Mercurial > libavcodec.hg
annotate common.h @ 471:d7f65ea52aaa libavcodec
* reimplemented remaing avg_ pixel functions
* modified PAVGB mmx macro to select mmx register with FE constant
author | kabi |
---|---|
date | Mon, 03 Jun 2002 17:35:35 +0000 |
parents | 1228bf19299b |
children | 90227eb98bf9 |
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 |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
13 //#define ALT_BITSTREAM_READER |
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
14 //#define ALIGNED_BITSTREAM |
199
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
15 #define FAST_GET_FIRST_VLC |
306 | 16 //#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
|
17 |
10 | 18 #ifdef HAVE_AV_CONFIG_H |
64 | 19 /* only include the following when compiling package */ |
408 | 20 #include "config.h" |
64 | 21 |
22 #include <stdlib.h> | |
23 #include <stdio.h> | |
24 #include <string.h> | |
25 #include <errno.h> | |
390
48e08d9871da
added proper memory handling functions - fixed include paths
glantau
parents:
370
diff
changeset
|
26 #include <math.h> |
64 | 27 |
28 #ifndef ENODATA | |
29 #define ENODATA 61 | |
10 | 30 #endif |
0 | 31 |
390
48e08d9871da
added proper memory handling functions - fixed include paths
glantau
parents:
370
diff
changeset
|
32 #endif /* HAVE_AV_CONFIG_H */ |
64 | 33 |
34 #ifdef CONFIG_WIN32 | |
35 | |
36 /* windows */ | |
37 | |
38 typedef unsigned short UINT16; | |
39 typedef signed short INT16; | |
40 typedef unsigned char UINT8; | |
41 typedef unsigned int UINT32; | |
42 typedef unsigned __int64 UINT64; | |
43 typedef signed char INT8; | |
44 typedef signed int INT32; | |
45 typedef signed __int64 INT64; | |
46 | |
47 typedef UINT8 uint8_t; | |
48 typedef INT8 int8_t; | |
49 typedef UINT16 uint16_t; | |
50 typedef INT16 int16_t; | |
51 typedef UINT32 uint32_t; | |
52 typedef INT32 int32_t; | |
426 | 53 typedef UINT64 uint64_t; |
54 typedef INT64 int64_t; | |
64 | 55 |
76 | 56 #ifndef __MINGW32__ |
64 | 57 #define INT64_C(c) (c ## i64) |
58 #define UINT64_C(c) (c ## i64) | |
59 | |
60 #define inline __inline | |
61 | |
76 | 62 #else |
63 #define INT64_C(c) (c ## LL) | |
64 #define UINT64_C(c) (c ## ULL) | |
65 #endif /* __MINGW32__ */ | |
66 | |
64 | 67 #define M_PI 3.14159265358979323846 |
68 #define M_SQRT2 1.41421356237309504880 /* sqrt(2) */ | |
69 | |
70 #ifdef _DEBUG | |
71 #define DEBUG | |
72 #endif | |
73 | |
74 #define snprintf _snprintf | |
75 | |
426 | 76 #else /* CONFIG_WIN32 */ |
64 | 77 |
78 /* unix */ | |
79 | |
80 #include <inttypes.h> | |
81 | |
2 | 82 #ifndef __WINE_WINDEF16_H |
83 /* workaround for typedef conflict in MPlayer (wine typedefs) */ | |
0 | 84 typedef unsigned short UINT16; |
85 typedef signed short INT16; | |
86 #endif | |
87 | |
88 typedef unsigned char UINT8; | |
89 typedef unsigned int UINT32; | |
90 typedef unsigned long long UINT64; | |
91 typedef signed char INT8; | |
92 typedef signed int INT32; | |
93 typedef signed long long INT64; | |
94 | |
64 | 95 #ifdef HAVE_AV_CONFIG_H |
96 | |
97 #ifdef __FreeBSD__ | |
98 #include <sys/param.h> | |
99 #endif | |
100 | |
101 #ifndef INT64_C | |
102 #define INT64_C(c) (c ## LL) | |
103 #define UINT64_C(c) (c ## ULL) | |
104 #endif | |
105 | |
106 #ifdef USE_FASTMEMCPY | |
107 #include "fastmemcpy.h" | |
108 #endif | |
109 | |
76 | 110 #endif /* HAVE_AV_CONFIG_H */ |
111 | |
112 #endif /* !CONFIG_WIN32 */ | |
113 | |
437
86f68813b97f
* for this moment bswap.h included ahead of HAVE_AV_CONFIG_H check
kabi
parents:
432
diff
changeset
|
114 #include "bswap.h" |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
115 |
426 | 116 #ifdef HAVE_AV_CONFIG_H |
117 | |
432 | 118 #if defined(__MINGW32__) || defined(__CYGWIN__) || \ |
119 defined(__OS2__) || defined (__OpenBSD__) | |
120 #define MANGLE(a) "_" #a | |
121 #else | |
122 #define MANGLE(a) #a | |
123 #endif | |
124 | |
76 | 125 /* debug stuff */ |
126 | |
64 | 127 #ifndef DEBUG |
128 #define NDEBUG | |
129 #endif | |
130 #include <assert.h> | |
131 | |
76 | 132 /* dprintf macros */ |
133 #if defined(CONFIG_WIN32) && !defined(__MINGW32__) | |
134 | |
135 inline void dprintf(const char* fmt,...) {} | |
136 | |
137 #else | |
138 | |
139 #ifdef DEBUG | |
140 #define dprintf(fmt,args...) printf(fmt, ## args) | |
141 #else | |
142 #define dprintf(fmt,args...) | |
143 #endif | |
64 | 144 |
145 #endif /* !CONFIG_WIN32 */ | |
146 | |
76 | 147 #endif /* HAVE_AV_CONFIG_H */ |
148 | |
423
ed6098c4216a
Add an av_abort macro that aborts, but also prints out the location of the abort.
philipjsg
parents:
408
diff
changeset
|
149 #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
|
150 |
344 | 151 /* assume b>0 */ |
152 #define ROUNDED_DIV(a,b) (((a)>0 ? (a) + ((b)>>1) : (a) - ((b)>>1))/(b)) | |
359 | 153 #define ABS(a) ((a) >= 0 ? (a) : (-(a))) |
451 | 154 #define MAX(a,b) ((a) > (b) ? (a) : (b)) |
155 #define MIN(a,b) ((a) > (b) ? (b) : (a)) | |
344 | 156 |
0 | 157 /* bit output */ |
158 | |
159 struct PutBitContext; | |
160 | |
161 typedef void (*WriteDataFunc)(void *, UINT8 *, int); | |
162 | |
163 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
|
164 #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
|
165 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
|
166 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
|
167 #else |
20
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
168 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
|
169 int bit_left; |
20
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
170 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
|
171 #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
|
172 INT64 data_out_size; /* in bytes */ |
0 | 173 } PutBitContext; |
174 | |
175 void init_put_bits(PutBitContext *s, | |
176 UINT8 *buffer, int buffer_size, | |
177 void *opaque, | |
178 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
|
179 |
85 | 180 INT64 get_bit_count(PutBitContext *s); /* XXX: change function name */ |
0 | 181 void align_put_bits(PutBitContext *s); |
182 void flush_put_bits(PutBitContext *s); | |
358 | 183 void put_string(PutBitContext * pbc, char *s); |
0 | 184 |
185 /* jpeg specific put_bits */ | |
186 void jflush_put_bits(PutBitContext *s); | |
187 | |
188 /* bit input */ | |
189 | |
190 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
|
191 #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
|
192 int index; |
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
193 UINT8 *buffer; |
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
194 #else |
20
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
195 UINT32 bit_buf; |
0 | 196 int bit_cnt; |
20
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
197 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
|
198 #endif |
277
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
199 int size; |
0 | 200 } GetBitContext; |
201 | |
290 | 202 static inline int get_bits_count(GetBitContext *s); |
203 | |
0 | 204 typedef struct VLC { |
205 int bits; | |
206 INT16 *table_codes; | |
207 INT8 *table_bits; | |
208 int table_size, table_allocated; | |
209 } VLC; | |
210 | |
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
211 /* 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
|
212 #ifdef ARCH_X86 |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
213 #define unaligned32(a) (*(UINT32*)(a)) |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
214 #else |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
215 #ifdef __GNUC__ |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
216 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
|
217 struct Unaligned { |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
218 uint32_t i; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
219 } __attribute__((packed)); |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
220 |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
221 return ((const struct Unaligned *) v)->i; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
222 } |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
223 #elif defined(__DECC) |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
224 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
|
225 return *(const __unaligned uint32_t *) v; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
226 } |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
227 #else |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
228 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
|
229 return *(const uint32_t *) v; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
230 } |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
231 #endif |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
232 #endif //!ARCH_X86 |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
233 |
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
|
234 #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
|
235 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
|
236 { |
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
|
237 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
|
238 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
|
239 |
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
|
240 #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
|
241 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
|
242 #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
|
243 // 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
|
244 assert(n == 32 || value < (1U << n)); |
306 | 245 |
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
|
246 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
|
247 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
|
248 |
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 // 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
|
250 /* 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
|
251 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
|
252 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
|
253 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
|
254 } 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
|
255 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
|
256 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
|
257 *(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
|
258 //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
|
259 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
|
260 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
|
261 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
|
262 } |
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 |
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 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
|
265 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
|
266 } |
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 #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
|
268 |
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 |
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
|
270 #ifdef ALT_BITSTREAM_WRITER |
235
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
271 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
|
272 { |
235
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
273 #ifdef ALIGNED_BITSTREAM_WRITER |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
274 #ifdef ARCH_X86 |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
275 asm volatile( |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
276 "movl %0, %%ecx \n\t" |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
277 "xorl %%eax, %%eax \n\t" |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
278 "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
|
279 "shrl %%cl, %1 \n\t" |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
280 "movl %0, %%ecx \n\t" |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
281 "shrl $3, %%ecx \n\t" |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
282 "andl $0xFFFFFFFC, %%ecx \n\t" |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
283 "bswapl %1 \n\t" |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
284 "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
|
285 "bswapl %%eax \n\t" |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
286 "addl %3, %0 \n\t" |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
287 "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
|
288 : "=&r" (s->index), "=&r" (value) |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
289 : "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
|
290 : "%eax", "%ecx" |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
291 ); |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
292 #else |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
293 int index= s->index; |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
294 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
|
295 |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
296 value<<= 32-n; |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
297 |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
298 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
|
299 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
|
300 //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
|
301 index+= n; |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
302 s->index= index; |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
303 #endif |
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
304 #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
|
305 #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
|
306 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
|
307 "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
|
308 "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
|
309 "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
|
310 "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
|
311 "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
|
312 "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
|
313 "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
|
314 "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
|
315 "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
|
316 "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
|
317 "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
|
318 : "=&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
|
319 : "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
|
320 : "%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
|
321 ); |
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 #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
|
323 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
|
324 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
|
325 |
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 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
|
327 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
|
328 //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
|
329 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
|
330 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
|
331 #endif |
235
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
332 #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
|
333 } |
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 #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
|
335 |
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
|
336 #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
|
337 /* 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
|
338 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
|
339 { |
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
|
340 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
|
341 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
|
342 |
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 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
|
344 |
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 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
|
346 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
|
347 |
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 //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
|
349 /* 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
|
350 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
|
351 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
|
352 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
|
353 } 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
|
354 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
|
355 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
|
356 /* 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
|
357 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
|
358 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
|
359 *(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
|
360 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
|
361 *(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
|
362 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
|
363 } |
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 |
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 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
|
366 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
|
367 } |
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 |
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 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
|
370 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
|
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 #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
|
373 |
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 |
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
|
375 #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
|
376 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
|
377 { |
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
|
378 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
|
379 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
|
380 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
|
381 //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
|
382 |
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 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
|
384 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
|
385 { |
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 /* 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
|
387 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
|
388 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
|
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(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
|
391 |
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 *(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
|
393 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
|
394 *(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
|
395 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
|
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 *(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
|
398 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
|
399 *(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
|
400 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
|
401 } |
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 *(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
|
403 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
|
404 *(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
|
405 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
|
406 } |
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 *(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
|
408 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
|
409 *(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
|
410 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
|
411 } |
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 *((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
|
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 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
|
415 { |
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 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
|
417 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
|
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 |
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 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
|
421 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
|
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 #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
|
424 |
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 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
|
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 #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
|
428 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
|
429 #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
|
430 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
|
431 #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
|
432 } |
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 |
0 | 434 void init_get_bits(GetBitContext *s, |
435 UINT8 *buffer, int buffer_size); | |
436 | |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
437 #ifndef ALT_BITSTREAM_READER |
20
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
438 unsigned int get_bits_long(GetBitContext *s, int n); |
144 | 439 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
|
440 #endif |
20
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
441 |
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
442 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
|
443 #ifdef ALT_BITSTREAM_READER |
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
444 #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
|
445 int index= s->index; |
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
446 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
|
447 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
|
448 #ifdef ARCH_X86 |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
449 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
|
450 : "=r" (result1) |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
451 : "0" (result1), "r" (result2), "c" (index)); |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
452 #else |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
453 result1<<= (index&0x1F); |
199
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
454 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
|
455 result1|= result2; |
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
456 #endif |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
457 result1>>= 32 - n; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
458 index+= n; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
459 s->index= index; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
460 |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
461 return result1; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
462 #else //ALIGNED_BITSTREAM |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
463 int index= s->index; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
464 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
|
465 |
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
466 result<<= (index&0x07); |
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
467 result>>= 32 - n; |
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
468 index+= n; |
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
469 s->index= index; |
306 | 470 #ifdef DUMP_STREAM |
471 while(n){ | |
472 printf("%d", (result>>(n-1))&1); | |
473 n--; | |
474 } | |
475 printf(" "); | |
476 #endif | |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
477 return result; |
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
478 #endif //!ALIGNED_BITSTREAM |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
479 #else //ALT_BITSTREAM_READER |
20
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
480 if(s->bit_cnt>=n){ |
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
481 /* most common case here */ |
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
482 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
|
483 s->bit_buf <<= n; |
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
484 s->bit_cnt -= n; |
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
485 #ifdef STATS |
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
486 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
|
487 #endif |
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
488 return val; |
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
489 } |
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
490 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
|
491 #endif //!ALT_BITSTREAM_READER |
20
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
492 } |
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
493 |
21 | 494 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
|
495 #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
|
496 int index= s->index; |
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
497 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
|
498 result<<= (index&0x07); |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
499 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
|
500 index++; |
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
501 s->index= index; |
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
502 |
306 | 503 #ifdef DUMP_STREAM |
504 printf("%d ", result); | |
505 #endif | |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
506 return result; |
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
507 #else |
21 | 508 if(s->bit_cnt>0){ |
509 /* most common case here */ | |
510 unsigned int val = s->bit_buf >> 31; | |
511 s->bit_buf <<= 1; | |
512 s->bit_cnt--; | |
513 #ifdef STATS | |
514 st_bit_counts[st_current_index]++; | |
515 #endif | |
516 return val; | |
517 } | |
518 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
|
519 #endif |
21 | 520 } |
521 | |
144 | 522 /* This function is identical to get_bits(), the only */ |
523 /* diference is that it doesn't touch the buffer */ | |
524 /* it is usefull to see the buffer. */ | |
525 static inline unsigned int show_bits(GetBitContext *s, int n) | |
526 { | |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
527 #ifdef ALT_BITSTREAM_READER |
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
528 #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
|
529 int index= s->index; |
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
530 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
|
531 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
|
532 #ifdef ARCH_X86 |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
533 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
|
534 : "=r" (result1) |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
535 : "0" (result1), "r" (result2), "c" (index)); |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
536 #else |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
537 result1<<= (index&0x1F); |
199
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
538 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
|
539 result1|= result2; |
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
540 #endif |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
541 result1>>= 32 - n; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
542 |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
543 return result1; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
544 #else //ALIGNED_BITSTREAM |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
545 int index= s->index; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
546 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
|
547 |
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
548 result<<= (index&0x07); |
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
549 result>>= 32 - n; |
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
550 |
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
551 return result; |
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
552 #endif //!ALIGNED_BITSTREAM |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
553 #else //ALT_BITSTREAM_READER |
144 | 554 if(s->bit_cnt>=n) { |
555 /* most common case here */ | |
556 unsigned int val = s->bit_buf >> (32 - n); | |
557 return val; | |
558 } | |
559 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
|
560 #endif //!ALT_BITSTREAM_READER |
144 | 561 } |
562 | |
290 | 563 static inline int show_aligned_bits(GetBitContext *s, int offset, int n) |
564 { | |
565 #ifdef ALT_BITSTREAM_READER | |
566 #ifdef ALIGNED_BITSTREAM | |
567 int index= (s->index + offset + 7)&(~7); | |
568 uint32_t result1= be2me_32( ((uint32_t *)s->buffer)[index>>5] ); | |
569 uint32_t result2= be2me_32( ((uint32_t *)s->buffer)[(index>>5) + 1] ); | |
570 #ifdef ARCH_X86 | |
571 asm ("shldl %%cl, %2, %0\n\t" | |
572 : "=r" (result1) | |
573 : "0" (result1), "r" (result2), "c" (index)); | |
574 #else | |
575 result1<<= (index&0x1F); | |
576 result2= (result2>>1) >> (31-(index&0x1F)); | |
577 result1|= result2; | |
578 #endif | |
579 result1>>= 32 - n; | |
580 | |
581 return result1; | |
582 #else //ALIGNED_BITSTREAM | |
583 int index= (s->index + offset + 7)>>3; | |
584 uint32_t result= be2me_32( unaligned32( ((uint8_t *)s->buffer)+index ) ); | |
585 | |
586 result>>= 32 - n; | |
587 | |
588 return result; | |
589 #endif //!ALIGNED_BITSTREAM | |
590 #else //ALT_BITSTREAM_READER | |
591 int index= (get_bits_count(s) + offset + 7)>>3; | |
592 uint32_t result= be2me_32( unaligned32( ((uint8_t *)s->buf)+index ) ); | |
593 | |
594 result>>= 32 - n; | |
595 //printf(" %X %X %d \n", (int)(((uint8_t *)s->buf)+index ), (int)s->buf_ptr, s->bit_cnt); | |
596 return result; | |
597 #endif //!ALT_BITSTREAM_READER | |
598 } | |
599 | |
21 | 600 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
|
601 #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
|
602 s->index+= n; |
306 | 603 #ifdef DUMP_STREAM |
604 { | |
605 int result; | |
606 s->index-= n; | |
607 result= get_bits(s, n); | |
608 } | |
609 #endif | |
610 | |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
611 #else |
21 | 612 if(s->bit_cnt>=n){ |
613 /* most common case here */ | |
614 s->bit_buf <<= n; | |
615 s->bit_cnt -= n; | |
616 #ifdef STATS | |
617 st_bit_counts[st_current_index] += n; | |
618 #endif | |
619 } else { | |
620 get_bits_long(s,n); | |
621 } | |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
622 #endif |
21 | 623 } |
624 | |
625 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
|
626 #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
|
627 s->index++; |
306 | 628 #ifdef DUMP_STREAM |
629 s->index--; | |
630 printf("%d ", get_bits1(s)); | |
631 #endif | |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
632 #else |
21 | 633 if(s->bit_cnt>0){ |
634 /* most common case here */ | |
635 s->bit_buf <<= 1; | |
636 s->bit_cnt--; | |
637 #ifdef STATS | |
638 st_bit_counts[st_current_index]++; | |
639 #endif | |
640 } else { | |
641 get_bits_long(s,1); | |
642 } | |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
643 #endif |
21 | 644 } |
645 | |
85 | 646 static inline int get_bits_count(GetBitContext *s) |
647 { | |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
648 #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
|
649 return s->index; |
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
650 #else |
85 | 651 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
|
652 #endif |
85 | 653 } |
21 | 654 |
264
28c5c62b1c4c
support decoding (with mplayer) of 3 .mp4 files from mphq
michaelni
parents:
238
diff
changeset
|
655 int check_marker(GetBitContext *s, char *msg); |
0 | 656 void align_get_bits(GetBitContext *s); |
657 int init_vlc(VLC *vlc, int nb_bits, int nb_codes, | |
658 const void *bits, int bits_wrap, int bits_size, | |
659 const void *codes, int codes_wrap, int codes_size); | |
660 void free_vlc(VLC *vlc); | |
661 | |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
662 #ifdef ALT_BITSTREAM_READER |
199
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
663 #ifdef ALIGNED_BITSTREAM |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
664 #ifdef ARCH_X86 |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
665 #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
|
666 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
|
667 {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
|
668 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
|
669 : "=r" (val)\ |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
670 : "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
|
671 ((uint32_t)val)>>= 32 - n;} |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
672 #else //ARCH_X86 |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
673 #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
|
674 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
|
675 {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
|
676 val<<= (bit_cnt&0x1F);\ |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
677 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
|
678 val|= result2;\ |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
679 ((uint32_t)val)>>= 32 - n;} |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
680 #endif //!ARCH_X86 |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
681 #else //ALIGNED_BITSTREAM |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
682 #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
|
683 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
|
684 val<<= (bit_cnt&0x07);\ |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
685 ((uint32_t)val)>>= 32 - n; |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
686 #endif // !ALIGNED_BITSTREAM |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
687 #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
|
688 #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
|
689 #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
|
690 #else |
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
691 |
0 | 692 /* macro to go faster */ |
693 /* n must be <= 24 */ | |
694 /* XXX: optimize buffer end test */ | |
695 #define SHOW_BITS(s, val, n)\ | |
696 {\ | |
697 if (bit_cnt < n && buf_ptr < (s)->buf_end) {\ | |
698 bit_buf |= *buf_ptr++ << (24 - bit_cnt);\ | |
699 bit_cnt += 8;\ | |
700 if (bit_cnt < n && buf_ptr < (s)->buf_end) {\ | |
701 bit_buf |= *buf_ptr++ << (24 - bit_cnt);\ | |
702 bit_cnt += 8;\ | |
703 if (bit_cnt < n && buf_ptr < (s)->buf_end) {\ | |
704 bit_buf |= *buf_ptr++ << (24 - bit_cnt);\ | |
705 bit_cnt += 8;\ | |
706 }\ | |
707 }\ | |
708 }\ | |
709 val = bit_buf >> (32 - n);\ | |
710 } | |
711 | |
712 /* SHOW_BITS with n1 >= n must be been done before */ | |
713 #define FLUSH_BITS(n)\ | |
714 {\ | |
715 bit_buf <<= n;\ | |
716 bit_cnt -= n;\ | |
717 } | |
718 | |
719 #define SAVE_BITS(s) \ | |
720 {\ | |
721 bit_cnt = (s)->bit_cnt;\ | |
722 bit_buf = (s)->bit_buf;\ | |
723 buf_ptr = (s)->buf_ptr;\ | |
724 } | |
725 | |
726 #define RESTORE_BITS(s) \ | |
727 {\ | |
728 (s)->buf_ptr = buf_ptr;\ | |
729 (s)->bit_buf = bit_buf;\ | |
730 (s)->bit_cnt = bit_cnt;\ | |
731 } | |
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
732 #endif // !ALT_BITSTREAM_READER |
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
733 |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
734 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
|
735 { |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
736 int code, n, nb_bits, index; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
737 INT16 *table_codes; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
738 INT8 *table_bits; |
199
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
739 int bit_cnt; |
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
740 #ifndef ALT_BITSTREAM_READER |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
741 UINT32 bit_buf; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
742 UINT8 *buf_ptr; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
743 #endif |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
744 |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
745 SAVE_BITS(s); |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
746 nb_bits = vlc->bits; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
747 table_codes = vlc->table_codes; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
748 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
|
749 |
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
750 #ifdef FAST_GET_FIRST_VLC |
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
751 SHOW_BITS(s, index, nb_bits); |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
752 code = table_codes[index]; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
753 n = table_bits[index]; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
754 if (n > 0) { |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
755 /* most common case (90%)*/ |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
756 FLUSH_BITS(n); |
306 | 757 #ifdef DUMP_STREAM |
758 { | |
759 int n= bit_cnt - s->index; | |
760 skip_bits(s, n); | |
761 RESTORE_BITS(s); | |
762 } | |
763 #endif | |
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
764 RESTORE_BITS(s); |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
765 return code; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
766 } else if (n == 0) { |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
767 return -1; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
768 } else { |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
769 FLUSH_BITS(nb_bits); |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
770 nb_bits = -n; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
771 table_codes = vlc->table_codes + code; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
772 table_bits = vlc->table_bits + code; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
773 } |
199
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
774 #endif |
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
775 for(;;) { |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
776 SHOW_BITS(s, index, nb_bits); |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
777 code = table_codes[index]; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
778 n = table_bits[index]; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
779 if (n > 0) { |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
780 /* most common case */ |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
781 FLUSH_BITS(n); |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
782 #ifdef STATS |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
783 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
|
784 #endif |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
785 break; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
786 } else if (n == 0) { |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
787 return -1; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
788 } else { |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
789 FLUSH_BITS(nb_bits); |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
790 #ifdef STATS |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
791 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
|
792 #endif |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
793 nb_bits = -n; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
794 table_codes = vlc->table_codes + code; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
795 table_bits = vlc->table_bits + code; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
796 } |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
797 } |
306 | 798 #ifdef DUMP_STREAM |
799 { | |
800 int n= bit_cnt - s->index; | |
801 skip_bits(s, n); | |
802 RESTORE_BITS(s); | |
803 } | |
804 #endif | |
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
805 RESTORE_BITS(s); |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
806 return code; |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
807 } |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
808 |
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
809 |
0 | 810 /* define it to include statistics code (useful only for optimizing |
811 codec efficiency */ | |
812 //#define STATS | |
813 | |
814 #ifdef STATS | |
815 | |
816 enum { | |
817 ST_UNKNOWN, | |
818 ST_DC, | |
819 ST_INTRA_AC, | |
820 ST_INTER_AC, | |
821 ST_INTRA_MB, | |
822 ST_INTER_MB, | |
823 ST_MV, | |
824 ST_NB, | |
825 }; | |
826 | |
827 extern int st_current_index; | |
828 extern unsigned int st_bit_counts[ST_NB]; | |
829 extern unsigned int st_out_bit_counts[ST_NB]; | |
830 | |
831 void print_stats(void); | |
832 #endif | |
833 | |
834 /* misc math functions */ | |
835 | |
151
ae0516eadae2
fixed gcc-3.0.x compilation (by Michael Niedermayer)
nickols_k
parents:
144
diff
changeset
|
836 static inline int av_log2(unsigned int v) |
0 | 837 { |
838 int n; | |
839 | |
840 n = 0; | |
841 if (v & 0xffff0000) { | |
842 v >>= 16; | |
843 n += 16; | |
844 } | |
845 if (v & 0xff00) { | |
846 v >>= 8; | |
847 n += 8; | |
848 } | |
849 if (v & 0xf0) { | |
850 v >>= 4; | |
851 n += 4; | |
852 } | |
853 if (v & 0xc) { | |
854 v >>= 2; | |
855 n += 2; | |
856 } | |
857 if (v & 0x2) { | |
858 n++; | |
859 } | |
860 return n; | |
861 } | |
862 | |
277
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
863 /* median of 3 */ |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
864 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
|
865 { |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
866 int vmin, vmax; |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
867 vmax = vmin = a; |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
868 if (b < vmin) |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
869 vmin = b; |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
870 else |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
871 vmax = b; |
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 if (c < vmin) |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
874 vmin = c; |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
875 else if (c > vmax) |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
876 vmax = c; |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
877 |
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
878 return a + b + c - vmin - vmax; |
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 |
327 | 881 static inline int clip(int a, int amin, int amax) |
882 { | |
883 if (a < amin) | |
884 return amin; | |
885 else if (a > amax) | |
886 return amax; | |
887 else | |
888 return a; | |
889 } | |
890 | |
0 | 891 /* memory */ |
390
48e08d9871da
added proper memory handling functions - fixed include paths
glantau
parents:
370
diff
changeset
|
892 void *av_malloc(int size); |
0 | 893 void *av_mallocz(int size); |
390
48e08d9871da
added proper memory handling functions - fixed include paths
glantau
parents:
370
diff
changeset
|
894 void av_free(void *ptr); |
400 | 895 void __av_freep(void **ptr); |
896 #define av_freep(p) __av_freep((void **)(p)) | |
0 | 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 |
0 | 937 #endif |