %%ebx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
769 "xorl %%ecx, %0 \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
770 #else /* CMOV_IS_FAST */ |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
771 "movl %%edx, %%ecx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
772 "shl $17, %%edx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
773 "subl %%ebx, %%edx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
774 "sarl $31, %%edx \n\t" //lps_mask |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
775 "subl %%ecx, %%esi \n\t" //RangeLPS - range |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
776 "andl %%edx, %%esi \n\t" //(RangeLPS - range)&lps_mask |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
777 "addl %%ecx, %%esi \n\t" //new range |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
778 "shl $17, %%ecx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
779 "andl %%edx, %%ecx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
780 "subl %%ecx, %%ebx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
781 "xorl %%edx, %0 \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
782 #endif /* CMOV_IS_FAST */ |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
783 |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
784 "movzbl " MANGLE(ff_h264_norm_shift) "(%%esi), %%ecx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
785 "shll %%cl, %%esi \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
786 "movzbl "MANGLE(ff_h264_mlps_state)"+128(%0), %%edx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
787 "movb %%dl, 61(%1) \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
788 "shll %%cl, %%ebx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
789 "test %%bx, %%bx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
790 " jnz 1f \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
791 |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
792 "movl "BYTE "(%3), %%ecx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
793 "movzwl (%%ecx), %%edx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
794 "bswap %%edx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
795 "shrl $15, %%edx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
796 "subl $0xFFFF, %%edx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
797 "addl $2, %%ecx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
798 "movl %%ecx, "BYTE "(%3) \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
799 |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
800 "leal -1(%%ebx), %%ecx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
801 "xorl %%ebx, %%ecx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
802 "shrl $15, %%ecx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
803 "movzbl " MANGLE(ff_h264_norm_shift) "(%%ecx), %%ecx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
804 "neg %%ecx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
805 "add $7, %%ecx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
806 |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
807 "shll %%cl , %%edx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
808 "addl %%edx, %%ebx \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
809 "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
|
810 |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
811 "test $1, %%eax \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
812 " 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
|
813 |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
814 "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
|
815 "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
|
816 "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
|
817 " 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
|
818 "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
|
819 "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
|
820 "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
|
821 "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
|
822 "addl $4, %%eax \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
823 "movl %%eax, %2 \n\t" |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
824 "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
|
825 "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
|
826 "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
|
827 "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
|
828 |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
829 "movl %%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
|
830 "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
|
831 :"=&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
|
832 :"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
|
833 : "%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
|
834 ); |
53be304c7f54
x86 asm version of the decode significance loop (not 8x8) of decode_residual() 5% faster decode_residual() on P3
michael
parents:
4035
diff
changeset
|
835 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
|
836 } |
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
|
837 #endif |
1287 | 838 |
1300
e18667d1e94d
FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents:
1298
diff
changeset
|
839 /** |
e18667d1e94d
FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents:
1298
diff
changeset
|
840 * |
e18667d1e94d
FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents:
1298
diff
changeset
|
841 * @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
|
842 */ |
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
|
843 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
|
844 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
|
845 if(c->low < c->range<<17){ |
2323 | 846 renorm_cabac_decoder_once(c); |
1287 | 847 return 0; |
848 }else{ | |
1300
e18667d1e94d
FFV1 codec (our very simple lossless intra only codec, compresses much better then huffyuv)
michaelni
parents:
1298
diff
changeset
|
849 return c->bytestream - c->bytestream_start; |
2967 | 850 } |
1287 | 851 } |
852 | |
1290 | 853 /** |
854 * get (truncated) unnary binarization. | |
855 */ | |
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
|
856 static int get_cabac_u(CABACContext *c, uint8_t * state, int max, int max_index, int truncated){ |
1290 | 857 int i; |
2967 | 858 |
859 for(i=0; i<max; i++){ | |
1290 | 860 if(get_cabac(c, state)==0) |
861 return i; | |
2967 | 862 |
1290 | 863 if(i< max_index) state++; |
864 } | |
865 | |
866 return truncated ? max : -1; | |
867 } | |
868 | |
869 /** | |
870 * get unary exp golomb k-th order binarization. | |
871 */ | |
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
|
872 static int get_cabac_ueg(CABACContext *c, uint8_t * state, int max, int is_signed, int k, int max_index){ |
1290 | 873 int i, v; |
874 int m= 1<<k; | |
2967 | 875 |
876 if(get_cabac(c, state)==0) | |
1290 | 877 return 0; |
2967 | 878 |
1290 | 879 if(0 < max_index) state++; |
2967 | 880 |
881 for(i=1; i<max; i++){ | |
1290 | 882 if(get_cabac(c, state)==0){ |
883 if(is_signed && get_cabac_bypass(c)){ | |
884 return -i; | |
885 }else | |
886 return i; | |
887 } | |
888 | |
889 if(i < max_index) state++; | |
890 } | |
2967 | 891 |
1290 | 892 while(get_cabac_bypass(c)){ |
893 i+= m; | |
894 m+= m; | |
895 } | |
2967 | 896 |
1290 | 897 v=0; |
898 while(m>>=1){ | |
899 v+= v + get_cabac_bypass(c); | |
900 } | |
901 i += v; | |
902 | |
903 if(is_signed && get_cabac_bypass(c)){ | |
904 return -i; | |
905 }else | |
906 return i; | |
907 } |