Mercurial > libavcodec.hg
annotate common.h @ 517:389e30fe7269 libavcodec
* Move "restrict" fallback to common.h
* Don't include config.h in dsputil.c
author | mellum |
---|---|
date | Thu, 04 Jul 2002 13:41:27 +0000 |
parents | ba67fefada47 |
children | 19a5e2a81e1a |
rev | line source |
---|---|
0 | 1 #ifndef COMMON_H |
2 #define COMMON_H | |
3 | |
86 | 4 #define FFMPEG_VERSION_INT 0x000406 |
5 #define FFMPEG_VERSION "0.4.6" | |
64 | 6 |
213
e80ad397d30e
Cygwin's mangling by Felix Buenemann <atmosfear@users.sourceforge.net>
nickols_k
parents:
199
diff
changeset
|
7 #if defined(WIN32) && !defined(__MINGW32__) && !defined(__CYGWIN__) |
64 | 8 #define CONFIG_WIN32 |
9 #endif | |
10 | |
234
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
11 //#define ALT_BITSTREAM_WRITER |
235
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
12 //#define ALIGNED_BITSTREAM_WRITER |
512
ba67fefada47
Activate ALT_BITSTREAM_READER by default on Alpha, since it seems to
mellum
parents:
496
diff
changeset
|
13 |
ba67fefada47
Activate ALT_BITSTREAM_READER by default on Alpha, since it seems to
mellum
parents:
496
diff
changeset
|
14 #ifdef ARCH_ALPHA |
ba67fefada47
Activate ALT_BITSTREAM_READER by default on Alpha, since it seems to
mellum
parents:
496
diff
changeset
|
15 #define ALT_BITSTREAM_READER |
ba67fefada47
Activate ALT_BITSTREAM_READER by default on Alpha, since it seems to
mellum
parents:
496
diff
changeset
|
16 #endif |
ba67fefada47
Activate ALT_BITSTREAM_READER by default on Alpha, since it seems to
mellum
parents:
496
diff
changeset
|
17 |
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
18 //#define ALIGNED_BITSTREAM |
199
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
19 #define FAST_GET_FIRST_VLC |
306 | 20 //#define DUMP_STREAM // only works with the ALT_BITSTREAM_READER |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
21 |
10 | 22 #ifdef HAVE_AV_CONFIG_H |
64 | 23 /* only include the following when compiling package */ |
408 | 24 #include "config.h" |
64 | 25 |
26 #include <stdlib.h> | |
27 #include <stdio.h> | |
28 #include <string.h> | |
29 #include <errno.h> | |
390
48e08d9871da
added proper memory handling functions - fixed include paths
glantau
parents:
370
diff
changeset
|
30 #include <math.h> |
64 | 31 |
32 #ifndef ENODATA | |
33 #define ENODATA 61 | |
10 | 34 #endif |
0 | 35 |
390
48e08d9871da
added proper memory handling functions - fixed include paths
glantau
parents:
370
diff
changeset
|
36 #endif /* HAVE_AV_CONFIG_H */ |
64 | 37 |
517 | 38 /* Suppress restrict if it was not defined in config.h. */ |
39 #ifndef restrict | |
40 #define restrict | |
41 #endif | |
42 | |
64 | 43 #ifdef CONFIG_WIN32 |
44 | |
45 /* windows */ | |
46 | |
47 typedef unsigned short UINT16; | |
48 typedef signed short INT16; | |
49 typedef unsigned char UINT8; | |
50 typedef unsigned int UINT32; | |
51 typedef unsigned __int64 UINT64; | |
52 typedef signed char INT8; | |
53 typedef signed int INT32; | |
54 typedef signed __int64 INT64; | |
55 | |
56 typedef UINT8 uint8_t; | |
57 typedef INT8 int8_t; | |
58 typedef UINT16 uint16_t; | |
59 typedef INT16 int16_t; | |
60 typedef UINT32 uint32_t; | |
61 typedef INT32 int32_t; | |
426 | 62 typedef UINT64 uint64_t; |
63 typedef INT64 int64_t; | |
64 | 64 |
76 | 65 #ifndef __MINGW32__ |
64 | 66 #define INT64_C(c) (c ## i64) |
67 #define UINT64_C(c) (c ## i64) | |
68 | |
69 #define inline __inline | |
70 | |
76 | 71 #else |
72 #define INT64_C(c) (c ## LL) | |
73 #define UINT64_C(c) (c ## ULL) | |
74 #endif /* __MINGW32__ */ | |
75 | |
64 | 76 #define M_PI 3.14159265358979323846 |
77 #define M_SQRT2 1.41421356237309504880 /* sqrt(2) */ | |
78 | |
79 #ifdef _DEBUG | |
80 #define DEBUG | |
81 #endif | |
82 | |
83 #define snprintf _snprintf | |
84 | |
426 | 85 #else /* CONFIG_WIN32 */ |
64 | 86 |
87 /* unix */ | |
88 | |
496
09dbc9e0e6a5
reverted to inttypes.h since it gives problems on some unixes
bellard
parents:
488
diff
changeset
|
89 #include <inttypes.h> |
64 | 90 |
2 | 91 #ifndef __WINE_WINDEF16_H |
92 /* workaround for typedef conflict in MPlayer (wine typedefs) */ | |
0 | 93 typedef unsigned short UINT16; |
94 typedef signed short INT16; | |
95 #endif | |
96 | |
97 typedef unsigned char UINT8; | |
98 typedef unsigned int UINT32; | |
99 typedef unsigned long long UINT64; | |
100 typedef signed char INT8; | |
101 typedef signed int INT32; | |
102 typedef signed long long INT64; | |
103 | |
64 | 104 #ifdef HAVE_AV_CONFIG_H |
105 | |
106 #ifdef __FreeBSD__ | |
107 #include <sys/param.h> | |
108 #endif | |
109 | |
110 #ifndef INT64_C | |
111 #define INT64_C(c) (c ## LL) | |
112 #define UINT64_C(c) (c ## ULL) | |
113 #endif | |
114 | |
115 #ifdef USE_FASTMEMCPY | |
116 #include "fastmemcpy.h" | |
117 #endif | |
118 | |
76 | 119 #endif /* HAVE_AV_CONFIG_H */ |
120 | |
121 #endif /* !CONFIG_WIN32 */ | |
122 | |
488 | 123 #ifdef HAVE_AV_CONFIG_H |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
124 |
488 | 125 #include "bswap.h" |
426 | 126 |
432 | 127 #if defined(__MINGW32__) || defined(__CYGWIN__) || \ |
128 defined(__OS2__) || defined (__OpenBSD__) | |
129 #define MANGLE(a) "_" #a | |
130 #else | |
131 #define MANGLE(a) #a | |
132 #endif | |
133 | |
76 | 134 /* debug stuff */ |
135 | |
64 | 136 #ifndef DEBUG |
137 #define NDEBUG | |
138 #endif | |
139 #include <assert.h> | |
140 | |
76 | 141 /* dprintf macros */ |
142 #if defined(CONFIG_WIN32) && !defined(__MINGW32__) | |
143 | |
144 inline void dprintf(const char* fmt,...) {} | |
145 | |
146 #else | |
147 | |
148 #ifdef DEBUG | |
149 #define dprintf(fmt,args...) printf(fmt, ## args) | |
150 #else | |
151 #define dprintf(fmt,args...) | |
152 #endif | |
64 | 153 |
154 #endif /* !CONFIG_WIN32 */ | |
155 | |
423
ed6098c4216a
Add an av_abort macro that aborts, but also prints out the location of the abort.
philipjsg
parents:
408
diff
changeset
|
156 #define av_abort() do { fprintf(stderr, "Abort at %s:%d\n", __FILE__, __LINE__); abort(); } while (0) |
ed6098c4216a
Add an av_abort macro that aborts, but also prints out the location of the abort.
philipjsg
parents:
408
diff
changeset
|
157 |
344 | 158 /* assume b>0 */ |
159 #define ROUNDED_DIV(a,b) (((a)>0 ? (a) + ((b)>>1) : (a) - ((b)>>1))/(b)) | |
359 | 160 #define ABS(a) ((a) >= 0 ? (a) : (-(a))) |
451 | 161 #define MAX(a,b) ((a) > (b) ? (a) : (b)) |
162 #define MIN(a,b) ((a) > (b) ? (b) : (a)) | |
344 | 163 |
0 | 164 /* bit output */ |
165 | |
166 struct PutBitContext; | |
167 | |
168 typedef void (*WriteDataFunc)(void *, UINT8 *, int); | |
169 | |
170 typedef struct PutBitContext { | |
234
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
171 #ifdef ALT_BITSTREAM_WRITER |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
172 UINT8 *buf, *buf_end; |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
173 int index; |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
174 #else |
20
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
175 UINT32 bit_buf; |
238
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
176 int bit_left; |
20
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
177 UINT8 *buf, *buf_ptr, *buf_end; |
234
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
178 #endif |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
179 INT64 data_out_size; /* in bytes */ |
0 | 180 } PutBitContext; |
181 | |
182 void init_put_bits(PutBitContext *s, | |
183 UINT8 *buffer, int buffer_size, | |
184 void *opaque, | |
185 void (*write_data)(void *, UINT8 *, int)); | |
234
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
186 |
85 | 187 INT64 get_bit_count(PutBitContext *s); /* XXX: change function name */ |
0 | 188 void align_put_bits(PutBitContext *s); |
189 void flush_put_bits(PutBitContext *s); | |
358 | 190 void put_string(PutBitContext * pbc, char *s); |
0 | 191 |
192 /* jpeg specific put_bits */ | |
193 void jflush_put_bits(PutBitContext *s); | |
194 | |
195 /* bit input */ | |
196 | |
197 typedef struct GetBitContext { | |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
198 #ifdef ALT_BITSTREAM_READER |
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
199 int index; |
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
200 UINT8 *buffer; |
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
201 #else |
20
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
202 UINT32 bit_buf; |
0 | 203 int bit_cnt; |
20
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
204 UINT8 *buf, *buf_ptr, *buf_end; |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
205 #endif |
277
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
206 int size; |
0 | 207 } GetBitContext; |
208 | |
290 | 209 static inline int get_bits_count(GetBitContext *s); |
210 | |
0 | 211 typedef struct VLC { |
212 int bits; | |
213 INT16 *table_codes; | |
214 INT8 *table_bits; | |
215 int table_size, table_allocated; | |
216 } VLC; | |
217 | |
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
218 /* used to avoid missaligned exceptions on some archs (alpha, ...) */ |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
219 #ifdef ARCH_X86 |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
220 #define unaligned32(a) (*(UINT32*)(a)) |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
221 #else |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
222 #ifdef __GNUC__ |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
223 static inline uint32_t unaligned32(const void *v) { |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
224 struct Unaligned { |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
225 uint32_t i; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
226 } __attribute__((packed)); |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
227 |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
228 return ((const struct Unaligned *) v)->i; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
229 } |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
230 #elif defined(__DECC) |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
231 static inline uint32_t unaligned32(const void *v) { |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
232 return *(const __unaligned uint32_t *) v; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
233 } |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
234 #else |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
235 static inline uint32_t unaligned32(const void *v) { |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
236 return *(const uint32_t *) v; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
237 } |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
238 #endif |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
239 #endif //!ARCH_X86 |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
240 |
238
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
241 #ifndef ALT_BITSTREAM_WRITER |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
242 static inline void put_bits(PutBitContext *s, int n, unsigned int value) |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
243 { |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
244 unsigned int bit_buf; |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
245 int bit_left; |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
246 |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
247 #ifdef STATS |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
248 st_out_bit_counts[st_current_index] += n; |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
249 #endif |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
250 // printf("put_bits=%d %x\n", n, value); |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
251 assert(n == 32 || value < (1U << n)); |
306 | 252 |
238
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
253 bit_buf = s->bit_buf; |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
254 bit_left = s->bit_left; |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
255 |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
256 // printf("n=%d value=%x cnt=%d buf=%x\n", n, value, bit_cnt, bit_buf); |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
257 /* XXX: optimize */ |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
258 if (n < bit_left) { |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
259 bit_buf = (bit_buf<<n) | value; |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
260 bit_left-=n; |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
261 } else { |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
262 bit_buf<<=bit_left; |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
263 bit_buf |= value >> (n - bit_left); |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
264 *(UINT32 *)s->buf_ptr = be2me_32(bit_buf); |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
265 //printf("bitbuf = %08x\n", bit_buf); |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
266 s->buf_ptr+=4; |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
267 bit_left+=32 - n; |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
268 bit_buf = value; |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
269 } |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
270 |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
271 s->bit_buf = bit_buf; |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
272 s->bit_left = bit_left; |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
273 } |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
274 #endif |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
275 |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
276 |
234
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
277 #ifdef ALT_BITSTREAM_WRITER |
235
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
278 static inline void put_bits(PutBitContext *s, int n, unsigned int value) |
234
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
279 { |
235
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
280 #ifdef ALIGNED_BITSTREAM_WRITER |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
281 #ifdef ARCH_X86 |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
282 asm volatile( |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
283 "movl %0, %%ecx \n\t" |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
284 "xorl %%eax, %%eax \n\t" |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
285 "shrdl %%cl, %1, %%eax \n\t" |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
286 "shrl %%cl, %1 \n\t" |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
287 "movl %0, %%ecx \n\t" |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
288 "shrl $3, %%ecx \n\t" |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
289 "andl $0xFFFFFFFC, %%ecx \n\t" |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
290 "bswapl %1 \n\t" |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
291 "orl %1, (%2, %%ecx) \n\t" |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
292 "bswapl %%eax \n\t" |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
293 "addl %3, %0 \n\t" |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
294 "movl %%eax, 4(%2, %%ecx) \n\t" |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
295 : "=&r" (s->index), "=&r" (value) |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
296 : "r" (s->buf), "r" (n), "0" (s->index), "1" (value<<(-n)) |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
297 : "%eax", "%ecx" |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
298 ); |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
299 #else |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
300 int index= s->index; |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
301 uint32_t *ptr= ((uint32_t *)s->buf)+(index>>5); |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
302 |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
303 value<<= 32-n; |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
304 |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
305 ptr[0] |= be2me_32(value>>(index&31)); |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
306 ptr[1] = be2me_32(value<<(32-(index&31))); |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
307 //if(n>24) printf("%d %d\n", n, value); |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
308 index+= n; |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
309 s->index= index; |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
310 #endif |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
311 #else //ALIGNED_BITSTREAM_WRITER |
234
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
312 #ifdef ARCH_X86 |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
313 asm volatile( |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
314 "movl $7, %%ecx \n\t" |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
315 "andl %0, %%ecx \n\t" |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
316 "addl %3, %%ecx \n\t" |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
317 "negl %%ecx \n\t" |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
318 "shll %%cl, %1 \n\t" |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
319 "bswapl %1 \n\t" |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
320 "movl %0, %%ecx \n\t" |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
321 "shrl $3, %%ecx \n\t" |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
322 "orl %1, (%%ecx, %2) \n\t" |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
323 "addl %3, %0 \n\t" |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
324 "movl $0, 4(%%ecx, %2) \n\t" |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
325 : "=&r" (s->index), "=&r" (value) |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
326 : "r" (s->buf), "r" (n), "0" (s->index), "1" (value) |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
327 : "%ecx" |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
328 ); |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
329 #else |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
330 int index= s->index; |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
331 uint32_t *ptr= (uint32_t*)(((uint8_t *)s->buf)+(index>>3)); |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
332 |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
333 ptr[0] |= be2me_32(value<<(32-n-(index&7) )); |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
334 ptr[1] = 0; |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
335 //if(n>24) printf("%d %d\n", n, value); |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
336 index+= n; |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
337 s->index= index; |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
338 #endif |
235
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
339 #endif //!ALIGNED_BITSTREAM_WRITER |
234
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
340 } |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
341 #endif |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
342 |
238
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
343 #ifndef ALT_BITSTREAM_WRITER |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
344 /* for jpeg : escape 0xff with 0x00 after it */ |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
345 static inline void jput_bits(PutBitContext *s, int n, unsigned int value) |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
346 { |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
347 unsigned int bit_buf, b; |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
348 int bit_left, i; |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
349 |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
350 assert(n == 32 || value < (1U << n)); |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
351 |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
352 bit_buf = s->bit_buf; |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
353 bit_left = s->bit_left; |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
354 |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
355 //printf("n=%d value=%x cnt=%d buf=%x\n", n, value, bit_cnt, bit_buf); |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
356 /* XXX: optimize */ |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
357 if (n < bit_left) { |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
358 bit_buf = (bit_buf<<n) | value; |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
359 bit_left-=n; |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
360 } else { |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
361 bit_buf<<=bit_left; |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
362 bit_buf |= value >> (n - bit_left); |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
363 /* handle escape */ |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
364 for(i=0;i<4;i++) { |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
365 b = (bit_buf >> 24); |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
366 *(s->buf_ptr++) = b; |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
367 if (b == 0xff) |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
368 *(s->buf_ptr++) = 0; |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
369 bit_buf <<= 8; |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
370 } |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
371 |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
372 bit_left+= 32 - n; |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
373 bit_buf = value; |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
374 } |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
375 |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
376 s->bit_buf = bit_buf; |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
377 s->bit_left = bit_left; |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
378 } |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
379 #endif |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
380 |
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
381 |
234
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
382 #ifdef ALT_BITSTREAM_WRITER |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
383 static inline void jput_bits(PutBitContext *s, int n, int value) |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
384 { |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
385 int index= s->index; |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
386 uint32_t *ptr= (uint32_t*)(((uint8_t *)s->buf)+(index>>3)); |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
387 int v= ptr[0]; |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
388 //if(n>24) printf("%d %d\n", n, value); |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
389 |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
390 v |= be2me_32(value<<(32-n-(index&7) )); |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
391 if(((v+0x01010101)^0xFFFFFFFF)&v&0x80808080) |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
392 { |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
393 /* handle idiotic (m)jpeg escapes */ |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
394 uint8_t *bPtr= (uint8_t*)ptr; |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
395 int numChecked= ((index+n)>>3) - (index>>3); |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
396 |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
397 v= be2me_32(v); |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
398 |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
399 *(bPtr++)= v>>24; |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
400 if((v&0xFF000000)==0xFF000000 && numChecked>0){ |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
401 *(bPtr++)= 0x00; |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
402 index+=8; |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
403 } |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
404 *(bPtr++)= (v>>16)&0xFF; |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
405 if((v&0x00FF0000)==0x00FF0000 && numChecked>1){ |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
406 *(bPtr++)= 0x00; |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
407 index+=8; |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
408 } |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
409 *(bPtr++)= (v>>8)&0xFF; |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
410 if((v&0x0000FF00)==0x0000FF00 && numChecked>2){ |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
411 *(bPtr++)= 0x00; |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
412 index+=8; |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
413 } |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
414 *(bPtr++)= v&0xFF; |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
415 if((v&0x000000FF)==0x000000FF && numChecked>3){ |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
416 *(bPtr++)= 0x00; |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
417 index+=8; |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
418 } |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
419 *((uint32_t*)bPtr)= 0; |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
420 } |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
421 else |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
422 { |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
423 ptr[0] = v; |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
424 ptr[1] = 0; |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
425 } |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
426 |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
427 index+= n; |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
428 s->index= index; |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
429 } |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
430 #endif |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
431 |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
432 static inline uint8_t* pbBufPtr(PutBitContext *s) |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
433 { |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
434 #ifdef ALT_BITSTREAM_WRITER |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
435 return s->buf + (s->index>>3); |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
436 #else |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
437 return s->buf_ptr; |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
438 #endif |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
439 } |
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
440 |
0 | 441 void init_get_bits(GetBitContext *s, |
442 UINT8 *buffer, int buffer_size); | |
443 | |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
444 #ifndef ALT_BITSTREAM_READER |
20
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
445 unsigned int get_bits_long(GetBitContext *s, int n); |
144 | 446 unsigned int show_bits_long(GetBitContext *s, int n); |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
447 #endif |
20
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
448 |
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
449 static inline unsigned int get_bits(GetBitContext *s, int n){ |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
450 #ifdef ALT_BITSTREAM_READER |
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
451 #ifdef ALIGNED_BITSTREAM |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
452 int index= s->index; |
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
453 uint32_t result1= be2me_32( ((uint32_t *)s->buffer)[index>>5] ); |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
454 uint32_t result2= be2me_32( ((uint32_t *)s->buffer)[(index>>5) + 1] ); |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
455 #ifdef ARCH_X86 |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
456 asm ("shldl %%cl, %2, %0\n\t" |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
457 : "=r" (result1) |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
458 : "0" (result1), "r" (result2), "c" (index)); |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
459 #else |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
460 result1<<= (index&0x1F); |
199
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
461 result2= (result2>>1) >> (31-(index&0x1F)); |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
462 result1|= result2; |
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
463 #endif |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
464 result1>>= 32 - n; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
465 index+= n; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
466 s->index= index; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
467 |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
468 return result1; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
469 #else //ALIGNED_BITSTREAM |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
470 int index= s->index; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
471 uint32_t result= be2me_32( unaligned32( ((uint8_t *)s->buffer)+(index>>3) ) ); |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
472 |
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
473 result<<= (index&0x07); |
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
474 result>>= 32 - n; |
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
475 index+= n; |
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
476 s->index= index; |
306 | 477 #ifdef DUMP_STREAM |
478 while(n){ | |
479 printf("%d", (result>>(n-1))&1); | |
480 n--; | |
481 } | |
482 printf(" "); | |
483 #endif | |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
484 return result; |
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
485 #endif //!ALIGNED_BITSTREAM |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
486 #else //ALT_BITSTREAM_READER |
20
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
487 if(s->bit_cnt>=n){ |
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
488 /* most common case here */ |
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
489 unsigned int val = s->bit_buf >> (32 - n); |
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
490 s->bit_buf <<= n; |
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
491 s->bit_cnt -= n; |
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
492 #ifdef STATS |
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
493 st_bit_counts[st_current_index] += n; |
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
494 #endif |
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
495 return val; |
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
496 } |
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
497 return get_bits_long(s,n); |
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
498 #endif //!ALT_BITSTREAM_READER |
20
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
499 } |
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
500 |
21 | 501 static inline unsigned int get_bits1(GetBitContext *s){ |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
502 #ifdef ALT_BITSTREAM_READER |
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
503 int index= s->index; |
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
504 uint8_t result= s->buffer[ index>>3 ]; |
199
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
505 result<<= (index&0x07); |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
506 result>>= 8 - 1; |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
507 index++; |
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
508 s->index= index; |
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
509 |
306 | 510 #ifdef DUMP_STREAM |
511 printf("%d ", result); | |
512 #endif | |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
513 return result; |
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
514 #else |
21 | 515 if(s->bit_cnt>0){ |
516 /* most common case here */ | |
517 unsigned int val = s->bit_buf >> 31; | |
518 s->bit_buf <<= 1; | |
519 s->bit_cnt--; | |
520 #ifdef STATS | |
521 st_bit_counts[st_current_index]++; | |
522 #endif | |
523 return val; | |
524 } | |
525 return get_bits_long(s,1); | |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
526 #endif |
21 | 527 } |
528 | |
144 | 529 /* This function is identical to get_bits(), the only */ |
530 /* diference is that it doesn't touch the buffer */ | |
531 /* it is usefull to see the buffer. */ | |
532 static inline unsigned int show_bits(GetBitContext *s, int n) | |
533 { | |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
534 #ifdef ALT_BITSTREAM_READER |
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
535 #ifdef ALIGNED_BITSTREAM |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
536 int index= s->index; |
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
537 uint32_t result1= be2me_32( ((uint32_t *)s->buffer)[index>>5] ); |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
538 uint32_t result2= be2me_32( ((uint32_t *)s->buffer)[(index>>5) + 1] ); |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
539 #ifdef ARCH_X86 |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
540 asm ("shldl %%cl, %2, %0\n\t" |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
541 : "=r" (result1) |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
542 : "0" (result1), "r" (result2), "c" (index)); |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
543 #else |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
544 result1<<= (index&0x1F); |
199
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
545 result2= (result2>>1) >> (31-(index&0x1F)); |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
546 result1|= result2; |
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
547 #endif |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
548 result1>>= 32 - n; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
549 |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
550 return result1; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
551 #else //ALIGNED_BITSTREAM |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
552 int index= s->index; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
553 uint32_t result= be2me_32( unaligned32( ((uint8_t *)s->buffer)+(index>>3) ) ); |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
554 |
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
555 result<<= (index&0x07); |
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
556 result>>= 32 - n; |
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
557 |
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
558 return result; |
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
559 #endif //!ALIGNED_BITSTREAM |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
560 #else //ALT_BITSTREAM_READER |
144 | 561 if(s->bit_cnt>=n) { |
562 /* most common case here */ | |
563 unsigned int val = s->bit_buf >> (32 - n); | |
564 return val; | |
565 } | |
566 return show_bits_long(s,n); | |
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
567 #endif //!ALT_BITSTREAM_READER |
144 | 568 } |
569 | |
290 | 570 static inline int show_aligned_bits(GetBitContext *s, int offset, int n) |
571 { | |
572 #ifdef ALT_BITSTREAM_READER | |
573 #ifdef ALIGNED_BITSTREAM | |
574 int index= (s->index + offset + 7)&(~7); | |
575 uint32_t result1= be2me_32( ((uint32_t *)s->buffer)[index>>5] ); | |
576 uint32_t result2= be2me_32( ((uint32_t *)s->buffer)[(index>>5) + 1] ); | |
577 #ifdef ARCH_X86 | |
578 asm ("shldl %%cl, %2, %0\n\t" | |
579 : "=r" (result1) | |
580 : "0" (result1), "r" (result2), "c" (index)); | |
581 #else | |
582 result1<<= (index&0x1F); | |
583 result2= (result2>>1) >> (31-(index&0x1F)); | |
584 result1|= result2; | |
585 #endif | |
586 result1>>= 32 - n; | |
587 | |
588 return result1; | |
589 #else //ALIGNED_BITSTREAM | |
590 int index= (s->index + offset + 7)>>3; | |
591 uint32_t result= be2me_32( unaligned32( ((uint8_t *)s->buffer)+index ) ); | |
592 | |
593 result>>= 32 - n; | |
594 | |
595 return result; | |
596 #endif //!ALIGNED_BITSTREAM | |
597 #else //ALT_BITSTREAM_READER | |
598 int index= (get_bits_count(s) + offset + 7)>>3; | |
599 uint32_t result= be2me_32( unaligned32( ((uint8_t *)s->buf)+index ) ); | |
600 | |
601 result>>= 32 - n; | |
602 //printf(" %X %X %d \n", (int)(((uint8_t *)s->buf)+index ), (int)s->buf_ptr, s->bit_cnt); | |
603 return result; | |
604 #endif //!ALT_BITSTREAM_READER | |
605 } | |
606 | |
21 | 607 static inline void skip_bits(GetBitContext *s, int n){ |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
608 #ifdef ALT_BITSTREAM_READER |
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
609 s->index+= n; |
306 | 610 #ifdef DUMP_STREAM |
611 { | |
612 int result; | |
613 s->index-= n; | |
614 result= get_bits(s, n); | |
615 } | |
616 #endif | |
617 | |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
618 #else |
21 | 619 if(s->bit_cnt>=n){ |
620 /* most common case here */ | |
621 s->bit_buf <<= n; | |
622 s->bit_cnt -= n; | |
623 #ifdef STATS | |
624 st_bit_counts[st_current_index] += n; | |
625 #endif | |
626 } else { | |
627 get_bits_long(s,n); | |
628 } | |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
629 #endif |
21 | 630 } |
631 | |
632 static inline void skip_bits1(GetBitContext *s){ | |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
633 #ifdef ALT_BITSTREAM_READER |
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
634 s->index++; |
306 | 635 #ifdef DUMP_STREAM |
636 s->index--; | |
637 printf("%d ", get_bits1(s)); | |
638 #endif | |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
639 #else |
21 | 640 if(s->bit_cnt>0){ |
641 /* most common case here */ | |
642 s->bit_buf <<= 1; | |
643 s->bit_cnt--; | |
644 #ifdef STATS | |
645 st_bit_counts[st_current_index]++; | |
646 #endif | |
647 } else { | |
648 get_bits_long(s,1); | |
649 } | |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
650 #endif |
21 | 651 } |
652 | |
85 | 653 static inline int get_bits_count(GetBitContext *s) |
654 { | |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
655 #ifdef ALT_BITSTREAM_READER |
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
656 return s->index; |
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
657 #else |
85 | 658 return (s->buf_ptr - s->buf) * 8 - s->bit_cnt; |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
659 #endif |
85 | 660 } |
21 | 661 |
264
28c5c62b1c4c
support decoding (with mplayer) of 3 .mp4 files from mphq
michaelni
parents:
238
diff
changeset
|
662 int check_marker(GetBitContext *s, char *msg); |
0 | 663 void align_get_bits(GetBitContext *s); |
664 int init_vlc(VLC *vlc, int nb_bits, int nb_codes, | |
665 const void *bits, int bits_wrap, int bits_size, | |
666 const void *codes, int codes_wrap, int codes_size); | |
667 void free_vlc(VLC *vlc); | |
668 | |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
669 #ifdef ALT_BITSTREAM_READER |
199
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
670 #ifdef ALIGNED_BITSTREAM |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
671 #ifdef ARCH_X86 |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
672 #define SHOW_BITS(s, val, n) \ |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
673 val= be2me_32( ((uint32_t *)(s)->buffer)[bit_cnt>>5] );\ |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
674 {uint32_t result2= be2me_32( ((uint32_t *)(s)->buffer)[(bit_cnt>>5) + 1] );\ |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
675 asm ("shldl %%cl, %2, %0\n\t"\ |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
676 : "=r" (val)\ |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
677 : "0" (val), "r" (result2), "c" (bit_cnt));\ |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
678 ((uint32_t)val)>>= 32 - n;} |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
679 #else //ARCH_X86 |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
680 #define SHOW_BITS(s, val, n) \ |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
681 val= be2me_32( ((uint32_t *)(s)->buffer)[bit_cnt>>5] );\ |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
682 {uint32_t result2= be2me_32( ((uint32_t *)(s)->buffer)[(bit_cnt>>5) + 1] );\ |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
683 val<<= (bit_cnt&0x1F);\ |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
684 result2= (result2>>1) >> (31-(bit_cnt&0x1F));\ |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
685 val|= result2;\ |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
686 ((uint32_t)val)>>= 32 - n;} |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
687 #endif //!ARCH_X86 |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
688 #else //ALIGNED_BITSTREAM |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
689 #define SHOW_BITS(s, val, n) \ |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
690 val= be2me_32( unaligned32( ((uint8_t *)(s)->buffer)+(bit_cnt>>3) ) );\ |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
691 val<<= (bit_cnt&0x07);\ |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
692 ((uint32_t)val)>>= 32 - n; |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
693 #endif // !ALIGNED_BITSTREAM |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
694 #define FLUSH_BITS(n) bit_cnt+=n; |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
695 #define SAVE_BITS(s) bit_cnt= (s)->index; |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
696 #define RESTORE_BITS(s) (s)->index= bit_cnt; |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
697 #else |
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
698 |
0 | 699 /* macro to go faster */ |
700 /* n must be <= 24 */ | |
701 /* XXX: optimize buffer end test */ | |
702 #define SHOW_BITS(s, val, n)\ | |
703 {\ | |
704 if (bit_cnt < n && buf_ptr < (s)->buf_end) {\ | |
705 bit_buf |= *buf_ptr++ << (24 - bit_cnt);\ | |
706 bit_cnt += 8;\ | |
707 if (bit_cnt < n && buf_ptr < (s)->buf_end) {\ | |
708 bit_buf |= *buf_ptr++ << (24 - bit_cnt);\ | |
709 bit_cnt += 8;\ | |
710 if (bit_cnt < n && buf_ptr < (s)->buf_end) {\ | |
711 bit_buf |= *buf_ptr++ << (24 - bit_cnt);\ | |
712 bit_cnt += 8;\ | |
713 }\ | |
714 }\ | |
715 }\ | |
716 val = bit_buf >> (32 - n);\ | |
717 } | |
718 | |
719 /* SHOW_BITS with n1 >= n must be been done before */ | |
720 #define FLUSH_BITS(n)\ | |
721 {\ | |
722 bit_buf <<= n;\ | |
723 bit_cnt -= n;\ | |
724 } | |
725 | |
726 #define SAVE_BITS(s) \ | |
727 {\ | |
728 bit_cnt = (s)->bit_cnt;\ | |
729 bit_buf = (s)->bit_buf;\ | |
730 buf_ptr = (s)->buf_ptr;\ | |
731 } | |
732 | |
733 #define RESTORE_BITS(s) \ | |
734 {\ | |
735 (s)->buf_ptr = buf_ptr;\ | |
736 (s)->bit_buf = bit_buf;\ | |
737 (s)->bit_cnt = bit_cnt;\ | |
738 } | |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
739 #endif // !ALT_BITSTREAM_READER |
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
740 |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
741 static inline int get_vlc(GetBitContext *s, VLC *vlc) |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
742 { |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
743 int code, n, nb_bits, index; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
744 INT16 *table_codes; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
745 INT8 *table_bits; |
199
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
746 int bit_cnt; |
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
747 #ifndef ALT_BITSTREAM_READER |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
748 UINT32 bit_buf; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
749 UINT8 *buf_ptr; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
750 #endif |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
751 |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
752 SAVE_BITS(s); |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
753 nb_bits = vlc->bits; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
754 table_codes = vlc->table_codes; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
755 table_bits = vlc->table_bits; |
199
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
756 |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
757 #ifdef FAST_GET_FIRST_VLC |
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
758 SHOW_BITS(s, index, nb_bits); |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
759 code = table_codes[index]; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
760 n = table_bits[index]; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
761 if (n > 0) { |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
762 /* most common case (90%)*/ |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
763 FLUSH_BITS(n); |
306 | 764 #ifdef DUMP_STREAM |
765 { | |
766 int n= bit_cnt - s->index; | |
767 skip_bits(s, n); | |
768 RESTORE_BITS(s); | |
769 } | |
770 #endif | |
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
771 RESTORE_BITS(s); |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
772 return code; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
773 } else if (n == 0) { |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
774 return -1; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
775 } else { |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
776 FLUSH_BITS(nb_bits); |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
777 nb_bits = -n; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
778 table_codes = vlc->table_codes + code; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
779 table_bits = vlc->table_bits + code; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
780 } |
199
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
781 #endif |
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
782 for(;;) { |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
783 SHOW_BITS(s, index, nb_bits); |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
784 code = table_codes[index]; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
785 n = table_bits[index]; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
786 if (n > 0) { |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
787 /* most common case */ |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
788 FLUSH_BITS(n); |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
789 #ifdef STATS |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
790 st_bit_counts[st_current_index] += n; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
791 #endif |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
792 break; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
793 } else if (n == 0) { |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
794 return -1; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
795 } else { |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
796 FLUSH_BITS(nb_bits); |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
797 #ifdef STATS |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
798 st_bit_counts[st_current_index] += nb_bits; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
799 #endif |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
800 nb_bits = -n; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
801 table_codes = vlc->table_codes + code; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
802 table_bits = vlc->table_bits + code; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
803 } |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
804 } |
306 | 805 #ifdef DUMP_STREAM |
806 { | |
807 int n= bit_cnt - s->index; | |
808 skip_bits(s, n); | |
809 RESTORE_BITS(s); | |
810 } | |
811 #endif | |
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
812 RESTORE_BITS(s); |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
813 return code; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
814 } |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
815 |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
816 |
0 | 817 /* define it to include statistics code (useful only for optimizing |
818 codec efficiency */ | |
819 //#define STATS | |
820 | |
821 #ifdef STATS | |
822 | |
823 enum { | |
824 ST_UNKNOWN, | |
825 ST_DC, | |
826 ST_INTRA_AC, | |
827 ST_INTER_AC, | |
828 ST_INTRA_MB, | |
829 ST_INTER_MB, | |
830 ST_MV, | |
831 ST_NB, | |
832 }; | |
833 | |
834 extern int st_current_index; | |
835 extern unsigned int st_bit_counts[ST_NB]; | |
836 extern unsigned int st_out_bit_counts[ST_NB]; | |
837 | |
838 void print_stats(void); | |
839 #endif | |
840 | |
841 /* misc math functions */ | |
842 | |
151
ae0516eadae2
fixed gcc-3.0.x compilation (by Michael Niedermayer)
nickols_k
parents:
144
diff
changeset
|
843 static inline int av_log2(unsigned int v) |
0 | 844 { |
845 int n; | |
846 | |
847 n = 0; | |
848 if (v & 0xffff0000) { | |
849 v >>= 16; | |
850 n += 16; | |
851 } | |
852 if (v & 0xff00) { | |
853 v >>= 8; | |
854 n += 8; | |
855 } | |
856 if (v & 0xf0) { | |
857 v >>= 4; | |
858 n += 4; | |
859 } | |
860 if (v & 0xc) { | |
861 v >>= 2; | |
862 n += 2; | |
863 } | |
864 if (v & 0x2) { | |
865 n++; | |
866 } | |
867 return n; | |
868 } | |
869 | |
277
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
870 /* median of 3 */ |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
871 static inline int mid_pred(int a, int b, int c) |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
872 { |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
873 int vmin, vmax; |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
874 vmax = vmin = a; |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
875 if (b < vmin) |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
876 vmin = b; |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
877 else |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
878 vmax = b; |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
879 |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
880 if (c < vmin) |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
881 vmin = c; |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
882 else if (c > vmax) |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
883 vmax = c; |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
884 |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
885 return a + b + c - vmin - vmax; |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
886 } |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
887 |
327 | 888 static inline int clip(int a, int amin, int amax) |
889 { | |
890 if (a < amin) | |
891 return amin; | |
892 else if (a > amax) | |
893 return amax; | |
894 else | |
895 return a; | |
896 } | |
897 | |
324 | 898 /* math */ |
899 int ff_gcd(int a, int b); | |
900 | |
451 | 901 static inline int ff_sqrt(int a) |
902 { | |
903 int ret=0; | |
904 int s; | |
905 int ret_sq=0; | |
906 | |
907 for(s=15; s>=0; s--){ | |
908 int b= ret_sq + (1<<(s*2)) + (ret<<s)*2; | |
909 if(b<=a){ | |
910 ret_sq=b; | |
911 ret+= 1<<s; | |
912 } | |
913 } | |
914 return ret; | |
915 } | |
916 #if __CPU__ >= 686 && !defined(RUNTIME_CPUDETECT) | |
917 #define COPY3_IF_LT(x,y,a,b,c,d)\ | |
918 asm volatile (\ | |
919 "cmpl %0, %3 \n\t"\ | |
920 "cmovl %3, %0 \n\t"\ | |
921 "cmovl %4, %1 \n\t"\ | |
922 "cmovl %5, %2 \n\t"\ | |
923 : "+r" (x), "+r" (a), "+r" (c)\ | |
924 : "r" (y), "r" (b), "r" (d)\ | |
925 ); | |
926 #else | |
927 #define COPY3_IF_LT(x,y,a,b,c,d)\ | |
928 if((y)<(x)){\ | |
929 (x)=(y);\ | |
930 (a)=(b);\ | |
931 (c)=(d);\ | |
932 } | |
933 #endif | |
934 | |
370
0eca28d16cbd
clamp intra matrix to 8bit for mjpeg (workaround for qscale>=25)
al3x
parents:
359
diff
changeset
|
935 #define CLAMP_TO_8BIT(d) ((d > 0xff) ? 0xff : (d < 0) ? 0 : d) |
0eca28d16cbd
clamp intra matrix to 8bit for mjpeg (workaround for qscale>=25)
al3x
parents:
359
diff
changeset
|
936 |
488 | 937 #endif /* HAVE_AV_CONFIG_H */ |
938 | |
939 #endif /* COMMON_H */ |