Mercurial > libavcodec.hg
annotate common.h @ 390:48e08d9871da libavcodec
added proper memory handling functions - fixed include paths
| author | glantau |
|---|---|
| date | Sat, 18 May 2002 22:56:25 +0000 |
| parents | 0eca28d16cbd |
| children | b0aed401a756 |
| 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 */ |
|
390
48e08d9871da
added proper memory handling functions - fixed include paths
glantau
parents:
370
diff
changeset
|
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; | |
| 53 | |
| 76 | 54 #ifndef __MINGW32__ |
| 64 | 55 #define INT64_C(c) (c ## i64) |
| 56 #define UINT64_C(c) (c ## i64) | |
| 57 | |
| 58 #define inline __inline | |
| 59 | |
| 60 /* | |
| 61 Disable warning messages: | |
| 62 warning C4244: '=' : conversion from 'double' to 'float', possible loss of data | |
| 63 warning C4305: 'argument' : truncation from 'const double' to 'float' | |
| 64 */ | |
| 65 #pragma warning( disable : 4244 ) | |
| 66 #pragma warning( disable : 4305 ) | |
| 67 | |
| 76 | 68 #else |
| 69 #define INT64_C(c) (c ## LL) | |
| 70 #define UINT64_C(c) (c ## ULL) | |
| 71 #endif /* __MINGW32__ */ | |
| 72 | |
| 64 | 73 #define M_PI 3.14159265358979323846 |
| 74 #define M_SQRT2 1.41421356237309504880 /* sqrt(2) */ | |
| 75 | |
| 76 #ifdef _DEBUG | |
| 77 #define DEBUG | |
| 78 #endif | |
| 79 | |
| 80 // code from bits/byteswap.h (C) 1997, 1998 Free Software Foundation, Inc. | |
| 81 #define bswap_32(x) \ | |
| 82 ((((x) & 0xff000000) >> 24) | (((x) & 0x00ff0000) >> 8) | \ | |
| 83 (((x) & 0x0000ff00) << 8) | (((x) & 0x000000ff) << 24)) | |
| 84 #define be2me_32(x) bswap_32(x) | |
| 85 | |
| 86 #define snprintf _snprintf | |
| 87 | |
| 76 | 88 #ifndef __MINGW32__ |
| 89 /* no config.h with VC */ | |
| 64 | 90 #define CONFIG_ENCODERS 1 |
| 91 #define CONFIG_DECODERS 1 | |
| 92 #define CONFIG_AC3 1 | |
| 76 | 93 #endif |
| 64 | 94 |
| 95 #else | |
| 96 | |
| 97 /* unix */ | |
| 98 | |
| 99 #include <inttypes.h> | |
| 100 | |
| 2 | 101 #ifndef __WINE_WINDEF16_H |
| 102 /* workaround for typedef conflict in MPlayer (wine typedefs) */ | |
| 0 | 103 typedef unsigned short UINT16; |
| 104 typedef signed short INT16; | |
| 105 #endif | |
| 106 | |
| 107 typedef unsigned char UINT8; | |
| 108 typedef unsigned int UINT32; | |
| 109 typedef unsigned long long UINT64; | |
| 110 typedef signed char INT8; | |
| 111 typedef signed int INT32; | |
| 112 typedef signed long long INT64; | |
| 113 | |
| 64 | 114 #ifdef HAVE_AV_CONFIG_H |
| 115 | |
| 116 #ifdef __FreeBSD__ | |
| 117 #include <sys/param.h> | |
| 118 #endif | |
| 119 | |
| 120 #ifndef INT64_C | |
| 121 #define INT64_C(c) (c ## LL) | |
| 122 #define UINT64_C(c) (c ## ULL) | |
| 123 #endif | |
| 124 | |
|
390
48e08d9871da
added proper memory handling functions - fixed include paths
glantau
parents:
370
diff
changeset
|
125 #include "bswap.h" |
| 64 | 126 |
| 127 #ifdef USE_FASTMEMCPY | |
| 128 #include "fastmemcpy.h" | |
| 129 #endif | |
| 130 | |
| 76 | 131 #endif /* HAVE_AV_CONFIG_H */ |
| 132 | |
| 133 #endif /* !CONFIG_WIN32 */ | |
| 134 | |
|
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
135 |
| 76 | 136 /* debug stuff */ |
| 137 #ifdef HAVE_AV_CONFIG_H | |
| 138 | |
| 64 | 139 #ifndef DEBUG |
| 140 #define NDEBUG | |
| 141 #endif | |
| 142 #include <assert.h> | |
| 143 | |
| 76 | 144 /* dprintf macros */ |
| 145 #if defined(CONFIG_WIN32) && !defined(__MINGW32__) | |
| 146 | |
| 147 inline void dprintf(const char* fmt,...) {} | |
| 148 | |
| 149 #else | |
| 150 | |
| 151 #ifdef DEBUG | |
| 152 #define dprintf(fmt,args...) printf(fmt, ## args) | |
| 153 #else | |
| 154 #define dprintf(fmt,args...) | |
| 155 #endif | |
| 64 | 156 |
| 157 #endif /* !CONFIG_WIN32 */ | |
| 158 | |
| 76 | 159 #endif /* HAVE_AV_CONFIG_H */ |
| 160 | |
| 344 | 161 /* assume b>0 */ |
| 162 #define ROUNDED_DIV(a,b) (((a)>0 ? (a) + ((b)>>1) : (a) - ((b)>>1))/(b)) | |
| 359 | 163 #define ABS(a) ((a) >= 0 ? (a) : (-(a))) |
| 344 | 164 |
| 0 | 165 /* bit output */ |
| 166 | |
| 167 struct PutBitContext; | |
| 168 | |
| 169 typedef void (*WriteDataFunc)(void *, UINT8 *, int); | |
| 170 | |
| 171 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
|
172 #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
|
173 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
|
174 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
|
175 #else |
|
20
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
176 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
|
177 int bit_left; |
|
20
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
178 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
|
179 #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
|
180 INT64 data_out_size; /* in bytes */ |
| 0 | 181 } PutBitContext; |
| 182 | |
| 183 void init_put_bits(PutBitContext *s, | |
| 184 UINT8 *buffer, int buffer_size, | |
| 185 void *opaque, | |
| 186 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
|
187 |
| 85 | 188 INT64 get_bit_count(PutBitContext *s); /* XXX: change function name */ |
| 0 | 189 void align_put_bits(PutBitContext *s); |
| 190 void flush_put_bits(PutBitContext *s); | |
| 358 | 191 void put_string(PutBitContext * pbc, char *s); |
| 0 | 192 |
| 193 /* jpeg specific put_bits */ | |
| 194 void jflush_put_bits(PutBitContext *s); | |
| 195 | |
| 196 /* bit input */ | |
| 197 | |
| 198 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
|
199 #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
|
200 int index; |
|
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
201 UINT8 *buffer; |
|
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
202 #else |
|
20
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
203 UINT32 bit_buf; |
| 0 | 204 int bit_cnt; |
|
20
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
205 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
|
206 #endif |
|
277
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
207 int size; |
| 0 | 208 } GetBitContext; |
| 209 | |
| 290 | 210 static inline int get_bits_count(GetBitContext *s); |
| 211 | |
| 0 | 212 typedef struct VLC { |
| 213 int bits; | |
| 214 INT16 *table_codes; | |
| 215 INT8 *table_bits; | |
| 216 int table_size, table_allocated; | |
| 217 } VLC; | |
| 218 | |
|
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
219 /* 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
|
220 #ifdef ARCH_X86 |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
221 #define unaligned32(a) (*(UINT32*)(a)) |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
222 #else |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
223 #ifdef __GNUC__ |
|
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 struct Unaligned { |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
226 uint32_t i; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
227 } __attribute__((packed)); |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
228 |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
229 return ((const struct Unaligned *) v)->i; |
|
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 #elif defined(__DECC) |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
232 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
|
233 return *(const __unaligned uint32_t *) v; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
234 } |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
235 #else |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
236 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
|
237 return *(const uint32_t *) v; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
238 } |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
239 #endif |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
240 #endif //!ARCH_X86 |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
241 |
|
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
|
242 #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
|
243 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
|
244 { |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
245 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
|
246 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
|
247 |
|
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 #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
|
249 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
|
250 #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
|
251 // 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
|
252 assert(n == 32 || value < (1U << n)); |
| 306 | 253 |
|
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
|
254 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
|
255 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
|
256 |
|
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 // 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
|
258 /* 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
|
259 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
|
260 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
|
261 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
|
262 } 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
|
263 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
|
264 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
|
265 *(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
|
266 //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
|
267 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
|
268 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
|
269 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
|
270 } |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
271 |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
272 s->bit_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
|
273 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
|
274 } |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
275 #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
|
276 |
|
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
|
277 |
|
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
|
278 #ifdef ALT_BITSTREAM_WRITER |
|
235
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
279 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
|
280 { |
|
235
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
281 #ifdef ALIGNED_BITSTREAM_WRITER |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
282 #ifdef ARCH_X86 |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
283 asm volatile( |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
284 "movl %0, %%ecx \n\t" |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
285 "xorl %%eax, %%eax \n\t" |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
286 "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
|
287 "shrl %%cl, %1 \n\t" |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
288 "movl %0, %%ecx \n\t" |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
289 "shrl $3, %%ecx \n\t" |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
290 "andl $0xFFFFFFFC, %%ecx \n\t" |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
291 "bswapl %1 \n\t" |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
292 "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
|
293 "bswapl %%eax \n\t" |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
294 "addl %3, %0 \n\t" |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
295 "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
|
296 : "=&r" (s->index), "=&r" (value) |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
297 : "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
|
298 : "%eax", "%ecx" |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
299 ); |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
300 #else |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
301 int index= s->index; |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
302 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
|
303 |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
304 value<<= 32-n; |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
305 |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
306 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
|
307 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
|
308 //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
|
309 index+= n; |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
310 s->index= index; |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
311 #endif |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
312 #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
|
313 #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
|
314 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
|
315 "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
|
316 "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
|
317 "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
|
318 "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
|
319 "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
|
320 "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
|
321 "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
|
322 "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
|
323 "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
|
324 "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
|
325 "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
|
326 : "=&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
|
327 : "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
|
328 : "%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
|
329 ); |
|
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 #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
|
331 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
|
332 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
|
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 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
|
335 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
|
336 //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
|
337 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
|
338 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
|
339 #endif |
|
235
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
340 #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
|
341 } |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
342 #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
|
343 |
|
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
|
344 #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
|
345 /* 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
|
346 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
|
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 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
|
349 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
|
350 |
|
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 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
|
352 |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
353 bit_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
|
354 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
|
355 |
|
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 //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
|
357 /* 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
|
358 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
|
359 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
|
360 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
|
361 } 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
|
362 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
|
363 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
|
364 /* 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
|
365 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
|
366 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
|
367 *(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
|
368 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
|
369 *(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
|
370 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
|
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 |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
373 bit_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
|
374 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
|
375 } |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
376 |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
377 s->bit_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
|
378 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
|
379 } |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
380 #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
|
381 |
|
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
|
382 |
|
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
|
383 #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
|
384 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
|
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 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
|
387 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
|
388 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
|
389 //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
|
390 |
|
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 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
|
392 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
|
393 { |
|
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 /* 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
|
395 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
|
396 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
|
397 |
|
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 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
|
399 |
|
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 *(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
|
401 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
|
402 *(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
|
403 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
|
404 } |
|
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 *(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
|
406 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
|
407 *(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
|
408 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
|
409 } |
|
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 *(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
|
411 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
|
412 *(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
|
413 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
|
414 } |
|
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 *(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
|
416 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
|
417 *(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
|
418 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
|
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 *((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
|
421 } |
|
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 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
|
423 { |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
424 ptr[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
|
425 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
|
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 |
|
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 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
|
429 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
|
430 } |
|
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 |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
434 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
|
435 { |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
436 #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
|
437 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
|
438 #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
|
439 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
|
440 #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
|
441 } |
|
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
|
442 |
| 0 | 443 void init_get_bits(GetBitContext *s, |
| 444 UINT8 *buffer, int buffer_size); | |
| 445 | |
|
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
446 #ifndef ALT_BITSTREAM_READER |
|
20
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
447 unsigned int get_bits_long(GetBitContext *s, int n); |
| 144 | 448 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
|
449 #endif |
|
20
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
450 |
|
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
451 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
|
452 #ifdef ALT_BITSTREAM_READER |
|
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
453 #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
|
454 int index= s->index; |
|
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
455 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
|
456 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
|
457 #ifdef ARCH_X86 |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
458 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
|
459 : "=r" (result1) |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
460 : "0" (result1), "r" (result2), "c" (index)); |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
461 #else |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
462 result1<<= (index&0x1F); |
|
199
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
463 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
|
464 result1|= result2; |
|
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
465 #endif |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
466 result1>>= 32 - n; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
467 index+= n; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
468 s->index= index; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
469 |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
470 return result1; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
471 #else //ALIGNED_BITSTREAM |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
472 int index= s->index; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
473 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
|
474 |
|
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
475 result<<= (index&0x07); |
|
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
476 result>>= 32 - n; |
|
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
477 index+= n; |
|
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
478 s->index= index; |
| 306 | 479 #ifdef DUMP_STREAM |
| 480 while(n){ | |
| 481 printf("%d", (result>>(n-1))&1); | |
| 482 n--; | |
| 483 } | |
| 484 printf(" "); | |
| 485 #endif | |
|
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
486 return result; |
|
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
487 #endif //!ALIGNED_BITSTREAM |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
488 #else //ALT_BITSTREAM_READER |
|
20
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
489 if(s->bit_cnt>=n){ |
|
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
490 /* most common case here */ |
|
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
491 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
|
492 s->bit_buf <<= n; |
|
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
493 s->bit_cnt -= n; |
|
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
494 #ifdef STATS |
|
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
495 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
|
496 #endif |
|
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
497 return val; |
|
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
498 } |
|
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
499 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
|
500 #endif //!ALT_BITSTREAM_READER |
|
20
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
501 } |
|
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
502 |
| 21 | 503 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
|
504 #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
|
505 int index= s->index; |
|
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
506 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
|
507 result<<= (index&0x07); |
|
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
508 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
|
509 index++; |
|
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
510 s->index= index; |
|
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
511 |
| 306 | 512 #ifdef DUMP_STREAM |
| 513 printf("%d ", result); | |
| 514 #endif | |
|
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
515 return result; |
|
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
516 #else |
| 21 | 517 if(s->bit_cnt>0){ |
| 518 /* most common case here */ | |
| 519 unsigned int val = s->bit_buf >> 31; | |
| 520 s->bit_buf <<= 1; | |
| 521 s->bit_cnt--; | |
| 522 #ifdef STATS | |
| 523 st_bit_counts[st_current_index]++; | |
| 524 #endif | |
| 525 return val; | |
| 526 } | |
| 527 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
|
528 #endif |
| 21 | 529 } |
| 530 | |
| 144 | 531 /* This function is identical to get_bits(), the only */ |
| 532 /* diference is that it doesn't touch the buffer */ | |
| 533 /* it is usefull to see the buffer. */ | |
| 534 static inline unsigned int show_bits(GetBitContext *s, int n) | |
| 535 { | |
|
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
536 #ifdef ALT_BITSTREAM_READER |
|
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
537 #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
|
538 int index= s->index; |
|
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
539 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
|
540 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
|
541 #ifdef ARCH_X86 |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
542 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
|
543 : "=r" (result1) |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
544 : "0" (result1), "r" (result2), "c" (index)); |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
545 #else |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
546 result1<<= (index&0x1F); |
|
199
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
547 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
|
548 result1|= result2; |
|
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
549 #endif |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
550 result1>>= 32 - n; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
551 |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
552 return result1; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
553 #else //ALIGNED_BITSTREAM |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
554 int index= s->index; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
555 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
|
556 |
|
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
557 result<<= (index&0x07); |
|
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
558 result>>= 32 - n; |
|
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
559 |
|
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
560 return result; |
|
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
561 #endif //!ALIGNED_BITSTREAM |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
562 #else //ALT_BITSTREAM_READER |
| 144 | 563 if(s->bit_cnt>=n) { |
| 564 /* most common case here */ | |
| 565 unsigned int val = s->bit_buf >> (32 - n); | |
| 566 return val; | |
| 567 } | |
| 568 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
|
569 #endif //!ALT_BITSTREAM_READER |
| 144 | 570 } |
| 571 | |
| 290 | 572 static inline int show_aligned_bits(GetBitContext *s, int offset, int n) |
| 573 { | |
| 574 #ifdef ALT_BITSTREAM_READER | |
| 575 #ifdef ALIGNED_BITSTREAM | |
| 576 int index= (s->index + offset + 7)&(~7); | |
| 577 uint32_t result1= be2me_32( ((uint32_t *)s->buffer)[index>>5] ); | |
| 578 uint32_t result2= be2me_32( ((uint32_t *)s->buffer)[(index>>5) + 1] ); | |
| 579 #ifdef ARCH_X86 | |
| 580 asm ("shldl %%cl, %2, %0\n\t" | |
| 581 : "=r" (result1) | |
| 582 : "0" (result1), "r" (result2), "c" (index)); | |
| 583 #else | |
| 584 result1<<= (index&0x1F); | |
| 585 result2= (result2>>1) >> (31-(index&0x1F)); | |
| 586 result1|= result2; | |
| 587 #endif | |
| 588 result1>>= 32 - n; | |
| 589 | |
| 590 return result1; | |
| 591 #else //ALIGNED_BITSTREAM | |
| 592 int index= (s->index + offset + 7)>>3; | |
| 593 uint32_t result= be2me_32( unaligned32( ((uint8_t *)s->buffer)+index ) ); | |
| 594 | |
| 595 result>>= 32 - n; | |
| 596 | |
| 597 return result; | |
| 598 #endif //!ALIGNED_BITSTREAM | |
| 599 #else //ALT_BITSTREAM_READER | |
| 600 int index= (get_bits_count(s) + offset + 7)>>3; | |
| 601 uint32_t result= be2me_32( unaligned32( ((uint8_t *)s->buf)+index ) ); | |
| 602 | |
| 603 result>>= 32 - n; | |
| 604 //printf(" %X %X %d \n", (int)(((uint8_t *)s->buf)+index ), (int)s->buf_ptr, s->bit_cnt); | |
| 605 return result; | |
| 606 #endif //!ALT_BITSTREAM_READER | |
| 607 } | |
| 608 | |
| 21 | 609 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
|
610 #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
|
611 s->index+= n; |
| 306 | 612 #ifdef DUMP_STREAM |
| 613 { | |
| 614 int result; | |
| 615 s->index-= n; | |
| 616 result= get_bits(s, n); | |
| 617 } | |
| 618 #endif | |
| 619 | |
|
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
620 #else |
| 21 | 621 if(s->bit_cnt>=n){ |
| 622 /* most common case here */ | |
| 623 s->bit_buf <<= n; | |
| 624 s->bit_cnt -= n; | |
| 625 #ifdef STATS | |
| 626 st_bit_counts[st_current_index] += n; | |
| 627 #endif | |
| 628 } else { | |
| 629 get_bits_long(s,n); | |
| 630 } | |
|
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
631 #endif |
| 21 | 632 } |
| 633 | |
| 634 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
|
635 #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
|
636 s->index++; |
| 306 | 637 #ifdef DUMP_STREAM |
| 638 s->index--; | |
| 639 printf("%d ", get_bits1(s)); | |
| 640 #endif | |
|
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
641 #else |
| 21 | 642 if(s->bit_cnt>0){ |
| 643 /* most common case here */ | |
| 644 s->bit_buf <<= 1; | |
| 645 s->bit_cnt--; | |
| 646 #ifdef STATS | |
| 647 st_bit_counts[st_current_index]++; | |
| 648 #endif | |
| 649 } else { | |
| 650 get_bits_long(s,1); | |
| 651 } | |
|
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
652 #endif |
| 21 | 653 } |
| 654 | |
| 85 | 655 static inline int get_bits_count(GetBitContext *s) |
| 656 { | |
|
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
657 #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
|
658 return s->index; |
|
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
659 #else |
| 85 | 660 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
|
661 #endif |
| 85 | 662 } |
| 21 | 663 |
|
264
28c5c62b1c4c
support decoding (with mplayer) of 3 .mp4 files from mphq
michaelni
parents:
238
diff
changeset
|
664 int check_marker(GetBitContext *s, char *msg); |
| 0 | 665 void align_get_bits(GetBitContext *s); |
| 666 int init_vlc(VLC *vlc, int nb_bits, int nb_codes, | |
| 667 const void *bits, int bits_wrap, int bits_size, | |
| 668 const void *codes, int codes_wrap, int codes_size); | |
| 669 void free_vlc(VLC *vlc); | |
| 670 | |
|
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
671 #ifdef ALT_BITSTREAM_READER |
|
199
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
672 #ifdef ALIGNED_BITSTREAM |
|
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
673 #ifdef ARCH_X86 |
|
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
674 #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
|
675 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
|
676 {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
|
677 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
|
678 : "=r" (val)\ |
|
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
679 : "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
|
680 ((uint32_t)val)>>= 32 - n;} |
|
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
681 #else //ARCH_X86 |
|
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( ((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
|
684 {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
|
685 val<<= (bit_cnt&0x1F);\ |
|
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
686 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
|
687 val|= result2;\ |
|
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
688 ((uint32_t)val)>>= 32 - n;} |
|
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
689 #endif //!ARCH_X86 |
|
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
690 #else //ALIGNED_BITSTREAM |
|
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
691 #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
|
692 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
|
693 val<<= (bit_cnt&0x07);\ |
|
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
694 ((uint32_t)val)>>= 32 - n; |
|
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
695 #endif // !ALIGNED_BITSTREAM |
|
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
696 #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
|
697 #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
|
698 #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
|
699 #else |
|
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
700 |
| 0 | 701 /* macro to go faster */ |
| 702 /* n must be <= 24 */ | |
| 703 /* XXX: optimize buffer end test */ | |
| 704 #define SHOW_BITS(s, val, n)\ | |
| 705 {\ | |
| 706 if (bit_cnt < n && buf_ptr < (s)->buf_end) {\ | |
| 707 bit_buf |= *buf_ptr++ << (24 - bit_cnt);\ | |
| 708 bit_cnt += 8;\ | |
| 709 if (bit_cnt < n && buf_ptr < (s)->buf_end) {\ | |
| 710 bit_buf |= *buf_ptr++ << (24 - bit_cnt);\ | |
| 711 bit_cnt += 8;\ | |
| 712 if (bit_cnt < n && buf_ptr < (s)->buf_end) {\ | |
| 713 bit_buf |= *buf_ptr++ << (24 - bit_cnt);\ | |
| 714 bit_cnt += 8;\ | |
| 715 }\ | |
| 716 }\ | |
| 717 }\ | |
| 718 val = bit_buf >> (32 - n);\ | |
| 719 } | |
| 720 | |
| 721 /* SHOW_BITS with n1 >= n must be been done before */ | |
| 722 #define FLUSH_BITS(n)\ | |
| 723 {\ | |
| 724 bit_buf <<= n;\ | |
| 725 bit_cnt -= n;\ | |
| 726 } | |
| 727 | |
| 728 #define SAVE_BITS(s) \ | |
| 729 {\ | |
| 730 bit_cnt = (s)->bit_cnt;\ | |
| 731 bit_buf = (s)->bit_buf;\ | |
| 732 buf_ptr = (s)->buf_ptr;\ | |
| 733 } | |
| 734 | |
| 735 #define RESTORE_BITS(s) \ | |
| 736 {\ | |
| 737 (s)->buf_ptr = buf_ptr;\ | |
| 738 (s)->bit_buf = bit_buf;\ | |
| 739 (s)->bit_cnt = bit_cnt;\ | |
| 740 } | |
|
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
741 #endif // !ALT_BITSTREAM_READER |
|
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
742 |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
743 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
|
744 { |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
745 int code, n, nb_bits, index; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
746 INT16 *table_codes; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
747 INT8 *table_bits; |
|
199
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
748 int bit_cnt; |
|
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
749 #ifndef ALT_BITSTREAM_READER |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
750 UINT32 bit_buf; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
751 UINT8 *buf_ptr; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
752 #endif |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
753 |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
754 SAVE_BITS(s); |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
755 nb_bits = vlc->bits; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
756 table_codes = vlc->table_codes; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
757 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
|
758 |
|
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
759 #ifdef FAST_GET_FIRST_VLC |
|
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
760 SHOW_BITS(s, index, nb_bits); |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
761 code = table_codes[index]; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
762 n = table_bits[index]; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
763 if (n > 0) { |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
764 /* most common case (90%)*/ |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
765 FLUSH_BITS(n); |
| 306 | 766 #ifdef DUMP_STREAM |
| 767 { | |
| 768 int n= bit_cnt - s->index; | |
| 769 skip_bits(s, n); | |
| 770 RESTORE_BITS(s); | |
| 771 } | |
| 772 #endif | |
|
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
773 RESTORE_BITS(s); |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
774 return code; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
775 } else if (n == 0) { |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
776 return -1; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
777 } else { |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
778 FLUSH_BITS(nb_bits); |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
779 nb_bits = -n; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
780 table_codes = vlc->table_codes + code; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
781 table_bits = vlc->table_bits + code; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
782 } |
|
199
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
783 #endif |
|
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
784 for(;;) { |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
785 SHOW_BITS(s, index, nb_bits); |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
786 code = table_codes[index]; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
787 n = table_bits[index]; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
788 if (n > 0) { |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
789 /* most common case */ |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
790 FLUSH_BITS(n); |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
791 #ifdef STATS |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
792 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
|
793 #endif |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
794 break; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
795 } else if (n == 0) { |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
796 return -1; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
797 } else { |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
798 FLUSH_BITS(nb_bits); |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
799 #ifdef STATS |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
800 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
|
801 #endif |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
802 nb_bits = -n; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
803 table_codes = vlc->table_codes + code; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
804 table_bits = vlc->table_bits + code; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
805 } |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
806 } |
| 306 | 807 #ifdef DUMP_STREAM |
| 808 { | |
| 809 int n= bit_cnt - s->index; | |
| 810 skip_bits(s, n); | |
| 811 RESTORE_BITS(s); | |
| 812 } | |
| 813 #endif | |
|
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
814 RESTORE_BITS(s); |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
815 return code; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
816 } |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
817 |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
818 |
| 0 | 819 /* define it to include statistics code (useful only for optimizing |
| 820 codec efficiency */ | |
| 821 //#define STATS | |
| 822 | |
| 823 #ifdef STATS | |
| 824 | |
| 825 enum { | |
| 826 ST_UNKNOWN, | |
| 827 ST_DC, | |
| 828 ST_INTRA_AC, | |
| 829 ST_INTER_AC, | |
| 830 ST_INTRA_MB, | |
| 831 ST_INTER_MB, | |
| 832 ST_MV, | |
| 833 ST_NB, | |
| 834 }; | |
| 835 | |
| 836 extern int st_current_index; | |
| 837 extern unsigned int st_bit_counts[ST_NB]; | |
| 838 extern unsigned int st_out_bit_counts[ST_NB]; | |
| 839 | |
| 840 void print_stats(void); | |
| 841 #endif | |
| 842 | |
| 843 /* misc math functions */ | |
| 844 | |
|
151
ae0516eadae2
fixed gcc-3.0.x compilation (by Michael Niedermayer)
nickols_k
parents:
144
diff
changeset
|
845 static inline int av_log2(unsigned int v) |
| 0 | 846 { |
| 847 int n; | |
| 848 | |
| 849 n = 0; | |
| 850 if (v & 0xffff0000) { | |
| 851 v >>= 16; | |
| 852 n += 16; | |
| 853 } | |
| 854 if (v & 0xff00) { | |
| 855 v >>= 8; | |
| 856 n += 8; | |
| 857 } | |
| 858 if (v & 0xf0) { | |
| 859 v >>= 4; | |
| 860 n += 4; | |
| 861 } | |
| 862 if (v & 0xc) { | |
| 863 v >>= 2; | |
| 864 n += 2; | |
| 865 } | |
| 866 if (v & 0x2) { | |
| 867 n++; | |
| 868 } | |
| 869 return n; | |
| 870 } | |
| 871 | |
|
277
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
872 /* median of 3 */ |
|
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
873 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
|
874 { |
|
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
875 int vmin, vmax; |
|
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
876 vmax = vmin = a; |
|
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
877 if (b < vmin) |
|
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
878 vmin = b; |
|
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
879 else |
|
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
880 vmax = b; |
|
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
881 |
|
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
882 if (c < vmin) |
|
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
883 vmin = c; |
|
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
884 else if (c > vmax) |
|
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
885 vmax = c; |
|
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
886 |
|
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
887 return a + b + c - vmin - vmax; |
|
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
888 } |
|
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
889 |
| 327 | 890 static inline int clip(int a, int amin, int amax) |
| 891 { | |
| 892 if (a < amin) | |
| 893 return amin; | |
| 894 else if (a > amax) | |
| 895 return amax; | |
| 896 else | |
| 897 return a; | |
| 898 } | |
| 899 | |
| 0 | 900 /* memory */ |
|
390
48e08d9871da
added proper memory handling functions - fixed include paths
glantau
parents:
370
diff
changeset
|
901 void *av_malloc(int size); |
| 0 | 902 void *av_mallocz(int size); |
|
390
48e08d9871da
added proper memory handling functions - fixed include paths
glantau
parents:
370
diff
changeset
|
903 void av_free(void *ptr); |
|
48e08d9871da
added proper memory handling functions - fixed include paths
glantau
parents:
370
diff
changeset
|
904 #define av_freep(p) do { av_free(*p); *p = NULL; } while (0) |
| 0 | 905 |
| 324 | 906 /* math */ |
| 907 int ff_gcd(int a, int b); | |
| 908 | |
|
370
0eca28d16cbd
clamp intra matrix to 8bit for mjpeg (workaround for qscale>=25)
al3x
parents:
359
diff
changeset
|
909 #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
|
910 |
| 0 | 911 #endif |
