Mercurial > libavcodec.hg
annotate common.h @ 370:0eca28d16cbd libavcodec
clamp intra matrix to 8bit for mjpeg (workaround for qscale>=25)
| author | al3x |
|---|---|
| date | Tue, 07 May 2002 17:55:02 +0000 |
| parents | df0736462b9f |
| children | 48e08d9871da |
| 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 */ |
| 0 | 20 #include "../config.h" |
| 64 | 21 |
| 22 #include <stdlib.h> | |
| 23 #include <stdio.h> | |
| 24 #include <string.h> | |
| 25 #include <errno.h> | |
| 26 | |
| 27 #ifndef ENODATA | |
| 28 #define ENODATA 61 | |
| 10 | 29 #endif |
| 0 | 30 |
| 64 | 31 #endif |
| 32 | |
| 33 #ifdef CONFIG_WIN32 | |
| 34 | |
| 35 /* windows */ | |
| 36 | |
| 37 typedef unsigned short UINT16; | |
| 38 typedef signed short INT16; | |
| 39 typedef unsigned char UINT8; | |
| 40 typedef unsigned int UINT32; | |
| 41 typedef unsigned __int64 UINT64; | |
| 42 typedef signed char INT8; | |
| 43 typedef signed int INT32; | |
| 44 typedef signed __int64 INT64; | |
| 45 | |
| 46 typedef UINT8 uint8_t; | |
| 47 typedef INT8 int8_t; | |
| 48 typedef UINT16 uint16_t; | |
| 49 typedef INT16 int16_t; | |
| 50 typedef UINT32 uint32_t; | |
| 51 typedef INT32 int32_t; | |
| 52 | |
| 76 | 53 #ifndef __MINGW32__ |
| 64 | 54 #define INT64_C(c) (c ## i64) |
| 55 #define UINT64_C(c) (c ## i64) | |
| 56 | |
| 57 #define inline __inline | |
| 58 | |
| 59 /* | |
| 60 Disable warning messages: | |
| 61 warning C4244: '=' : conversion from 'double' to 'float', possible loss of data | |
| 62 warning C4305: 'argument' : truncation from 'const double' to 'float' | |
| 63 */ | |
| 64 #pragma warning( disable : 4244 ) | |
| 65 #pragma warning( disable : 4305 ) | |
| 66 | |
| 76 | 67 #else |
| 68 #define INT64_C(c) (c ## LL) | |
| 69 #define UINT64_C(c) (c ## ULL) | |
| 70 #endif /* __MINGW32__ */ | |
| 71 | |
| 64 | 72 #define M_PI 3.14159265358979323846 |
| 73 #define M_SQRT2 1.41421356237309504880 /* sqrt(2) */ | |
| 74 | |
| 75 #ifdef _DEBUG | |
| 76 #define DEBUG | |
| 77 #endif | |
| 78 | |
| 79 // code from bits/byteswap.h (C) 1997, 1998 Free Software Foundation, Inc. | |
| 80 #define bswap_32(x) \ | |
| 81 ((((x) & 0xff000000) >> 24) | (((x) & 0x00ff0000) >> 8) | \ | |
| 82 (((x) & 0x0000ff00) << 8) | (((x) & 0x000000ff) << 24)) | |
| 83 #define be2me_32(x) bswap_32(x) | |
| 84 | |
| 85 #define snprintf _snprintf | |
| 86 | |
| 76 | 87 #ifndef __MINGW32__ |
| 88 /* no config.h with VC */ | |
| 64 | 89 #define CONFIG_ENCODERS 1 |
| 90 #define CONFIG_DECODERS 1 | |
| 91 #define CONFIG_AC3 1 | |
| 76 | 92 #endif |
| 64 | 93 |
| 94 #else | |
| 95 | |
| 96 /* unix */ | |
| 97 | |
| 98 #include <inttypes.h> | |
| 99 | |
| 2 | 100 #ifndef __WINE_WINDEF16_H |
| 101 /* workaround for typedef conflict in MPlayer (wine typedefs) */ | |
| 0 | 102 typedef unsigned short UINT16; |
| 103 typedef signed short INT16; | |
| 104 #endif | |
| 105 | |
| 106 typedef unsigned char UINT8; | |
| 107 typedef unsigned int UINT32; | |
| 108 typedef unsigned long long UINT64; | |
| 109 typedef signed char INT8; | |
| 110 typedef signed int INT32; | |
| 111 typedef signed long long INT64; | |
| 112 | |
| 64 | 113 #ifdef HAVE_AV_CONFIG_H |
| 114 | |
| 115 #ifdef __FreeBSD__ | |
| 116 #include <sys/param.h> | |
| 117 #endif | |
| 118 | |
| 119 #ifndef INT64_C | |
| 120 #define INT64_C(c) (c ## LL) | |
| 121 #define UINT64_C(c) (c ## ULL) | |
| 122 #endif | |
| 123 | |
| 124 #include "../bswap.h" | |
| 125 | |
| 126 #ifdef USE_FASTMEMCPY | |
| 127 #include "fastmemcpy.h" | |
| 128 #endif | |
| 129 | |
| 76 | 130 #endif /* HAVE_AV_CONFIG_H */ |
| 131 | |
| 132 #endif /* !CONFIG_WIN32 */ | |
| 133 | |
|
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
134 |
| 76 | 135 /* debug stuff */ |
| 136 #ifdef HAVE_AV_CONFIG_H | |
| 137 | |
| 64 | 138 #ifndef DEBUG |
| 139 #define NDEBUG | |
| 140 #endif | |
| 141 #include <assert.h> | |
| 142 | |
| 76 | 143 /* dprintf macros */ |
| 144 #if defined(CONFIG_WIN32) && !defined(__MINGW32__) | |
| 145 | |
| 146 inline void dprintf(const char* fmt,...) {} | |
| 147 | |
| 148 #else | |
| 149 | |
| 150 #ifdef DEBUG | |
| 151 #define dprintf(fmt,args...) printf(fmt, ## args) | |
| 152 #else | |
| 153 #define dprintf(fmt,args...) | |
| 154 #endif | |
| 64 | 155 |
| 156 #endif /* !CONFIG_WIN32 */ | |
| 157 | |
| 76 | 158 #endif /* HAVE_AV_CONFIG_H */ |
| 159 | |
| 344 | 160 /* assume b>0 */ |
| 161 #define ROUNDED_DIV(a,b) (((a)>0 ? (a) + ((b)>>1) : (a) - ((b)>>1))/(b)) | |
| 359 | 162 #define ABS(a) ((a) >= 0 ? (a) : (-(a))) |
| 344 | 163 |
| 0 | 164 /* bit output */ |
| 165 | |
| 166 struct PutBitContext; | |
| 167 | |
| 168 typedef void (*WriteDataFunc)(void *, UINT8 *, int); | |
| 169 | |
| 170 typedef struct PutBitContext { | |
|
234
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
171 #ifdef ALT_BITSTREAM_WRITER |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
172 UINT8 *buf, *buf_end; |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
173 int index; |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
174 #else |
|
20
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
175 UINT32 bit_buf; |
|
238
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
176 int bit_left; |
|
20
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
177 UINT8 *buf, *buf_ptr, *buf_end; |
|
234
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
178 #endif |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
179 INT64 data_out_size; /* in bytes */ |
| 0 | 180 } PutBitContext; |
| 181 | |
| 182 void init_put_bits(PutBitContext *s, | |
| 183 UINT8 *buffer, int buffer_size, | |
| 184 void *opaque, | |
| 185 void (*write_data)(void *, UINT8 *, int)); | |
|
234
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
186 |
| 85 | 187 INT64 get_bit_count(PutBitContext *s); /* XXX: change function name */ |
| 0 | 188 void align_put_bits(PutBitContext *s); |
| 189 void flush_put_bits(PutBitContext *s); | |
| 358 | 190 void put_string(PutBitContext * pbc, char *s); |
| 0 | 191 |
| 192 /* jpeg specific put_bits */ | |
| 193 void jflush_put_bits(PutBitContext *s); | |
| 194 | |
| 195 /* bit input */ | |
| 196 | |
| 197 typedef struct GetBitContext { | |
|
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
198 #ifdef ALT_BITSTREAM_READER |
|
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
199 int index; |
|
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
200 UINT8 *buffer; |
|
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
201 #else |
|
20
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
202 UINT32 bit_buf; |
| 0 | 203 int bit_cnt; |
|
20
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
204 UINT8 *buf, *buf_ptr, *buf_end; |
|
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
205 #endif |
|
277
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
206 int size; |
| 0 | 207 } GetBitContext; |
| 208 | |
| 290 | 209 static inline int get_bits_count(GetBitContext *s); |
| 210 | |
| 0 | 211 typedef struct VLC { |
| 212 int bits; | |
| 213 INT16 *table_codes; | |
| 214 INT8 *table_bits; | |
| 215 int table_size, table_allocated; | |
| 216 } VLC; | |
| 217 | |
|
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
218 /* used to avoid missaligned exceptions on some archs (alpha, ...) */ |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
219 #ifdef ARCH_X86 |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
220 #define unaligned32(a) (*(UINT32*)(a)) |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
221 #else |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
222 #ifdef __GNUC__ |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
223 static inline uint32_t unaligned32(const void *v) { |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
224 struct Unaligned { |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
225 uint32_t i; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
226 } __attribute__((packed)); |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
227 |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
228 return ((const struct Unaligned *) v)->i; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
229 } |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
230 #elif defined(__DECC) |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
231 static inline uint32_t unaligned32(const void *v) { |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
232 return *(const __unaligned uint32_t *) v; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
233 } |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
234 #else |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
235 static inline uint32_t unaligned32(const void *v) { |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
236 return *(const uint32_t *) v; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
237 } |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
238 #endif |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
239 #endif //!ARCH_X86 |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
240 |
|
238
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
241 #ifndef ALT_BITSTREAM_WRITER |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
242 static inline void put_bits(PutBitContext *s, int n, unsigned int value) |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
243 { |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
244 unsigned int bit_buf; |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
245 int bit_left; |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
246 |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
247 #ifdef STATS |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
248 st_out_bit_counts[st_current_index] += n; |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
249 #endif |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
250 // printf("put_bits=%d %x\n", n, value); |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
251 assert(n == 32 || value < (1U << n)); |
| 306 | 252 |
|
238
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
253 bit_buf = s->bit_buf; |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
254 bit_left = s->bit_left; |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
255 |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
256 // printf("n=%d value=%x cnt=%d buf=%x\n", n, value, bit_cnt, bit_buf); |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
257 /* XXX: optimize */ |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
258 if (n < bit_left) { |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
259 bit_buf = (bit_buf<<n) | value; |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
260 bit_left-=n; |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
261 } else { |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
262 bit_buf<<=bit_left; |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
263 bit_buf |= value >> (n - bit_left); |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
264 *(UINT32 *)s->buf_ptr = be2me_32(bit_buf); |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
265 //printf("bitbuf = %08x\n", bit_buf); |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
266 s->buf_ptr+=4; |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
267 bit_left+=32 - n; |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
268 bit_buf = value; |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
269 } |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
270 |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
271 s->bit_buf = bit_buf; |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
272 s->bit_left = bit_left; |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
273 } |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
274 #endif |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
275 |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
276 |
|
234
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
277 #ifdef ALT_BITSTREAM_WRITER |
|
235
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
278 static inline void put_bits(PutBitContext *s, int n, unsigned int value) |
|
234
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
279 { |
|
235
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
280 #ifdef ALIGNED_BITSTREAM_WRITER |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
281 #ifdef ARCH_X86 |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
282 asm volatile( |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
283 "movl %0, %%ecx \n\t" |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
284 "xorl %%eax, %%eax \n\t" |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
285 "shrdl %%cl, %1, %%eax \n\t" |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
286 "shrl %%cl, %1 \n\t" |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
287 "movl %0, %%ecx \n\t" |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
288 "shrl $3, %%ecx \n\t" |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
289 "andl $0xFFFFFFFC, %%ecx \n\t" |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
290 "bswapl %1 \n\t" |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
291 "orl %1, (%2, %%ecx) \n\t" |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
292 "bswapl %%eax \n\t" |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
293 "addl %3, %0 \n\t" |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
294 "movl %%eax, 4(%2, %%ecx) \n\t" |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
295 : "=&r" (s->index), "=&r" (value) |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
296 : "r" (s->buf), "r" (n), "0" (s->index), "1" (value<<(-n)) |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
297 : "%eax", "%ecx" |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
298 ); |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
299 #else |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
300 int index= s->index; |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
301 uint32_t *ptr= ((uint32_t *)s->buf)+(index>>5); |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
302 |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
303 value<<= 32-n; |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
304 |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
305 ptr[0] |= be2me_32(value>>(index&31)); |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
306 ptr[1] = be2me_32(value<<(32-(index&31))); |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
307 //if(n>24) printf("%d %d\n", n, value); |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
308 index+= n; |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
309 s->index= index; |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
310 #endif |
|
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
311 #else //ALIGNED_BITSTREAM_WRITER |
|
234
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
312 #ifdef ARCH_X86 |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
313 asm volatile( |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
314 "movl $7, %%ecx \n\t" |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
315 "andl %0, %%ecx \n\t" |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
316 "addl %3, %%ecx \n\t" |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
317 "negl %%ecx \n\t" |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
318 "shll %%cl, %1 \n\t" |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
319 "bswapl %1 \n\t" |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
320 "movl %0, %%ecx \n\t" |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
321 "shrl $3, %%ecx \n\t" |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
322 "orl %1, (%%ecx, %2) \n\t" |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
323 "addl %3, %0 \n\t" |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
324 "movl $0, 4(%%ecx, %2) \n\t" |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
325 : "=&r" (s->index), "=&r" (value) |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
326 : "r" (s->buf), "r" (n), "0" (s->index), "1" (value) |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
327 : "%ecx" |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
328 ); |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
329 #else |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
330 int index= s->index; |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
331 uint32_t *ptr= (uint32_t*)(((uint8_t *)s->buf)+(index>>3)); |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
332 |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
333 ptr[0] |= be2me_32(value<<(32-n-(index&7) )); |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
334 ptr[1] = 0; |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
335 //if(n>24) printf("%d %d\n", n, value); |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
336 index+= n; |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
337 s->index= index; |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
338 #endif |
|
235
41f0ef2cd942
aligned bitstream writer (1% slower on p3 but perhaps its faster on p4?)
michaelni
parents:
234
diff
changeset
|
339 #endif //!ALIGNED_BITSTREAM_WRITER |
|
234
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
340 } |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
341 #endif |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
342 |
|
238
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
343 #ifndef ALT_BITSTREAM_WRITER |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
344 /* for jpeg : escape 0xff with 0x00 after it */ |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
345 static inline void jput_bits(PutBitContext *s, int n, unsigned int value) |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
346 { |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
347 unsigned int bit_buf, b; |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
348 int bit_left, i; |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
349 |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
350 assert(n == 32 || value < (1U << n)); |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
351 |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
352 bit_buf = s->bit_buf; |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
353 bit_left = s->bit_left; |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
354 |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
355 //printf("n=%d value=%x cnt=%d buf=%x\n", n, value, bit_cnt, bit_buf); |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
356 /* XXX: optimize */ |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
357 if (n < bit_left) { |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
358 bit_buf = (bit_buf<<n) | value; |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
359 bit_left-=n; |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
360 } else { |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
361 bit_buf<<=bit_left; |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
362 bit_buf |= value >> (n - bit_left); |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
363 /* handle escape */ |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
364 for(i=0;i<4;i++) { |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
365 b = (bit_buf >> 24); |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
366 *(s->buf_ptr++) = b; |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
367 if (b == 0xff) |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
368 *(s->buf_ptr++) = 0; |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
369 bit_buf <<= 8; |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
370 } |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
371 |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
372 bit_left+= 32 - n; |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
373 bit_buf = value; |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
374 } |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
375 |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
376 s->bit_buf = bit_buf; |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
377 s->bit_left = bit_left; |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
378 } |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
379 #endif |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
380 |
|
99a9f903f0e3
optimized the normal bitstream writer, its faster than the alternative one on p3 now ... lets hope its at least not slower on p4 & k7
michaelni
parents:
235
diff
changeset
|
381 |
|
234
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
382 #ifdef ALT_BITSTREAM_WRITER |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
383 static inline void jput_bits(PutBitContext *s, int n, int value) |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
384 { |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
385 int index= s->index; |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
386 uint32_t *ptr= (uint32_t*)(((uint8_t *)s->buf)+(index>>3)); |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
387 int v= ptr[0]; |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
388 //if(n>24) printf("%d %d\n", n, value); |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
389 |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
390 v |= be2me_32(value<<(32-n-(index&7) )); |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
391 if(((v+0x01010101)^0xFFFFFFFF)&v&0x80808080) |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
392 { |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
393 /* handle idiotic (m)jpeg escapes */ |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
394 uint8_t *bPtr= (uint8_t*)ptr; |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
395 int numChecked= ((index+n)>>3) - (index>>3); |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
396 |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
397 v= be2me_32(v); |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
398 |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
399 *(bPtr++)= v>>24; |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
400 if((v&0xFF000000)==0xFF000000 && numChecked>0){ |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
401 *(bPtr++)= 0x00; |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
402 index+=8; |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
403 } |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
404 *(bPtr++)= (v>>16)&0xFF; |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
405 if((v&0x00FF0000)==0x00FF0000 && numChecked>1){ |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
406 *(bPtr++)= 0x00; |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
407 index+=8; |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
408 } |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
409 *(bPtr++)= (v>>8)&0xFF; |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
410 if((v&0x0000FF00)==0x0000FF00 && numChecked>2){ |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
411 *(bPtr++)= 0x00; |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
412 index+=8; |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
413 } |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
414 *(bPtr++)= v&0xFF; |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
415 if((v&0x000000FF)==0x000000FF && numChecked>3){ |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
416 *(bPtr++)= 0x00; |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
417 index+=8; |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
418 } |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
419 *((uint32_t*)bPtr)= 0; |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
420 } |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
421 else |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
422 { |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
423 ptr[0] = v; |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
424 ptr[1] = 0; |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
425 } |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
426 |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
427 index+= n; |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
428 s->index= index; |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
429 } |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
430 #endif |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
431 |
|
5fc0c3af3fe4
alternative bitstream writer (disabled by default, uncomment #define ALT_BISTREAM_WRITER in common.h if u want to try it)
michaelni
parents:
213
diff
changeset
|
432 |
|
5fc0c3af3fe4
alternative 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 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
|
434 { |
|
5fc0c3af3fe4
alternative 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 #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
|
436 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
|
437 #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
|
438 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
|
439 #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
|
440 } |
|
5fc0c3af3fe4
alternative 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 |
| 0 | 442 void init_get_bits(GetBitContext *s, |
| 443 UINT8 *buffer, int buffer_size); | |
| 444 | |
|
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
445 #ifndef ALT_BITSTREAM_READER |
|
20
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
446 unsigned int get_bits_long(GetBitContext *s, int n); |
| 144 | 447 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
|
448 #endif |
|
20
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
449 |
|
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
450 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
|
451 #ifdef ALT_BITSTREAM_READER |
|
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
452 #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
|
453 int index= s->index; |
|
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
454 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
|
455 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
|
456 #ifdef ARCH_X86 |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
457 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
|
458 : "=r" (result1) |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
459 : "0" (result1), "r" (result2), "c" (index)); |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
460 #else |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
461 result1<<= (index&0x1F); |
|
199
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
462 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
|
463 result1|= result2; |
|
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
464 #endif |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
465 result1>>= 32 - n; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
466 index+= n; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
467 s->index= index; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
468 |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
469 return result1; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
470 #else //ALIGNED_BITSTREAM |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
471 int index= s->index; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
472 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
|
473 |
|
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
474 result<<= (index&0x07); |
|
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
475 result>>= 32 - n; |
|
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
476 index+= n; |
|
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
477 s->index= index; |
| 306 | 478 #ifdef DUMP_STREAM |
| 479 while(n){ | |
| 480 printf("%d", (result>>(n-1))&1); | |
| 481 n--; | |
| 482 } | |
| 483 printf(" "); | |
| 484 #endif | |
|
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
485 return result; |
|
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
486 #endif //!ALIGNED_BITSTREAM |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
487 #else //ALT_BITSTREAM_READER |
|
20
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
488 if(s->bit_cnt>=n){ |
|
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
489 /* most common case here */ |
|
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
490 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
|
491 s->bit_buf <<= n; |
|
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
492 s->bit_cnt -= n; |
|
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
493 #ifdef STATS |
|
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
494 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
|
495 #endif |
|
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
496 return val; |
|
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
497 } |
|
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
498 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
|
499 #endif //!ALT_BITSTREAM_READER |
|
20
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
500 } |
|
907b67420d84
inlineing common case of get_bits() -> gives 2speedup. more optim coming soon...
arpi_esp
parents:
10
diff
changeset
|
501 |
| 21 | 502 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
|
503 #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
|
504 int index= s->index; |
|
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
505 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
|
506 result<<= (index&0x07); |
|
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
507 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
|
508 index++; |
|
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
509 s->index= index; |
|
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
510 |
| 306 | 511 #ifdef DUMP_STREAM |
| 512 printf("%d ", result); | |
| 513 #endif | |
|
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
514 return result; |
|
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
515 #else |
| 21 | 516 if(s->bit_cnt>0){ |
| 517 /* most common case here */ | |
| 518 unsigned int val = s->bit_buf >> 31; | |
| 519 s->bit_buf <<= 1; | |
| 520 s->bit_cnt--; | |
| 521 #ifdef STATS | |
| 522 st_bit_counts[st_current_index]++; | |
| 523 #endif | |
| 524 return val; | |
| 525 } | |
| 526 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
|
527 #endif |
| 21 | 528 } |
| 529 | |
| 144 | 530 /* This function is identical to get_bits(), the only */ |
| 531 /* diference is that it doesn't touch the buffer */ | |
| 532 /* it is usefull to see the buffer. */ | |
| 533 static inline unsigned int show_bits(GetBitContext *s, int n) | |
| 534 { | |
|
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
535 #ifdef ALT_BITSTREAM_READER |
|
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
536 #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
|
537 int index= s->index; |
|
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
538 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
|
539 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
|
540 #ifdef ARCH_X86 |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
541 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
|
542 : "=r" (result1) |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
543 : "0" (result1), "r" (result2), "c" (index)); |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
544 #else |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
545 result1<<= (index&0x1F); |
|
199
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
546 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
|
547 result1|= result2; |
|
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
548 #endif |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
549 result1>>= 32 - n; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
550 |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
551 return result1; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
552 #else //ALIGNED_BITSTREAM |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
553 int index= s->index; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
554 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
|
555 |
|
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
556 result<<= (index&0x07); |
|
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
557 result>>= 32 - n; |
|
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
558 |
|
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
559 return result; |
|
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
560 #endif //!ALIGNED_BITSTREAM |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
561 #else //ALT_BITSTREAM_READER |
| 144 | 562 if(s->bit_cnt>=n) { |
| 563 /* most common case here */ | |
| 564 unsigned int val = s->bit_buf >> (32 - n); | |
| 565 return val; | |
| 566 } | |
| 567 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
|
568 #endif //!ALT_BITSTREAM_READER |
| 144 | 569 } |
| 570 | |
| 290 | 571 static inline int show_aligned_bits(GetBitContext *s, int offset, int n) |
| 572 { | |
| 573 #ifdef ALT_BITSTREAM_READER | |
| 574 #ifdef ALIGNED_BITSTREAM | |
| 575 int index= (s->index + offset + 7)&(~7); | |
| 576 uint32_t result1= be2me_32( ((uint32_t *)s->buffer)[index>>5] ); | |
| 577 uint32_t result2= be2me_32( ((uint32_t *)s->buffer)[(index>>5) + 1] ); | |
| 578 #ifdef ARCH_X86 | |
| 579 asm ("shldl %%cl, %2, %0\n\t" | |
| 580 : "=r" (result1) | |
| 581 : "0" (result1), "r" (result2), "c" (index)); | |
| 582 #else | |
| 583 result1<<= (index&0x1F); | |
| 584 result2= (result2>>1) >> (31-(index&0x1F)); | |
| 585 result1|= result2; | |
| 586 #endif | |
| 587 result1>>= 32 - n; | |
| 588 | |
| 589 return result1; | |
| 590 #else //ALIGNED_BITSTREAM | |
| 591 int index= (s->index + offset + 7)>>3; | |
| 592 uint32_t result= be2me_32( unaligned32( ((uint8_t *)s->buffer)+index ) ); | |
| 593 | |
| 594 result>>= 32 - n; | |
| 595 | |
| 596 return result; | |
| 597 #endif //!ALIGNED_BITSTREAM | |
| 598 #else //ALT_BITSTREAM_READER | |
| 599 int index= (get_bits_count(s) + offset + 7)>>3; | |
| 600 uint32_t result= be2me_32( unaligned32( ((uint8_t *)s->buf)+index ) ); | |
| 601 | |
| 602 result>>= 32 - n; | |
| 603 //printf(" %X %X %d \n", (int)(((uint8_t *)s->buf)+index ), (int)s->buf_ptr, s->bit_cnt); | |
| 604 return result; | |
| 605 #endif //!ALT_BITSTREAM_READER | |
| 606 } | |
| 607 | |
| 21 | 608 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
|
609 #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
|
610 s->index+= n; |
| 306 | 611 #ifdef DUMP_STREAM |
| 612 { | |
| 613 int result; | |
| 614 s->index-= n; | |
| 615 result= get_bits(s, n); | |
| 616 } | |
| 617 #endif | |
| 618 | |
|
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
619 #else |
| 21 | 620 if(s->bit_cnt>=n){ |
| 621 /* most common case here */ | |
| 622 s->bit_buf <<= n; | |
| 623 s->bit_cnt -= n; | |
| 624 #ifdef STATS | |
| 625 st_bit_counts[st_current_index] += n; | |
| 626 #endif | |
| 627 } else { | |
| 628 get_bits_long(s,n); | |
| 629 } | |
|
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
630 #endif |
| 21 | 631 } |
| 632 | |
| 633 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
|
634 #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
|
635 s->index++; |
| 306 | 636 #ifdef DUMP_STREAM |
| 637 s->index--; | |
| 638 printf("%d ", get_bits1(s)); | |
| 639 #endif | |
|
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
640 #else |
| 21 | 641 if(s->bit_cnt>0){ |
| 642 /* most common case here */ | |
| 643 s->bit_buf <<= 1; | |
| 644 s->bit_cnt--; | |
| 645 #ifdef STATS | |
| 646 st_bit_counts[st_current_index]++; | |
| 647 #endif | |
| 648 } else { | |
| 649 get_bits_long(s,1); | |
| 650 } | |
|
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
651 #endif |
| 21 | 652 } |
| 653 | |
| 85 | 654 static inline int get_bits_count(GetBitContext *s) |
| 655 { | |
|
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
656 #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
|
657 return s->index; |
|
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
658 #else |
| 85 | 659 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
|
660 #endif |
| 85 | 661 } |
| 21 | 662 |
|
264
28c5c62b1c4c
support decoding (with mplayer) of 3 .mp4 files from mphq
michaelni
parents:
238
diff
changeset
|
663 int check_marker(GetBitContext *s, char *msg); |
| 0 | 664 void align_get_bits(GetBitContext *s); |
| 665 int init_vlc(VLC *vlc, int nb_bits, int nb_codes, | |
| 666 const void *bits, int bits_wrap, int bits_size, | |
| 667 const void *codes, int codes_wrap, int codes_size); | |
| 668 void free_vlc(VLC *vlc); | |
| 669 | |
|
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
670 #ifdef ALT_BITSTREAM_READER |
|
199
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
671 #ifdef ALIGNED_BITSTREAM |
|
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
672 #ifdef ARCH_X86 |
|
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
673 #define SHOW_BITS(s, val, n) \ |
|
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
674 val= be2me_32( ((uint32_t *)(s)->buffer)[bit_cnt>>5] );\ |
|
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
675 {uint32_t result2= be2me_32( ((uint32_t *)(s)->buffer)[(bit_cnt>>5) + 1] );\ |
|
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
676 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
|
677 : "=r" (val)\ |
|
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
678 : "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
|
679 ((uint32_t)val)>>= 32 - n;} |
|
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
680 #else //ARCH_X86 |
|
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
681 #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
|
682 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
|
683 {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
|
684 val<<= (bit_cnt&0x1F);\ |
|
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
685 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
|
686 val|= result2;\ |
|
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
687 ((uint32_t)val)>>= 32 - n;} |
|
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
688 #endif //!ARCH_X86 |
|
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
689 #else //ALIGNED_BITSTREAM |
|
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
690 #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
|
691 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
|
692 val<<= (bit_cnt&0x07);\ |
|
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
693 ((uint32_t)val)>>= 32 - n; |
|
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
694 #endif // !ALIGNED_BITSTREAM |
|
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
695 #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
|
696 #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
|
697 #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
|
698 #else |
|
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
699 |
| 0 | 700 /* macro to go faster */ |
| 701 /* n must be <= 24 */ | |
| 702 /* XXX: optimize buffer end test */ | |
| 703 #define SHOW_BITS(s, val, n)\ | |
| 704 {\ | |
| 705 if (bit_cnt < n && buf_ptr < (s)->buf_end) {\ | |
| 706 bit_buf |= *buf_ptr++ << (24 - bit_cnt);\ | |
| 707 bit_cnt += 8;\ | |
| 708 if (bit_cnt < n && buf_ptr < (s)->buf_end) {\ | |
| 709 bit_buf |= *buf_ptr++ << (24 - bit_cnt);\ | |
| 710 bit_cnt += 8;\ | |
| 711 if (bit_cnt < n && buf_ptr < (s)->buf_end) {\ | |
| 712 bit_buf |= *buf_ptr++ << (24 - bit_cnt);\ | |
| 713 bit_cnt += 8;\ | |
| 714 }\ | |
| 715 }\ | |
| 716 }\ | |
| 717 val = bit_buf >> (32 - n);\ | |
| 718 } | |
| 719 | |
| 720 /* SHOW_BITS with n1 >= n must be been done before */ | |
| 721 #define FLUSH_BITS(n)\ | |
| 722 {\ | |
| 723 bit_buf <<= n;\ | |
| 724 bit_cnt -= n;\ | |
| 725 } | |
| 726 | |
| 727 #define SAVE_BITS(s) \ | |
| 728 {\ | |
| 729 bit_cnt = (s)->bit_cnt;\ | |
| 730 bit_buf = (s)->bit_buf;\ | |
| 731 buf_ptr = (s)->buf_ptr;\ | |
| 732 } | |
| 733 | |
| 734 #define RESTORE_BITS(s) \ | |
| 735 {\ | |
| 736 (s)->buf_ptr = buf_ptr;\ | |
| 737 (s)->bit_buf = bit_buf;\ | |
| 738 (s)->bit_cnt = bit_cnt;\ | |
| 739 } | |
|
192
1e5f64be86fc
another bitstream reader code (faster on intel cpus) - patch by Michael Niedermayer <michaelni@gmx.at>
uid46427
parents:
151
diff
changeset
|
740 #endif // !ALT_BITSTREAM_READER |
|
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
741 |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
742 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
|
743 { |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
744 int code, n, nb_bits, index; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
745 INT16 *table_codes; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
746 INT8 *table_bits; |
|
199
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
747 int bit_cnt; |
|
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
748 #ifndef ALT_BITSTREAM_READER |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
749 UINT32 bit_buf; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
750 UINT8 *buf_ptr; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
751 #endif |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
752 |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
753 SAVE_BITS(s); |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
754 nb_bits = vlc->bits; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
755 table_codes = vlc->table_codes; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
756 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
|
757 |
|
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
758 #ifdef FAST_GET_FIRST_VLC |
|
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
759 SHOW_BITS(s, index, nb_bits); |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
760 code = table_codes[index]; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
761 n = table_bits[index]; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
762 if (n > 0) { |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
763 /* most common case (90%)*/ |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
764 FLUSH_BITS(n); |
| 306 | 765 #ifdef DUMP_STREAM |
| 766 { | |
| 767 int n= bit_cnt - s->index; | |
| 768 skip_bits(s, n); | |
| 769 RESTORE_BITS(s); | |
| 770 } | |
| 771 #endif | |
|
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
772 RESTORE_BITS(s); |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
773 return code; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
774 } else if (n == 0) { |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
775 return -1; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
776 } else { |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
777 FLUSH_BITS(nb_bits); |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
778 nb_bits = -n; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
779 table_codes = vlc->table_codes + code; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
780 table_bits = vlc->table_bits + code; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
781 } |
|
199
0f1dba8fc617
(commited by michael / arpi was crazy enough to give me his password)
arpi_esp
parents:
193
diff
changeset
|
782 #endif |
|
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
783 for(;;) { |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
784 SHOW_BITS(s, index, nb_bits); |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
785 code = table_codes[index]; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
786 n = table_bits[index]; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
787 if (n > 0) { |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
788 /* most common case */ |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
789 FLUSH_BITS(n); |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
790 #ifdef STATS |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
791 st_bit_counts[st_current_index] += n; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
792 #endif |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
793 break; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
794 } else if (n == 0) { |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
795 return -1; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
796 } else { |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
797 FLUSH_BITS(nb_bits); |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
798 #ifdef STATS |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
799 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
|
800 #endif |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
801 nb_bits = -n; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
802 table_codes = vlc->table_codes + code; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
803 table_bits = vlc->table_bits + code; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
804 } |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
805 } |
| 306 | 806 #ifdef DUMP_STREAM |
| 807 { | |
| 808 int n= bit_cnt - s->index; | |
| 809 skip_bits(s, n); | |
| 810 RESTORE_BITS(s); | |
| 811 } | |
| 812 #endif | |
|
193
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
813 RESTORE_BITS(s); |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
814 return code; |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
815 } |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
816 |
|
b691dd3e9088
aligned bitstream support (optional) - patch by ichael Niedermayer <michaelni@gmx.at>
arpi_esp
parents:
192
diff
changeset
|
817 |
| 0 | 818 /* define it to include statistics code (useful only for optimizing |
| 819 codec efficiency */ | |
| 820 //#define STATS | |
| 821 | |
| 822 #ifdef STATS | |
| 823 | |
| 824 enum { | |
| 825 ST_UNKNOWN, | |
| 826 ST_DC, | |
| 827 ST_INTRA_AC, | |
| 828 ST_INTER_AC, | |
| 829 ST_INTRA_MB, | |
| 830 ST_INTER_MB, | |
| 831 ST_MV, | |
| 832 ST_NB, | |
| 833 }; | |
| 834 | |
| 835 extern int st_current_index; | |
| 836 extern unsigned int st_bit_counts[ST_NB]; | |
| 837 extern unsigned int st_out_bit_counts[ST_NB]; | |
| 838 | |
| 839 void print_stats(void); | |
| 840 #endif | |
| 841 | |
| 842 /* misc math functions */ | |
| 843 | |
|
151
ae0516eadae2
fixed gcc-3.0.x compilation (by Michael Niedermayer)
nickols_k
parents:
144
diff
changeset
|
844 static inline int av_log2(unsigned int v) |
| 0 | 845 { |
| 846 int n; | |
| 847 | |
| 848 n = 0; | |
| 849 if (v & 0xffff0000) { | |
| 850 v >>= 16; | |
| 851 n += 16; | |
| 852 } | |
| 853 if (v & 0xff00) { | |
| 854 v >>= 8; | |
| 855 n += 8; | |
| 856 } | |
| 857 if (v & 0xf0) { | |
| 858 v >>= 4; | |
| 859 n += 4; | |
| 860 } | |
| 861 if (v & 0xc) { | |
| 862 v >>= 2; | |
| 863 n += 2; | |
| 864 } | |
| 865 if (v & 0x2) { | |
| 866 n++; | |
| 867 } | |
| 868 return n; | |
| 869 } | |
| 870 | |
|
277
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
871 /* median of 3 */ |
|
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
872 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
|
873 { |
|
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
874 int vmin, vmax; |
|
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
875 vmax = vmin = a; |
|
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
876 if (b < vmin) |
|
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
877 vmin = b; |
|
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
878 else |
|
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
879 vmax = b; |
|
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
880 |
|
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
881 if (c < vmin) |
|
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
882 vmin = c; |
|
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
883 else if (c > vmax) |
|
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
884 vmax = c; |
|
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
885 |
|
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
886 return a + b + c - vmin - vmax; |
|
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
887 } |
|
5cb2978e701f
new motion estimation (epzs) not complete yet but allready pretty good :)
michaelni
parents:
264
diff
changeset
|
888 |
| 327 | 889 static inline int clip(int a, int amin, int amax) |
| 890 { | |
| 891 if (a < amin) | |
| 892 return amin; | |
| 893 else if (a > amax) | |
| 894 return amax; | |
| 895 else | |
| 896 return a; | |
| 897 } | |
| 898 | |
| 0 | 899 /* memory */ |
| 900 void *av_mallocz(int size); | |
| 901 | |
| 324 | 902 /* math */ |
| 903 int ff_gcd(int a, int b); | |
| 904 | |
|
370
0eca28d16cbd
clamp intra matrix to 8bit for mjpeg (workaround for qscale>=25)
al3x
parents:
359
diff
changeset
|
905 #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
|
906 |
| 0 | 907 #endif |
