Mercurial > libavcodec.hg
annotate huffyuv.c @ 9906:fe4e6d8242d4 libavcodec
Make sure we dont read over the end.
Fixes issue1237.
author | michael |
---|---|
date | Wed, 01 Jul 2009 20:30:37 +0000 |
parents | f1f5e78cf0aa |
children | 2fb2e212318b |
rev | line source |
---|---|
866 | 1 /* |
2 * huffyuv codec for libavcodec | |
3 * | |
1006 | 4 * Copyright (c) 2002-2003 Michael Niedermayer <michaelni@gmx.at> |
866 | 5 * |
5214 | 6 * see http://www.pcisys.net/~melanson/codecs/huffyuv.txt for a description of |
7 * the algorithm used | |
8 * | |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3777
diff
changeset
|
9 * This file is part of FFmpeg. |
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3777
diff
changeset
|
10 * |
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3777
diff
changeset
|
11 * FFmpeg is free software; you can redistribute it and/or |
866 | 12 * modify it under the terms of the GNU Lesser General Public |
13 * License as published by the Free Software Foundation; either | |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3777
diff
changeset
|
14 * version 2.1 of the License, or (at your option) any later version. |
866 | 15 * |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3777
diff
changeset
|
16 * FFmpeg is distributed in the hope that it will be useful, |
866 | 17 * but WITHOUT ANY WARRANTY; without even the implied warranty of |
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
19 * Lesser General Public License for more details. | |
20 * | |
21 * You should have received a copy of the GNU Lesser General Public | |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3777
diff
changeset
|
22 * License along with FFmpeg; if not, write to the Free Software |
3036
0b546eab515d
Update licensing information: The FSF changed postal address.
diego
parents:
2967
diff
changeset
|
23 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
866 | 24 */ |
2967 | 25 |
1106 | 26 /** |
8718
e9d9d946f213
Use full internal pathname in doxygen @file directives.
diego
parents:
8627
diff
changeset
|
27 * @file libavcodec/huffyuv.c |
1106 | 28 * huffyuv codec for libavcodec. |
29 */ | |
866 | 30 |
4962
f99e40a7155b
Remove redundant #inclusion of common.h, avcodec.h already #includes it.
diego
parents:
4691
diff
changeset
|
31 #include "avcodec.h" |
9428 | 32 #include "get_bits.h" |
9411
4cb7c65fc775
Split bitstream.h, put the bitstream writer stuff in the new file
stefano
parents:
9355
diff
changeset
|
33 #include "put_bits.h" |
866 | 34 #include "dsputil.h" |
35 | |
36 #define VLC_BITS 11 | |
903 | 37 |
2176 | 38 #ifdef WORDS_BIGENDIAN |
39 #define B 3 | |
40 #define G 2 | |
41 #define R 1 | |
42 #else | |
43 #define B 0 | |
44 #define G 1 | |
45 #define R 2 | |
46 #endif | |
47 | |
866 | 48 typedef enum Predictor{ |
49 LEFT= 0, | |
50 PLANE, | |
51 MEDIAN, | |
52 } Predictor; | |
2967 | 53 |
866 | 54 typedef struct HYuvContext{ |
55 AVCodecContext *avctx; | |
56 Predictor predictor; | |
57 GetBitContext gb; | |
58 PutBitContext pb; | |
59 int interlaced; | |
60 int decorrelate; | |
61 int bitstream_bpp; | |
62 int version; | |
63 int yuy2; //use yuy2 instead of 422P | |
64 int bgr32; //use bgr32 instead of bgr24 | |
65 int width, height; | |
66 int flags; | |
2369 | 67 int context; |
866 | 68 int picture_number; |
871 | 69 int last_slice_end; |
2422 | 70 uint8_t *temp[3]; |
866 | 71 uint64_t stats[3][256]; |
72 uint8_t len[3][256]; | |
73 uint32_t bits[3][256]; | |
5074 | 74 uint32_t pix_bgr_map[1<<VLC_BITS]; |
5063
d5640ea6d4a6
merge huffman tables so that we read 2 symbols at a time. 30% faster huffyuv decoding.
lorenm
parents:
5040
diff
changeset
|
75 VLC vlc[6]; //Y,U,V,YY,YU,YV |
925 | 76 AVFrame picture; |
2422 | 77 uint8_t *bitstream_buffer; |
3066
04b924f8f5a5
warning fixes by Luca Abeni, lucabe72 ##@## email ##.## it
diego
parents:
3036
diff
changeset
|
78 unsigned int bitstream_buffer_size; |
2967 | 79 DSPContext dsp; |
866 | 80 }HYuvContext; |
81 | |
1082 | 82 static const unsigned char classic_shift_luma[] = { |
1080
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
83 34,36,35,69,135,232,9,16,10,24,11,23,12,16,13,10,14,8,15,8, |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
84 16,8,17,20,16,10,207,206,205,236,11,8,10,21,9,23,8,8,199,70, |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
85 69,68, 0 |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
86 }; |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
87 |
1082 | 88 static const unsigned char classic_shift_chroma[] = { |
1080
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
89 66,36,37,38,39,40,41,75,76,77,110,239,144,81,82,83,84,85,118,183, |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
90 56,57,88,89,56,89,154,57,58,57,26,141,57,56,58,57,58,57,184,119, |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
91 214,245,116,83,82,49,80,79,78,77,44,75,41,40,39,38,37,36,34, 0 |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
92 }; |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
93 |
1082 | 94 static const unsigned char classic_add_luma[256] = { |
1080
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
95 3, 9, 5, 12, 10, 35, 32, 29, 27, 50, 48, 45, 44, 41, 39, 37, |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
96 73, 70, 68, 65, 64, 61, 58, 56, 53, 50, 49, 46, 44, 41, 38, 36, |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
97 68, 65, 63, 61, 58, 55, 53, 51, 48, 46, 45, 43, 41, 39, 38, 36, |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
98 35, 33, 32, 30, 29, 27, 26, 25, 48, 47, 46, 44, 43, 41, 40, 39, |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
99 37, 36, 35, 34, 32, 31, 30, 28, 27, 26, 24, 23, 22, 20, 19, 37, |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
100 35, 34, 33, 31, 30, 29, 27, 26, 24, 23, 21, 20, 18, 17, 15, 29, |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
101 27, 26, 24, 22, 21, 19, 17, 16, 14, 26, 25, 23, 21, 19, 18, 16, |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
102 15, 27, 25, 23, 21, 19, 17, 16, 14, 26, 25, 23, 21, 18, 17, 14, |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
103 12, 17, 19, 13, 4, 9, 2, 11, 1, 7, 8, 0, 16, 3, 14, 6, |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
104 12, 10, 5, 15, 18, 11, 10, 13, 15, 16, 19, 20, 22, 24, 27, 15, |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
105 18, 20, 22, 24, 26, 14, 17, 20, 22, 24, 27, 15, 18, 20, 23, 25, |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
106 28, 16, 19, 22, 25, 28, 32, 36, 21, 25, 29, 33, 38, 42, 45, 49, |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
107 28, 31, 34, 37, 40, 42, 44, 47, 49, 50, 52, 54, 56, 57, 59, 60, |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
108 62, 64, 66, 67, 69, 35, 37, 39, 40, 42, 43, 45, 47, 48, 51, 52, |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
109 54, 55, 57, 59, 60, 62, 63, 66, 67, 69, 71, 72, 38, 40, 42, 43, |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
110 46, 47, 49, 51, 26, 28, 30, 31, 33, 34, 18, 19, 11, 13, 7, 8, |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
111 }; |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
112 |
1082 | 113 static const unsigned char classic_add_chroma[256] = { |
1080
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
114 3, 1, 2, 2, 2, 2, 3, 3, 7, 5, 7, 5, 8, 6, 11, 9, |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
115 7, 13, 11, 10, 9, 8, 7, 5, 9, 7, 6, 4, 7, 5, 8, 7, |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
116 11, 8, 13, 11, 19, 15, 22, 23, 20, 33, 32, 28, 27, 29, 51, 77, |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
117 43, 45, 76, 81, 46, 82, 75, 55, 56,144, 58, 80, 60, 74,147, 63, |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
118 143, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
119 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 27, 30, 21, 22, |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
120 17, 14, 5, 6,100, 54, 47, 50, 51, 53,106,107,108,109,110,111, |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
121 112,113,114,115, 4,117,118, 92, 94,121,122, 3,124,103, 2, 1, |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
122 0,129,130,131,120,119,126,125,136,137,138,139,140,141,142,134, |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
123 135,132,133,104, 64,101, 62, 57,102, 95, 93, 59, 61, 28, 97, 96, |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
124 52, 49, 48, 29, 32, 25, 24, 46, 23, 98, 45, 44, 43, 20, 42, 41, |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
125 19, 18, 99, 40, 15, 39, 38, 16, 13, 12, 11, 37, 10, 9, 8, 36, |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
126 7,128,127,105,123,116, 35, 34, 33,145, 31, 79, 42,146, 78, 26, |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
127 83, 48, 49, 50, 44, 47, 26, 31, 30, 18, 17, 19, 21, 24, 25, 13, |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
128 14, 16, 17, 18, 20, 21, 12, 14, 15, 9, 10, 6, 9, 6, 5, 8, |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
129 6, 12, 8, 10, 7, 9, 6, 4, 6, 2, 2, 3, 3, 3, 3, 2, |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
130 }; |
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
131 |
866 | 132 static inline int add_left_prediction(uint8_t *dst, uint8_t *src, int w, int acc){ |
133 int i; | |
134 | |
135 for(i=0; i<w-1; i++){ | |
136 acc+= src[i]; | |
137 dst[i]= acc; | |
138 i++; | |
139 acc+= src[i]; | |
140 dst[i]= acc; | |
141 } | |
142 | |
143 for(; i<w; i++){ | |
144 acc+= src[i]; | |
145 dst[i]= acc; | |
146 } | |
147 | |
148 return acc; | |
149 } | |
150 | |
151 static inline void add_left_prediction_bgr32(uint8_t *dst, uint8_t *src, int w, int *red, int *green, int *blue){ | |
152 int i; | |
153 int r,g,b; | |
154 r= *red; | |
155 g= *green; | |
156 b= *blue; | |
157 | |
158 for(i=0; i<w; i++){ | |
2176 | 159 b+= src[4*i+B]; |
160 g+= src[4*i+G]; | |
161 r+= src[4*i+R]; | |
2967 | 162 |
2176 | 163 dst[4*i+B]= b; |
164 dst[4*i+G]= g; | |
165 dst[4*i+R]= r; | |
866 | 166 } |
167 | |
168 *red= r; | |
169 *green= g; | |
170 *blue= b; | |
171 } | |
172 | |
871 | 173 static inline int sub_left_prediction(HYuvContext *s, uint8_t *dst, uint8_t *src, int w, int left){ |
866 | 174 int i; |
871 | 175 if(w<32){ |
176 for(i=0; i<w; i++){ | |
177 const int temp= src[i]; | |
178 dst[i]= temp - left; | |
179 left= temp; | |
180 } | |
181 return left; | |
182 }else{ | |
183 for(i=0; i<16; i++){ | |
184 const int temp= src[i]; | |
185 dst[i]= temp - left; | |
186 left= temp; | |
187 } | |
188 s->dsp.diff_bytes(dst+16, src+16, src+15, w-16); | |
189 return src[w-1]; | |
866 | 190 } |
191 } | |
1325 | 192 |
4682 | 193 static inline void sub_left_prediction_bgr32(HYuvContext *s, uint8_t *dst, uint8_t *src, int w, int *red, int *green, int *blue){ |
194 int i; | |
195 int r,g,b; | |
196 r= *red; | |
197 g= *green; | |
198 b= *blue; | |
199 for(i=0; i<FFMIN(w,4); i++){ | |
200 const int rt= src[i*4+R]; | |
201 const int gt= src[i*4+G]; | |
202 const int bt= src[i*4+B]; | |
203 dst[i*4+R]= rt - r; | |
204 dst[i*4+G]= gt - g; | |
205 dst[i*4+B]= bt - b; | |
206 r = rt; | |
207 g = gt; | |
208 b = bt; | |
209 } | |
210 s->dsp.diff_bytes(dst+16, src+16, src+12, w*4-16); | |
211 *red= src[(w-1)*4+R]; | |
212 *green= src[(w-1)*4+G]; | |
213 *blue= src[(w-1)*4+B]; | |
214 } | |
215 | |
866 | 216 static void read_len_table(uint8_t *dst, GetBitContext *gb){ |
217 int i, val, repeat; | |
2967 | 218 |
866 | 219 for(i=0; i<256;){ |
220 repeat= get_bits(gb, 3); | |
221 val = get_bits(gb, 5); | |
222 if(repeat==0) | |
223 repeat= get_bits(gb, 8); | |
224 //printf("%d %d\n", val, repeat); | |
225 while (repeat--) | |
226 dst[i++] = val; | |
227 } | |
228 } | |
229 | |
230 static int generate_bits_table(uint32_t *dst, uint8_t *len_table){ | |
231 int len, index; | |
232 uint32_t bits=0; | |
233 | |
234 for(len=32; len>0; len--){ | |
235 for(index=0; index<256; index++){ | |
1279 | 236 if(len_table[index]==len) |
237 dst[index]= bits++; | |
866 | 238 } |
1279 | 239 if(bits & 1){ |
1598
932d306bf1dc
av_log() patch by (Michel Bardiaux <mbardiaux at peaktime dot be>)
michael
parents:
1529
diff
changeset
|
240 av_log(NULL, AV_LOG_ERROR, "Error generating huffman table\n"); |
1279 | 241 return -1; |
242 } | |
243 bits >>= 1; | |
866 | 244 } |
245 return 0; | |
246 } | |
247 | |
8590 | 248 #if CONFIG_HUFFYUV_ENCODER || CONFIG_FFVHUFF_ENCODER |
5040
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
249 typedef struct { |
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
250 uint64_t val; |
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
251 int name; |
8308 | 252 } HeapElem; |
5040
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
253 |
8308 | 254 static void heap_sift(HeapElem *h, int root, int size) |
5040
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
255 { |
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
256 while(root*2+1 < size) { |
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
257 int child = root*2+1; |
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
258 if(child < size-1 && h[child].val > h[child+1].val) |
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
259 child++; |
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
260 if(h[root].val > h[child].val) { |
8308 | 261 FFSWAP(HeapElem, h[root], h[child]); |
5040
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
262 root = child; |
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
263 } else |
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
264 break; |
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
265 } |
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
266 } |
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
267 |
866 | 268 static void generate_len_table(uint8_t *dst, uint64_t *stats, int size){ |
8308 | 269 HeapElem h[size]; |
866 | 270 int up[2*size]; |
5040
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
271 int len[2*size]; |
866 | 272 int offset, i, next; |
2967 | 273 |
866 | 274 for(offset=1; ; offset<<=1){ |
275 for(i=0; i<size; i++){ | |
5040
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
276 h[i].name = i; |
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
277 h[i].val = (stats[i] << 8) + offset; |
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
278 } |
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
279 for(i=size/2-1; i>=0; i--) |
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
280 heap_sift(h, i, size); |
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
281 |
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
282 for(next=size; next<size*2-1; next++){ |
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
283 // merge the two smallest entries, and put it back in the heap |
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
284 uint64_t min1v = h[0].val; |
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
285 up[h[0].name] = next; |
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
286 h[0].val = INT64_MAX; |
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
287 heap_sift(h, 0, size); |
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
288 up[h[0].name] = next; |
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
289 h[0].name = next; |
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
290 h[0].val += min1v; |
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
291 heap_sift(h, 0, size); |
866 | 292 } |
2967 | 293 |
5040
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
294 len[2*size-2] = 0; |
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
295 for(i=2*size-3; i>=size; i--) |
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
296 len[i] = len[up[i]] + 1; |
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
297 for(i=0; i<size; i++) { |
5c6cd6601371
change brute force search to min-heap. 3.6x faster generate_len_table, 8% faster ffvhuff encoding.
lorenm
parents:
5026
diff
changeset
|
298 dst[i] = len[up[i]] + 1; |
5481
defae3a747d9
prevent huffyuv from generating codewords of length 32. (regression in r9069)
lorenm
parents:
5221
diff
changeset
|
299 if(dst[i] >= 32) break; |
866 | 300 } |
301 if(i==size) break; | |
302 } | |
303 } | |
8590 | 304 #endif /* CONFIG_HUFFYUV_ENCODER || CONFIG_FFVHUFF_ENCODER */ |
866 | 305 |
5063
d5640ea6d4a6
merge huffman tables so that we read 2 symbols at a time. 30% faster huffyuv decoding.
lorenm
parents:
5040
diff
changeset
|
306 static void generate_joint_tables(HYuvContext *s){ |
5074 | 307 uint16_t symbols[1<<VLC_BITS]; |
308 uint16_t bits[1<<VLC_BITS]; | |
309 uint8_t len[1<<VLC_BITS]; | |
5063
d5640ea6d4a6
merge huffman tables so that we read 2 symbols at a time. 30% faster huffyuv decoding.
lorenm
parents:
5040
diff
changeset
|
310 if(s->bitstream_bpp < 24){ |
d5640ea6d4a6
merge huffman tables so that we read 2 symbols at a time. 30% faster huffyuv decoding.
lorenm
parents:
5040
diff
changeset
|
311 int p, i, y, u; |
d5640ea6d4a6
merge huffman tables so that we read 2 symbols at a time. 30% faster huffyuv decoding.
lorenm
parents:
5040
diff
changeset
|
312 for(p=0; p<3; p++){ |
d5640ea6d4a6
merge huffman tables so that we read 2 symbols at a time. 30% faster huffyuv decoding.
lorenm
parents:
5040
diff
changeset
|
313 for(i=y=0; y<256; y++){ |
d5640ea6d4a6
merge huffman tables so that we read 2 symbols at a time. 30% faster huffyuv decoding.
lorenm
parents:
5040
diff
changeset
|
314 int len0 = s->len[0][y]; |
d5640ea6d4a6
merge huffman tables so that we read 2 symbols at a time. 30% faster huffyuv decoding.
lorenm
parents:
5040
diff
changeset
|
315 int limit = VLC_BITS - len0; |
5073 | 316 if(limit <= 0) |
317 continue; | |
318 for(u=0; u<256; u++){ | |
319 int len1 = s->len[p][u]; | |
320 if(len1 > limit) | |
321 continue; | |
322 len[i] = len0 + len1; | |
323 bits[i] = (s->bits[0][y] << len1) + s->bits[p][u]; | |
324 symbols[i] = (y<<8) + u; | |
325 if(symbols[i] != 0xffff) // reserved to mean "invalid" | |
326 i++; | |
5063
d5640ea6d4a6
merge huffman tables so that we read 2 symbols at a time. 30% faster huffyuv decoding.
lorenm
parents:
5040
diff
changeset
|
327 } |
d5640ea6d4a6
merge huffman tables so that we read 2 symbols at a time. 30% faster huffyuv decoding.
lorenm
parents:
5040
diff
changeset
|
328 } |
d5640ea6d4a6
merge huffman tables so that we read 2 symbols at a time. 30% faster huffyuv decoding.
lorenm
parents:
5040
diff
changeset
|
329 free_vlc(&s->vlc[3+p]); |
5072
1aec7fab94c4
use sparse huffman tables. 1.5% faster huffyuv decoding.
lorenm
parents:
5063
diff
changeset
|
330 init_vlc_sparse(&s->vlc[3+p], VLC_BITS, i, len, 1, 1, bits, 2, 2, symbols, 2, 2, 0); |
5063
d5640ea6d4a6
merge huffman tables so that we read 2 symbols at a time. 30% faster huffyuv decoding.
lorenm
parents:
5040
diff
changeset
|
331 } |
5074 | 332 }else{ |
333 uint8_t (*map)[4] = (uint8_t(*)[4])s->pix_bgr_map; | |
334 int i, b, g, r, code; | |
335 int p0 = s->decorrelate; | |
336 int p1 = !s->decorrelate; | |
337 // restrict the range to +/-16 becaues that's pretty much guaranteed to | |
338 // cover all the combinations that fit in 11 bits total, and it doesn't | |
339 // matter if we miss a few rare codes. | |
340 for(i=0, g=-16; g<16; g++){ | |
341 int len0 = s->len[p0][g&255]; | |
342 int limit0 = VLC_BITS - len0; | |
343 if(limit0 < 2) | |
344 continue; | |
345 for(b=-16; b<16; b++){ | |
346 int len1 = s->len[p1][b&255]; | |
347 int limit1 = limit0 - len1; | |
348 if(limit1 < 1) | |
349 continue; | |
350 code = (s->bits[p0][g&255] << len1) + s->bits[p1][b&255]; | |
351 for(r=-16; r<16; r++){ | |
352 int len2 = s->len[2][r&255]; | |
353 if(len2 > limit1) | |
354 continue; | |
355 len[i] = len0 + len1 + len2; | |
356 bits[i] = (code << len2) + s->bits[2][r&255]; | |
357 if(s->decorrelate){ | |
358 map[i][G] = g; | |
359 map[i][B] = g+b; | |
360 map[i][R] = g+r; | |
361 }else{ | |
362 map[i][B] = g; | |
363 map[i][G] = b; | |
364 map[i][R] = r; | |
365 } | |
366 i++; | |
367 } | |
368 } | |
369 } | |
370 free_vlc(&s->vlc[3]); | |
371 init_vlc(&s->vlc[3], VLC_BITS, i, len, 1, 1, bits, 2, 2, 0); | |
5063
d5640ea6d4a6
merge huffman tables so that we read 2 symbols at a time. 30% faster huffyuv decoding.
lorenm
parents:
5040
diff
changeset
|
372 } |
d5640ea6d4a6
merge huffman tables so that we read 2 symbols at a time. 30% faster huffyuv decoding.
lorenm
parents:
5040
diff
changeset
|
373 } |
d5640ea6d4a6
merge huffman tables so that we read 2 symbols at a time. 30% faster huffyuv decoding.
lorenm
parents:
5040
diff
changeset
|
374 |
866 | 375 static int read_huffman_tables(HYuvContext *s, uint8_t *src, int length){ |
376 GetBitContext gb; | |
377 int i; | |
2967 | 378 |
1025
1f9afd8b9131
GetBitContext.size is allways multiplied by 8 -> use size_in_bits to avoid useless *8 in a few inner loops
michaelni
parents:
1006
diff
changeset
|
379 init_get_bits(&gb, src, length*8); |
2967 | 380 |
866 | 381 for(i=0; i<3; i++){ |
382 read_len_table(s->len[i], &gb); | |
2967 | 383 |
866 | 384 if(generate_bits_table(s->bits[i], s->len[i])<0){ |
385 return -1; | |
386 } | |
387 #if 0 | |
388 for(j=0; j<256; j++){ | |
389 printf("%6X, %2d, %3d\n", s->bits[i][j], s->len[i][j], j); | |
390 } | |
391 #endif | |
2369 | 392 free_vlc(&s->vlc[i]); |
2370
26560d4fdb1f
Memory leak fix patch by (Burkhard Plaum <plaum >at< ipf.uni-stuttgart )dot( de>)
michael
parents:
2369
diff
changeset
|
393 init_vlc(&s->vlc[i], VLC_BITS, 256, s->len[i], 1, 1, s->bits[i], 4, 4, 0); |
866 | 394 } |
2967 | 395 |
5063
d5640ea6d4a6
merge huffman tables so that we read 2 symbols at a time. 30% faster huffyuv decoding.
lorenm
parents:
5040
diff
changeset
|
396 generate_joint_tables(s); |
d5640ea6d4a6
merge huffman tables so that we read 2 symbols at a time. 30% faster huffyuv decoding.
lorenm
parents:
5040
diff
changeset
|
397 |
2369 | 398 return (get_bits_count(&gb)+7)/8; |
866 | 399 } |
400 | |
401 static int read_old_huffman_tables(HYuvContext *s){ | |
1080
a150aba978de
huffyuv v1 tables, as they are essential and the only possible way for decding of v1 files they very likely cant be copyrighted ...
michaelni
parents:
1078
diff
changeset
|
402 #if 1 |
866 | 403 GetBitContext gb; |
404 int i; | |
405 | |
1025
1f9afd8b9131
GetBitContext.size is allways multiplied by 8 -> use size_in_bits to avoid useless *8 in a few inner loops
michaelni
parents:
1006
diff
changeset
|
406 init_get_bits(&gb, classic_shift_luma, sizeof(classic_shift_luma)*8); |
866 | 407 read_len_table(s->len[0], &gb); |
1025
1f9afd8b9131
GetBitContext.size is allways multiplied by 8 -> use size_in_bits to avoid useless *8 in a few inner loops
michaelni
parents:
1006
diff
changeset
|
408 init_get_bits(&gb, classic_shift_chroma, sizeof(classic_shift_chroma)*8); |
866 | 409 read_len_table(s->len[1], &gb); |
2967 | 410 |
866 | 411 for(i=0; i<256; i++) s->bits[0][i] = classic_add_luma [i]; |
412 for(i=0; i<256; i++) s->bits[1][i] = classic_add_chroma[i]; | |
413 | |
414 if(s->bitstream_bpp >= 24){ | |
415 memcpy(s->bits[1], s->bits[0], 256*sizeof(uint32_t)); | |
416 memcpy(s->len[1] , s->len [0], 256*sizeof(uint8_t)); | |
417 } | |
418 memcpy(s->bits[2], s->bits[1], 256*sizeof(uint32_t)); | |
419 memcpy(s->len[2] , s->len [1], 256*sizeof(uint8_t)); | |
2967 | 420 |
2369 | 421 for(i=0; i<3; i++){ |
422 free_vlc(&s->vlc[i]); | |
2370
26560d4fdb1f
Memory leak fix patch by (Burkhard Plaum <plaum >at< ipf.uni-stuttgart )dot( de>)
michael
parents:
2369
diff
changeset
|
423 init_vlc(&s->vlc[i], VLC_BITS, 256, s->len[i], 1, 1, s->bits[i], 4, 4, 0); |
2369 | 424 } |
2967 | 425 |
5063
d5640ea6d4a6
merge huffman tables so that we read 2 symbols at a time. 30% faster huffyuv decoding.
lorenm
parents:
5040
diff
changeset
|
426 generate_joint_tables(s); |
d5640ea6d4a6
merge huffman tables so that we read 2 symbols at a time. 30% faster huffyuv decoding.
lorenm
parents:
5040
diff
changeset
|
427 |
866 | 428 return 0; |
429 #else | |
3177 | 430 av_log(s->avctx, AV_LOG_DEBUG, "v1 huffyuv is not supported \n"); |
866 | 431 return -1; |
432 #endif | |
433 } | |
434 | |
9007
043574c5c153
Add missing av_cold in static init/close functions.
stefano
parents:
8760
diff
changeset
|
435 static av_cold void alloc_temp(HYuvContext *s){ |
2511 | 436 int i; |
2967 | 437 |
2511 | 438 if(s->bitstream_bpp<24){ |
439 for(i=0; i<3; i++){ | |
440 s->temp[i]= av_malloc(s->width + 16); | |
441 } | |
442 }else{ | |
4682 | 443 for(i=0; i<2; i++){ |
444 s->temp[i]= av_malloc(4*s->width + 16); | |
445 } | |
2511 | 446 } |
447 } | |
448 | |
9007
043574c5c153
Add missing av_cold in static init/close functions.
stefano
parents:
8760
diff
changeset
|
449 static av_cold int common_init(AVCodecContext *avctx){ |
866 | 450 HYuvContext *s = avctx->priv_data; |
451 | |
452 s->avctx= avctx; | |
453 s->flags= avctx->flags; | |
2967 | 454 |
1097 | 455 dsputil_init(&s->dsp, avctx); |
2967 | 456 |
2422 | 457 s->width= avctx->width; |
458 s->height= avctx->height; | |
459 assert(s->width>0 && s->height>0); | |
2967 | 460 |
2422 | 461 return 0; |
462 } | |
463 | |
8590 | 464 #if CONFIG_HUFFYUV_DECODER || CONFIG_FFVHUFF_DECODER |
6517
48759bfbd073
Apply 'cold' attribute to init/uninit functions in libavcodec
zuxy
parents:
6234
diff
changeset
|
465 static av_cold int decode_init(AVCodecContext *avctx) |
2422 | 466 { |
467 HYuvContext *s = avctx->priv_data; | |
468 | |
469 common_init(avctx); | |
2369 | 470 memset(s->vlc, 0, 3*sizeof(VLC)); |
2967 | 471 |
925 | 472 avctx->coded_frame= &s->picture; |
2422 | 473 s->interlaced= s->height > 288; |
903 | 474 |
866 | 475 s->bgr32=1; |
476 //if(avctx->extradata) | |
477 // printf("extradata:%X, extradata_size:%d\n", *(uint32_t*)avctx->extradata, avctx->extradata_size); | |
478 if(avctx->extradata_size){ | |
7823
4525dcd81357
Bump Major version, this commit is almost just renaming bits_per_sample to
michael
parents:
7782
diff
changeset
|
479 if((avctx->bits_per_coded_sample&7) && avctx->bits_per_coded_sample != 12) |
866 | 480 s->version=1; // do such files exist at all? |
481 else | |
482 s->version=2; | |
483 }else | |
484 s->version=0; | |
2967 | 485 |
866 | 486 if(s->version==2){ |
2374 | 487 int method, interlace; |
866 | 488 |
489 method= ((uint8_t*)avctx->extradata)[0]; | |
490 s->decorrelate= method&64 ? 1 : 0; | |
491 s->predictor= method&63; | |
492 s->bitstream_bpp= ((uint8_t*)avctx->extradata)[1]; | |
2967 | 493 if(s->bitstream_bpp==0) |
7823
4525dcd81357
Bump Major version, this commit is almost just renaming bits_per_sample to
michael
parents:
7782
diff
changeset
|
494 s->bitstream_bpp= avctx->bits_per_coded_sample&~7; |
2374 | 495 interlace= (((uint8_t*)avctx->extradata)[2] & 0x30) >> 4; |
496 s->interlaced= (interlace==1) ? 1 : (interlace==2) ? 0 : s->interlaced; | |
2369 | 497 s->context= ((uint8_t*)avctx->extradata)[2] & 0x40 ? 1 : 0; |
2967 | 498 |
866 | 499 if(read_huffman_tables(s, ((uint8_t*)avctx->extradata)+4, avctx->extradata_size) < 0) |
500 return -1; | |
501 }else{ | |
7823
4525dcd81357
Bump Major version, this commit is almost just renaming bits_per_sample to
michael
parents:
7782
diff
changeset
|
502 switch(avctx->bits_per_coded_sample&7){ |
866 | 503 case 1: |
504 s->predictor= LEFT; | |
505 s->decorrelate= 0; | |
506 break; | |
507 case 2: | |
508 s->predictor= LEFT; | |
509 s->decorrelate= 1; | |
510 break; | |
511 case 3: | |
512 s->predictor= PLANE; | |
7823
4525dcd81357
Bump Major version, this commit is almost just renaming bits_per_sample to
michael
parents:
7782
diff
changeset
|
513 s->decorrelate= avctx->bits_per_coded_sample >= 24; |
866 | 514 break; |
515 case 4: | |
516 s->predictor= MEDIAN; | |
517 s->decorrelate= 0; | |
518 break; | |
519 default: | |
520 s->predictor= LEFT; //OLD | |
521 s->decorrelate= 0; | |
522 break; | |
523 } | |
7823
4525dcd81357
Bump Major version, this commit is almost just renaming bits_per_sample to
michael
parents:
7782
diff
changeset
|
524 s->bitstream_bpp= avctx->bits_per_coded_sample & ~7; |
2369 | 525 s->context= 0; |
2967 | 526 |
866 | 527 if(read_old_huffman_tables(s) < 0) |
528 return -1; | |
529 } | |
2967 | 530 |
866 | 531 switch(s->bitstream_bpp){ |
532 case 12: | |
533 avctx->pix_fmt = PIX_FMT_YUV420P; | |
534 break; | |
535 case 16: | |
536 if(s->yuy2){ | |
4494
ce643a22f049
Replace deprecated PIX_FMT names by the newer variants.
diego
parents:
3947
diff
changeset
|
537 avctx->pix_fmt = PIX_FMT_YUYV422; |
866 | 538 }else{ |
539 avctx->pix_fmt = PIX_FMT_YUV422P; | |
540 } | |
541 break; | |
542 case 24: | |
543 case 32: | |
544 if(s->bgr32){ | |
4494
ce643a22f049
Replace deprecated PIX_FMT names by the newer variants.
diego
parents:
3947
diff
changeset
|
545 avctx->pix_fmt = PIX_FMT_RGB32; |
866 | 546 }else{ |
547 avctx->pix_fmt = PIX_FMT_BGR24; | |
548 } | |
549 break; | |
550 default: | |
551 assert(0); | |
552 } | |
2967 | 553 |
2511 | 554 alloc_temp(s); |
2967 | 555 |
7823
4525dcd81357
Bump Major version, this commit is almost just renaming bits_per_sample to
michael
parents:
7782
diff
changeset
|
556 // av_log(NULL, AV_LOG_DEBUG, "pred:%d bpp:%d hbpp:%d il:%d\n", s->predictor, s->bitstream_bpp, avctx->bits_per_coded_sample, s->interlaced); |
2190 | 557 |
866 | 558 return 0; |
559 } | |
8590 | 560 #endif /* CONFIG_HUFFYUV_DECODER || CONFIG_FFVHUFF_DECODER */ |
866 | 561 |
8590 | 562 #if CONFIG_HUFFYUV_ENCODER || CONFIG_FFVHUFF_ENCODER |
2369 | 563 static int store_table(HYuvContext *s, uint8_t *len, uint8_t *buf){ |
866 | 564 int i; |
2369 | 565 int index= 0; |
866 | 566 |
567 for(i=0; i<256;){ | |
568 int val= len[i]; | |
1529
cb523a2ca00f
fix the case where all vlc codes are 8 bits long (repeat=256)
michael
parents:
1528
diff
changeset
|
569 int repeat=0; |
2967 | 570 |
1529
cb523a2ca00f
fix the case where all vlc codes are 8 bits long (repeat=256)
michael
parents:
1528
diff
changeset
|
571 for(; i<256 && len[i]==val && repeat<255; i++) |
cb523a2ca00f
fix the case where all vlc codes are 8 bits long (repeat=256)
michael
parents:
1528
diff
changeset
|
572 repeat++; |
2967 | 573 |
1529
cb523a2ca00f
fix the case where all vlc codes are 8 bits long (repeat=256)
michael
parents:
1528
diff
changeset
|
574 assert(val < 32 && val >0 && repeat<256 && repeat>0); |
866 | 575 if(repeat>7){ |
2369 | 576 buf[index++]= val; |
577 buf[index++]= repeat; | |
866 | 578 }else{ |
2369 | 579 buf[index++]= val | (repeat<<5); |
866 | 580 } |
581 } | |
2967 | 582 |
2369 | 583 return index; |
866 | 584 } |
585 | |
6517
48759bfbd073
Apply 'cold' attribute to init/uninit functions in libavcodec
zuxy
parents:
6234
diff
changeset
|
586 static av_cold int encode_init(AVCodecContext *avctx) |
866 | 587 { |
588 HYuvContext *s = avctx->priv_data; | |
2422 | 589 int i, j; |
866 | 590 |
2422 | 591 common_init(avctx); |
2967 | 592 |
2422 | 593 avctx->extradata= av_mallocz(1024*30); // 256*3+4 == 772 |
594 avctx->stats_out= av_mallocz(1024*30); // 21*256*3(%llu ) + 3(\n) + 1(0) = 16132 | |
866 | 595 s->version=2; |
2967 | 596 |
925 | 597 avctx->coded_frame= &s->picture; |
2967 | 598 |
866 | 599 switch(avctx->pix_fmt){ |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
600 case PIX_FMT_YUV420P: |
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
601 s->bitstream_bpp= 12; |
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
602 break; |
866 | 603 case PIX_FMT_YUV422P: |
604 s->bitstream_bpp= 16; | |
605 break; | |
4682 | 606 case PIX_FMT_RGB32: |
607 s->bitstream_bpp= 24; | |
608 break; | |
866 | 609 default: |
1598
932d306bf1dc
av_log() patch by (Michel Bardiaux <mbardiaux at peaktime dot be>)
michael
parents:
1529
diff
changeset
|
610 av_log(avctx, AV_LOG_ERROR, "format not supported\n"); |
866 | 611 return -1; |
612 } | |
7823
4525dcd81357
Bump Major version, this commit is almost just renaming bits_per_sample to
michael
parents:
7782
diff
changeset
|
613 avctx->bits_per_coded_sample= s->bitstream_bpp; |
866 | 614 s->decorrelate= s->bitstream_bpp >= 24; |
615 s->predictor= avctx->prediction_method; | |
2237
d43321e67acd
(non)interlaced huffyuv patch by (Loren Merritt <lorenm at u dot washington dot edu>)
michael
parents:
2233
diff
changeset
|
616 s->interlaced= avctx->flags&CODEC_FLAG_INTERLACED_ME ? 1 : 0; |
2369 | 617 if(avctx->context_model==1){ |
618 s->context= avctx->context_model; | |
619 if(s->flags & (CODEC_FLAG_PASS1|CODEC_FLAG_PASS2)){ | |
620 av_log(avctx, AV_LOG_ERROR, "context=1 is not compatible with 2 pass huffyuv encoding\n"); | |
621 return -1; | |
622 } | |
623 }else s->context= 0; | |
2967 | 624 |
2373 | 625 if(avctx->codec->id==CODEC_ID_HUFFYUV){ |
626 if(avctx->pix_fmt==PIX_FMT_YUV420P){ | |
627 av_log(avctx, AV_LOG_ERROR, "Error: YV12 is not supported by huffyuv; use vcodec=ffvhuff or format=422p\n"); | |
628 return -1; | |
629 } | |
630 if(avctx->context_model){ | |
631 av_log(avctx, AV_LOG_ERROR, "Error: per-frame huffman tables are not supported by huffyuv; use vcodec=ffvhuff\n"); | |
632 return -1; | |
633 } | |
2422 | 634 if(s->interlaced != ( s->height > 288 )) |
2373 | 635 av_log(avctx, AV_LOG_INFO, "using huffyuv 2.2.0 or newer interlacing flag\n"); |
636 } | |
2967 | 637 |
4682 | 638 if(s->bitstream_bpp>=24 && s->predictor==MEDIAN){ |
639 av_log(avctx, AV_LOG_ERROR, "Error: RGB is incompatible with median predictor\n"); | |
640 return -1; | |
641 } | |
642 | |
643 ((uint8_t*)avctx->extradata)[0]= s->predictor | (s->decorrelate << 6); | |
866 | 644 ((uint8_t*)avctx->extradata)[1]= s->bitstream_bpp; |
2374 | 645 ((uint8_t*)avctx->extradata)[2]= s->interlaced ? 0x10 : 0x20; |
2369 | 646 if(s->context) |
647 ((uint8_t*)avctx->extradata)[2]|= 0x40; | |
866 | 648 ((uint8_t*)avctx->extradata)[3]= 0; |
649 s->avctx->extradata_size= 4; | |
2967 | 650 |
866 | 651 if(avctx->stats_in){ |
652 char *p= avctx->stats_in; | |
2967 | 653 |
866 | 654 for(i=0; i<3; i++) |
655 for(j=0; j<256; j++) | |
656 s->stats[i][j]= 1; | |
657 | |
658 for(;;){ | |
659 for(i=0; i<3; i++){ | |
660 char *next; | |
661 | |
662 for(j=0; j<256; j++){ | |
663 s->stats[i][j]+= strtol(p, &next, 0); | |
664 if(next==p) return -1; | |
665 p=next; | |
2967 | 666 } |
866 | 667 } |
668 if(p[0]==0 || p[1]==0 || p[2]==0) break; | |
669 } | |
670 }else{ | |
671 for(i=0; i<3; i++) | |
672 for(j=0; j<256; j++){ | |
673 int d= FFMIN(j, 256-j); | |
2967 | 674 |
866 | 675 s->stats[i][j]= 100000000/(d+1); |
676 } | |
677 } | |
2967 | 678 |
866 | 679 for(i=0; i<3; i++){ |
680 generate_len_table(s->len[i], s->stats[i], 256); | |
681 | |
682 if(generate_bits_table(s->bits[i], s->len[i])<0){ | |
683 return -1; | |
684 } | |
2967 | 685 |
2369 | 686 s->avctx->extradata_size+= |
687 store_table(s, s->len[i], &((uint8_t*)s->avctx->extradata)[s->avctx->extradata_size]); | |
866 | 688 } |
689 | |
2369 | 690 if(s->context){ |
691 for(i=0; i<3; i++){ | |
2422 | 692 int pels = s->width*s->height / (i?40:10); |
2369 | 693 for(j=0; j<256; j++){ |
694 int d= FFMIN(j, 256-j); | |
695 s->stats[i][j]= pels/(d+1); | |
696 } | |
697 } | |
698 }else{ | |
699 for(i=0; i<3; i++) | |
700 for(j=0; j<256; j++) | |
701 s->stats[i][j]= 0; | |
702 } | |
2967 | 703 |
7823
4525dcd81357
Bump Major version, this commit is almost just renaming bits_per_sample to
michael
parents:
7782
diff
changeset
|
704 // printf("pred:%d bpp:%d hbpp:%d il:%d\n", s->predictor, s->bitstream_bpp, avctx->bits_per_coded_sample, s->interlaced); |
1232
e88d3b1fb2a1
more #ifdef CONFIG_ENCODERS by (Wolfgang Hesseler <qv at multimediaware dot com>)
michaelni
parents:
1228
diff
changeset
|
705 |
2511 | 706 alloc_temp(s); |
707 | |
866 | 708 s->picture_number=0; |
1232
e88d3b1fb2a1
more #ifdef CONFIG_ENCODERS by (Wolfgang Hesseler <qv at multimediaware dot com>)
michaelni
parents:
1228
diff
changeset
|
709 |
866 | 710 return 0; |
711 } | |
8590 | 712 #endif /* CONFIG_HUFFYUV_ENCODER || CONFIG_FFVHUFF_ENCODER */ |
866 | 713 |
5063
d5640ea6d4a6
merge huffman tables so that we read 2 symbols at a time. 30% faster huffyuv decoding.
lorenm
parents:
5040
diff
changeset
|
714 /* TODO instead of restarting the read when the code isn't in the first level |
d5640ea6d4a6
merge huffman tables so that we read 2 symbols at a time. 30% faster huffyuv decoding.
lorenm
parents:
5040
diff
changeset
|
715 * of the joint table, jump into the 2nd level of the individual table. */ |
d5640ea6d4a6
merge huffman tables so that we read 2 symbols at a time. 30% faster huffyuv decoding.
lorenm
parents:
5040
diff
changeset
|
716 #define READ_2PIX(dst0, dst1, plane1){\ |
5072
1aec7fab94c4
use sparse huffman tables. 1.5% faster huffyuv decoding.
lorenm
parents:
5063
diff
changeset
|
717 uint16_t code = get_vlc2(&s->gb, s->vlc[3+plane1].table, VLC_BITS, 1);\ |
1aec7fab94c4
use sparse huffman tables. 1.5% faster huffyuv decoding.
lorenm
parents:
5063
diff
changeset
|
718 if(code != 0xffff){\ |
1aec7fab94c4
use sparse huffman tables. 1.5% faster huffyuv decoding.
lorenm
parents:
5063
diff
changeset
|
719 dst0 = code>>8;\ |
1aec7fab94c4
use sparse huffman tables. 1.5% faster huffyuv decoding.
lorenm
parents:
5063
diff
changeset
|
720 dst1 = code;\ |
5063
d5640ea6d4a6
merge huffman tables so that we read 2 symbols at a time. 30% faster huffyuv decoding.
lorenm
parents:
5040
diff
changeset
|
721 }else{\ |
d5640ea6d4a6
merge huffman tables so that we read 2 symbols at a time. 30% faster huffyuv decoding.
lorenm
parents:
5040
diff
changeset
|
722 dst0 = get_vlc2(&s->gb, s->vlc[0].table, VLC_BITS, 3);\ |
d5640ea6d4a6
merge huffman tables so that we read 2 symbols at a time. 30% faster huffyuv decoding.
lorenm
parents:
5040
diff
changeset
|
723 dst1 = get_vlc2(&s->gb, s->vlc[plane1].table, VLC_BITS, 3);\ |
d5640ea6d4a6
merge huffman tables so that we read 2 symbols at a time. 30% faster huffyuv decoding.
lorenm
parents:
5040
diff
changeset
|
724 }\ |
d5640ea6d4a6
merge huffman tables so that we read 2 symbols at a time. 30% faster huffyuv decoding.
lorenm
parents:
5040
diff
changeset
|
725 } |
d5640ea6d4a6
merge huffman tables so that we read 2 symbols at a time. 30% faster huffyuv decoding.
lorenm
parents:
5040
diff
changeset
|
726 |
866 | 727 static void decode_422_bitstream(HYuvContext *s, int count){ |
728 int i; | |
1232
e88d3b1fb2a1
more #ifdef CONFIG_ENCODERS by (Wolfgang Hesseler <qv at multimediaware dot com>)
michaelni
parents:
1228
diff
changeset
|
729 |
866 | 730 count/=2; |
2967 | 731 |
9906 | 732 if(count >= (s->gb.size_in_bits - get_bits_count(&s->gb))/(31*4)){ |
733 for(i=0; i<count && get_bits_count(&s->gb) < s->gb.size_in_bits; i++){ | |
734 READ_2PIX(s->temp[0][2*i ], s->temp[1][i], 1); | |
735 READ_2PIX(s->temp[0][2*i+1], s->temp[2][i], 2); | |
736 } | |
737 }else{ | |
866 | 738 for(i=0; i<count; i++){ |
5063
d5640ea6d4a6
merge huffman tables so that we read 2 symbols at a time. 30% faster huffyuv decoding.
lorenm
parents:
5040
diff
changeset
|
739 READ_2PIX(s->temp[0][2*i ], s->temp[1][i], 1); |
d5640ea6d4a6
merge huffman tables so that we read 2 symbols at a time. 30% faster huffyuv decoding.
lorenm
parents:
5040
diff
changeset
|
740 READ_2PIX(s->temp[0][2*i+1], s->temp[2][i], 2); |
866 | 741 } |
9906 | 742 } |
866 | 743 } |
744 | |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
745 static void decode_gray_bitstream(HYuvContext *s, int count){ |
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
746 int i; |
2967 | 747 |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
748 count/=2; |
2967 | 749 |
9906 | 750 if(count >= (s->gb.size_in_bits - get_bits_count(&s->gb))/(31*2)){ |
751 for(i=0; i<count && get_bits_count(&s->gb) < s->gb.size_in_bits; i++){ | |
752 READ_2PIX(s->temp[0][2*i ], s->temp[0][2*i+1], 0); | |
753 } | |
754 }else{ | |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
755 for(i=0; i<count; i++){ |
5063
d5640ea6d4a6
merge huffman tables so that we read 2 symbols at a time. 30% faster huffyuv decoding.
lorenm
parents:
5040
diff
changeset
|
756 READ_2PIX(s->temp[0][2*i ], s->temp[0][2*i+1], 0); |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
757 } |
9906 | 758 } |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
759 } |
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
760 |
8590 | 761 #if CONFIG_HUFFYUV_ENCODER || CONFIG_FFVHUFF_ENCODER |
9694
39f6d1f21ef8
Add offset argument to encode_422_bitstream() to control where
heydowns
parents:
9428
diff
changeset
|
762 static int encode_422_bitstream(HYuvContext *s, int offset, int count){ |
866 | 763 int i; |
9694
39f6d1f21ef8
Add offset argument to encode_422_bitstream() to control where
heydowns
parents:
9428
diff
changeset
|
764 const uint8_t *y = s->temp[0] + offset; |
39f6d1f21ef8
Add offset argument to encode_422_bitstream() to control where
heydowns
parents:
9428
diff
changeset
|
765 const uint8_t *u = s->temp[1] + offset/2; |
39f6d1f21ef8
Add offset argument to encode_422_bitstream() to control where
heydowns
parents:
9428
diff
changeset
|
766 const uint8_t *v = s->temp[2] + offset/2; |
2967 | 767 |
2422 | 768 if(s->pb.buf_end - s->pb.buf - (put_bits_count(&s->pb)>>3) < 2*4*count){ |
769 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n"); | |
770 return -1; | |
771 } | |
2967 | 772 |
5026 | 773 #define LOAD4\ |
9694
39f6d1f21ef8
Add offset argument to encode_422_bitstream() to control where
heydowns
parents:
9428
diff
changeset
|
774 int y0 = y[2*i];\ |
39f6d1f21ef8
Add offset argument to encode_422_bitstream() to control where
heydowns
parents:
9428
diff
changeset
|
775 int y1 = y[2*i+1];\ |
39f6d1f21ef8
Add offset argument to encode_422_bitstream() to control where
heydowns
parents:
9428
diff
changeset
|
776 int u0 = u[i];\ |
39f6d1f21ef8
Add offset argument to encode_422_bitstream() to control where
heydowns
parents:
9428
diff
changeset
|
777 int v0 = v[i]; |
5026 | 778 |
866 | 779 count/=2; |
780 if(s->flags&CODEC_FLAG_PASS1){ | |
781 for(i=0; i<count; i++){ | |
5026 | 782 LOAD4; |
783 s->stats[0][y0]++; | |
784 s->stats[1][u0]++; | |
785 s->stats[0][y1]++; | |
786 s->stats[2][v0]++; | |
866 | 787 } |
2501 | 788 } |
789 if(s->avctx->flags2&CODEC_FLAG2_NO_OUTPUT) | |
790 return 0; | |
791 if(s->context){ | |
2369 | 792 for(i=0; i<count; i++){ |
5026 | 793 LOAD4; |
794 s->stats[0][y0]++; | |
795 put_bits(&s->pb, s->len[0][y0], s->bits[0][y0]); | |
796 s->stats[1][u0]++; | |
797 put_bits(&s->pb, s->len[1][u0], s->bits[1][u0]); | |
798 s->stats[0][y1]++; | |
799 put_bits(&s->pb, s->len[0][y1], s->bits[0][y1]); | |
800 s->stats[2][v0]++; | |
801 put_bits(&s->pb, s->len[2][v0], s->bits[2][v0]); | |
2369 | 802 } |
866 | 803 }else{ |
804 for(i=0; i<count; i++){ | |
5026 | 805 LOAD4; |
806 put_bits(&s->pb, s->len[0][y0], s->bits[0][y0]); | |
807 put_bits(&s->pb, s->len[1][u0], s->bits[1][u0]); | |
808 put_bits(&s->pb, s->len[0][y1], s->bits[0][y1]); | |
809 put_bits(&s->pb, s->len[2][v0], s->bits[2][v0]); | |
866 | 810 } |
811 } | |
2422 | 812 return 0; |
866 | 813 } |
814 | |
2422 | 815 static int encode_gray_bitstream(HYuvContext *s, int count){ |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
816 int i; |
2967 | 817 |
2422 | 818 if(s->pb.buf_end - s->pb.buf - (put_bits_count(&s->pb)>>3) < 4*count){ |
819 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n"); | |
820 return -1; | |
821 } | |
822 | |
5026 | 823 #define LOAD2\ |
824 int y0 = s->temp[0][2*i];\ | |
825 int y1 = s->temp[0][2*i+1]; | |
826 #define STAT2\ | |
827 s->stats[0][y0]++;\ | |
828 s->stats[0][y1]++; | |
829 #define WRITE2\ | |
830 put_bits(&s->pb, s->len[0][y0], s->bits[0][y0]);\ | |
831 put_bits(&s->pb, s->len[0][y1], s->bits[0][y1]); | |
832 | |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
833 count/=2; |
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
834 if(s->flags&CODEC_FLAG_PASS1){ |
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
835 for(i=0; i<count; i++){ |
5026 | 836 LOAD2; |
837 STAT2; | |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
838 } |
2501 | 839 } |
840 if(s->avctx->flags2&CODEC_FLAG2_NO_OUTPUT) | |
841 return 0; | |
2967 | 842 |
2501 | 843 if(s->context){ |
2369 | 844 for(i=0; i<count; i++){ |
5026 | 845 LOAD2; |
846 STAT2; | |
847 WRITE2; | |
2369 | 848 } |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
849 }else{ |
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
850 for(i=0; i<count; i++){ |
5026 | 851 LOAD2; |
852 WRITE2; | |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
853 } |
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
854 } |
2422 | 855 return 0; |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
856 } |
8590 | 857 #endif /* CONFIG_HUFFYUV_ENCODER || CONFIG_FFVHUFF_ENCODER */ |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
858 |
5074 | 859 static av_always_inline void decode_bgr_1(HYuvContext *s, int count, int decorrelate, int alpha){ |
866 | 860 int i; |
5074 | 861 for(i=0; i<count; i++){ |
862 int code = get_vlc2(&s->gb, s->vlc[3].table, VLC_BITS, 1); | |
863 if(code != -1){ | |
864 *(uint32_t*)&s->temp[0][4*i] = s->pix_bgr_map[code]; | |
865 }else if(decorrelate){ | |
866 s->temp[0][4*i+G] = get_vlc2(&s->gb, s->vlc[1].table, VLC_BITS, 3); | |
867 s->temp[0][4*i+B] = get_vlc2(&s->gb, s->vlc[0].table, VLC_BITS, 3) + s->temp[0][4*i+G]; | |
868 s->temp[0][4*i+R] = get_vlc2(&s->gb, s->vlc[2].table, VLC_BITS, 3) + s->temp[0][4*i+G]; | |
866 | 869 }else{ |
5074 | 870 s->temp[0][4*i+B] = get_vlc2(&s->gb, s->vlc[0].table, VLC_BITS, 3); |
871 s->temp[0][4*i+G] = get_vlc2(&s->gb, s->vlc[1].table, VLC_BITS, 3); | |
872 s->temp[0][4*i+R] = get_vlc2(&s->gb, s->vlc[2].table, VLC_BITS, 3); | |
866 | 873 } |
5074 | 874 if(alpha) |
875 get_vlc2(&s->gb, s->vlc[2].table, VLC_BITS, 3); //?! | |
876 } | |
877 } | |
878 | |
879 static void decode_bgr_bitstream(HYuvContext *s, int count){ | |
880 if(s->decorrelate){ | |
881 if(s->bitstream_bpp==24) | |
882 decode_bgr_1(s, count, 1, 0); | |
883 else | |
884 decode_bgr_1(s, count, 1, 1); | |
866 | 885 }else{ |
5074 | 886 if(s->bitstream_bpp==24) |
887 decode_bgr_1(s, count, 0, 0); | |
888 else | |
889 decode_bgr_1(s, count, 0, 1); | |
866 | 890 } |
891 } | |
892 | |
4682 | 893 static int encode_bgr_bitstream(HYuvContext *s, int count){ |
894 int i; | |
895 | |
896 if(s->pb.buf_end - s->pb.buf - (put_bits_count(&s->pb)>>3) < 3*4*count){ | |
897 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n"); | |
898 return -1; | |
899 } | |
900 | |
5026 | 901 #define LOAD3\ |
902 int g= s->temp[0][4*i+G];\ | |
903 int b= (s->temp[0][4*i+B] - g) & 0xff;\ | |
904 int r= (s->temp[0][4*i+R] - g) & 0xff; | |
905 #define STAT3\ | |
906 s->stats[0][b]++;\ | |
907 s->stats[1][g]++;\ | |
908 s->stats[2][r]++; | |
909 #define WRITE3\ | |
910 put_bits(&s->pb, s->len[1][g], s->bits[1][g]);\ | |
911 put_bits(&s->pb, s->len[0][b], s->bits[0][b]);\ | |
912 put_bits(&s->pb, s->len[2][r], s->bits[2][r]); | |
913 | |
4682 | 914 if((s->flags&CODEC_FLAG_PASS1) && (s->avctx->flags2&CODEC_FLAG2_NO_OUTPUT)){ |
915 for(i=0; i<count; i++){ | |
5026 | 916 LOAD3; |
917 STAT3; | |
4682 | 918 } |
919 }else if(s->context || (s->flags&CODEC_FLAG_PASS1)){ | |
920 for(i=0; i<count; i++){ | |
5026 | 921 LOAD3; |
922 STAT3; | |
923 WRITE3; | |
4682 | 924 } |
925 }else{ | |
926 for(i=0; i<count; i++){ | |
5026 | 927 LOAD3; |
928 WRITE3; | |
4682 | 929 } |
930 } | |
931 return 0; | |
932 } | |
933 | |
8590 | 934 #if CONFIG_HUFFYUV_DECODER || CONFIG_FFVHUFF_DECODER |
871 | 935 static void draw_slice(HYuvContext *s, int y){ |
936 int h, cy; | |
1368 | 937 int offset[4]; |
2967 | 938 |
939 if(s->avctx->draw_horiz_band==NULL) | |
871 | 940 return; |
2967 | 941 |
871 | 942 h= y - s->last_slice_end; |
943 y -= h; | |
2967 | 944 |
871 | 945 if(s->bitstream_bpp==12){ |
946 cy= y>>1; | |
947 }else{ | |
948 cy= y; | |
949 } | |
1368 | 950 |
951 offset[0] = s->picture.linesize[0]*y; | |
952 offset[1] = s->picture.linesize[1]*cy; | |
953 offset[2] = s->picture.linesize[2]*cy; | |
954 offset[3] = 0; | |
871 | 955 emms_c(); |
956 | |
1370 | 957 s->avctx->draw_horiz_band(s->avctx, &s->picture, offset, y, 3, h); |
2967 | 958 |
871 | 959 s->last_slice_end= y + h; |
960 } | |
961 | |
9355
54bc8a2727b0
Implement avcodec_decode_video2(), _audio3() and _subtitle2() which takes an
rbultje
parents:
9346
diff
changeset
|
962 static int decode_frame(AVCodecContext *avctx, void *data, int *data_size, AVPacket *avpkt){ |
54bc8a2727b0
Implement avcodec_decode_video2(), _audio3() and _subtitle2() which takes an
rbultje
parents:
9346
diff
changeset
|
963 const uint8_t *buf = avpkt->data; |
54bc8a2727b0
Implement avcodec_decode_video2(), _audio3() and _subtitle2() which takes an
rbultje
parents:
9346
diff
changeset
|
964 int buf_size = avpkt->size; |
866 | 965 HYuvContext *s = avctx->priv_data; |
966 const int width= s->width; | |
967 const int width2= s->width>>1; | |
968 const int height= s->height; | |
870 | 969 int fake_ystride, fake_ustride, fake_vstride; |
925 | 970 AVFrame * const p= &s->picture; |
2369 | 971 int table_size= 0; |
866 | 972 |
925 | 973 AVFrame *picture = data; |
866 | 974 |
9415
141badec76fc
Add a av_fast_malloc function and replace several uses of av_fast_realloc,
reimar
parents:
9411
diff
changeset
|
975 av_fast_malloc(&s->bitstream_buffer, &s->bitstream_buffer_size, buf_size + FF_INPUT_BUFFER_PADDING_SIZE); |
141badec76fc
Add a av_fast_malloc function and replace several uses of av_fast_realloc,
reimar
parents:
9411
diff
changeset
|
976 if (!s->bitstream_buffer) |
141badec76fc
Add a av_fast_malloc function and replace several uses of av_fast_realloc,
reimar
parents:
9411
diff
changeset
|
977 return AVERROR(ENOMEM); |
866 | 978 |
6234 | 979 s->dsp.bswap_buf((uint32_t*)s->bitstream_buffer, (const uint32_t*)buf, buf_size/4); |
2967 | 980 |
1228 | 981 if(p->data[0]) |
982 avctx->release_buffer(avctx, p); | |
983 | |
903 | 984 p->reference= 0; |
985 if(avctx->get_buffer(avctx, p) < 0){ | |
1598
932d306bf1dc
av_log() patch by (Michel Bardiaux <mbardiaux at peaktime dot be>)
michael
parents:
1529
diff
changeset
|
986 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n"); |
903 | 987 return -1; |
988 } | |
2967 | 989 |
2369 | 990 if(s->context){ |
991 table_size = read_huffman_tables(s, s->bitstream_buffer, buf_size); | |
992 if(table_size < 0) | |
993 return -1; | |
994 } | |
995 | |
3201 | 996 if((unsigned)(buf_size-table_size) >= INT_MAX/8) |
997 return -1; | |
998 | |
2369 | 999 init_get_bits(&s->gb, s->bitstream_buffer+table_size, (buf_size-table_size)*8); |
870 | 1000 |
903 | 1001 fake_ystride= s->interlaced ? p->linesize[0]*2 : p->linesize[0]; |
1002 fake_ustride= s->interlaced ? p->linesize[1]*2 : p->linesize[1]; | |
1003 fake_vstride= s->interlaced ? p->linesize[2]*2 : p->linesize[2]; | |
2967 | 1004 |
871 | 1005 s->last_slice_end= 0; |
2967 | 1006 |
866 | 1007 if(s->bitstream_bpp<24){ |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1008 int y, cy; |
866 | 1009 int lefty, leftu, leftv; |
1010 int lefttopy, lefttopu, lefttopv; | |
2967 | 1011 |
866 | 1012 if(s->yuy2){ |
903 | 1013 p->data[0][3]= get_bits(&s->gb, 8); |
1014 p->data[0][2]= get_bits(&s->gb, 8); | |
1015 p->data[0][1]= get_bits(&s->gb, 8); | |
1016 p->data[0][0]= get_bits(&s->gb, 8); | |
2967 | 1017 |
2628
511e3afc43e1
Ministry of English Composition, reporting for duty (and the word is "skipped", not "skiped"; "skiped" would rhyme with "hyped")
melanson
parents:
2552
diff
changeset
|
1018 av_log(avctx, AV_LOG_ERROR, "YUY2 output is not implemented yet\n"); |
866 | 1019 return -1; |
1020 }else{ | |
2967 | 1021 |
903 | 1022 leftv= p->data[2][0]= get_bits(&s->gb, 8); |
1023 lefty= p->data[0][1]= get_bits(&s->gb, 8); | |
1024 leftu= p->data[1][0]= get_bits(&s->gb, 8); | |
1025 p->data[0][0]= get_bits(&s->gb, 8); | |
2967 | 1026 |
866 | 1027 switch(s->predictor){ |
1028 case LEFT: | |
1029 case PLANE: | |
1030 decode_422_bitstream(s, width-2); | |
903 | 1031 lefty= add_left_prediction(p->data[0] + 2, s->temp[0], width-2, lefty); |
866 | 1032 if(!(s->flags&CODEC_FLAG_GRAY)){ |
903 | 1033 leftu= add_left_prediction(p->data[1] + 1, s->temp[1], width2-1, leftu); |
1034 leftv= add_left_prediction(p->data[2] + 1, s->temp[2], width2-1, leftv); | |
866 | 1035 } |
1036 | |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1037 for(cy=y=1; y<s->height; y++,cy++){ |
866 | 1038 uint8_t *ydst, *udst, *vdst; |
2967 | 1039 |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1040 if(s->bitstream_bpp==12){ |
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1041 decode_gray_bitstream(s, width); |
2967 | 1042 |
903 | 1043 ydst= p->data[0] + p->linesize[0]*y; |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1044 |
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1045 lefty= add_left_prediction(ydst, s->temp[0], width, lefty); |
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1046 if(s->predictor == PLANE){ |
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1047 if(y>s->interlaced) |
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1048 s->dsp.add_bytes(ydst, ydst - fake_ystride, width); |
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1049 } |
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1050 y++; |
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1051 if(y>=s->height) break; |
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1052 } |
2967 | 1053 |
871 | 1054 draw_slice(s, y); |
2967 | 1055 |
903 | 1056 ydst= p->data[0] + p->linesize[0]*y; |
1057 udst= p->data[1] + p->linesize[1]*cy; | |
1058 vdst= p->data[2] + p->linesize[2]*cy; | |
2967 | 1059 |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1060 decode_422_bitstream(s, width); |
866 | 1061 lefty= add_left_prediction(ydst, s->temp[0], width, lefty); |
1062 if(!(s->flags&CODEC_FLAG_GRAY)){ | |
1063 leftu= add_left_prediction(udst, s->temp[1], width2, leftu); | |
1064 leftv= add_left_prediction(vdst, s->temp[2], width2, leftv); | |
1065 } | |
1066 if(s->predictor == PLANE){ | |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1067 if(cy>s->interlaced){ |
866 | 1068 s->dsp.add_bytes(ydst, ydst - fake_ystride, width); |
1069 if(!(s->flags&CODEC_FLAG_GRAY)){ | |
1070 s->dsp.add_bytes(udst, udst - fake_ustride, width2); | |
1071 s->dsp.add_bytes(vdst, vdst - fake_vstride, width2); | |
1072 } | |
1073 } | |
1074 } | |
1075 } | |
871 | 1076 draw_slice(s, height); |
2967 | 1077 |
866 | 1078 break; |
1079 case MEDIAN: | |
1080 /* first line except first 2 pixels is left predicted */ | |
1081 decode_422_bitstream(s, width-2); | |
903 | 1082 lefty= add_left_prediction(p->data[0] + 2, s->temp[0], width-2, lefty); |
866 | 1083 if(!(s->flags&CODEC_FLAG_GRAY)){ |
903 | 1084 leftu= add_left_prediction(p->data[1] + 1, s->temp[1], width2-1, leftu); |
1085 leftv= add_left_prediction(p->data[2] + 1, s->temp[2], width2-1, leftv); | |
866 | 1086 } |
2967 | 1087 |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1088 cy=y=1; |
2967 | 1089 |
866 | 1090 /* second line is left predicted for interlaced case */ |
1091 if(s->interlaced){ | |
1092 decode_422_bitstream(s, width); | |
903 | 1093 lefty= add_left_prediction(p->data[0] + p->linesize[0], s->temp[0], width, lefty); |
866 | 1094 if(!(s->flags&CODEC_FLAG_GRAY)){ |
903 | 1095 leftu= add_left_prediction(p->data[1] + p->linesize[2], s->temp[1], width2, leftu); |
1096 leftv= add_left_prediction(p->data[2] + p->linesize[1], s->temp[2], width2, leftv); | |
866 | 1097 } |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1098 y++; cy++; |
866 | 1099 } |
1100 | |
1101 /* next 4 pixels are left predicted too */ | |
1102 decode_422_bitstream(s, 4); | |
903 | 1103 lefty= add_left_prediction(p->data[0] + fake_ystride, s->temp[0], 4, lefty); |
866 | 1104 if(!(s->flags&CODEC_FLAG_GRAY)){ |
903 | 1105 leftu= add_left_prediction(p->data[1] + fake_ustride, s->temp[1], 2, leftu); |
1106 leftv= add_left_prediction(p->data[2] + fake_vstride, s->temp[2], 2, leftv); | |
866 | 1107 } |
1108 | |
1109 /* next line except the first 4 pixels is median predicted */ | |
903 | 1110 lefttopy= p->data[0][3]; |
866 | 1111 decode_422_bitstream(s, width-4); |
8760 | 1112 s->dsp.add_hfyu_median_prediction(p->data[0] + fake_ystride+4, p->data[0]+4, s->temp[0], width-4, &lefty, &lefttopy); |
866 | 1113 if(!(s->flags&CODEC_FLAG_GRAY)){ |
903 | 1114 lefttopu= p->data[1][1]; |
1115 lefttopv= p->data[2][1]; | |
8760 | 1116 s->dsp.add_hfyu_median_prediction(p->data[1] + fake_ustride+2, p->data[1]+2, s->temp[1], width2-2, &leftu, &lefttopu); |
1117 s->dsp.add_hfyu_median_prediction(p->data[2] + fake_vstride+2, p->data[2]+2, s->temp[2], width2-2, &leftv, &lefttopv); | |
866 | 1118 } |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1119 y++; cy++; |
2967 | 1120 |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1121 for(; y<height; y++,cy++){ |
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1122 uint8_t *ydst, *udst, *vdst; |
866 | 1123 |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1124 if(s->bitstream_bpp==12){ |
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1125 while(2*cy > y){ |
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1126 decode_gray_bitstream(s, width); |
903 | 1127 ydst= p->data[0] + p->linesize[0]*y; |
8760 | 1128 s->dsp.add_hfyu_median_prediction(ydst, ydst - fake_ystride, s->temp[0], width, &lefty, &lefttopy); |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1129 y++; |
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1130 } |
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1131 if(y>=height) break; |
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1132 } |
871 | 1133 draw_slice(s, y); |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1134 |
866 | 1135 decode_422_bitstream(s, width); |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1136 |
903 | 1137 ydst= p->data[0] + p->linesize[0]*y; |
1138 udst= p->data[1] + p->linesize[1]*cy; | |
1139 vdst= p->data[2] + p->linesize[2]*cy; | |
866 | 1140 |
8760 | 1141 s->dsp.add_hfyu_median_prediction(ydst, ydst - fake_ystride, s->temp[0], width, &lefty, &lefttopy); |
866 | 1142 if(!(s->flags&CODEC_FLAG_GRAY)){ |
8760 | 1143 s->dsp.add_hfyu_median_prediction(udst, udst - fake_ustride, s->temp[1], width2, &leftu, &lefttopu); |
1144 s->dsp.add_hfyu_median_prediction(vdst, vdst - fake_vstride, s->temp[2], width2, &leftv, &lefttopv); | |
866 | 1145 } |
1146 } | |
871 | 1147 |
1148 draw_slice(s, height); | |
866 | 1149 break; |
1150 } | |
1151 } | |
1152 }else{ | |
1153 int y; | |
1154 int leftr, leftg, leftb; | |
903 | 1155 const int last_line= (height-1)*p->linesize[0]; |
2967 | 1156 |
866 | 1157 if(s->bitstream_bpp==32){ |
2177 | 1158 skip_bits(&s->gb, 8); |
1159 leftr= p->data[0][last_line+R]= get_bits(&s->gb, 8); | |
1160 leftg= p->data[0][last_line+G]= get_bits(&s->gb, 8); | |
1161 leftb= p->data[0][last_line+B]= get_bits(&s->gb, 8); | |
866 | 1162 }else{ |
2177 | 1163 leftr= p->data[0][last_line+R]= get_bits(&s->gb, 8); |
1164 leftg= p->data[0][last_line+G]= get_bits(&s->gb, 8); | |
1165 leftb= p->data[0][last_line+B]= get_bits(&s->gb, 8); | |
866 | 1166 skip_bits(&s->gb, 8); |
1167 } | |
2967 | 1168 |
866 | 1169 if(s->bgr32){ |
1170 switch(s->predictor){ | |
1171 case LEFT: | |
1172 case PLANE: | |
1173 decode_bgr_bitstream(s, width-1); | |
903 | 1174 add_left_prediction_bgr32(p->data[0] + last_line+4, s->temp[0], width-1, &leftr, &leftg, &leftb); |
866 | 1175 |
5129 | 1176 for(y=s->height-2; y>=0; y--){ //Yes it is stored upside down. |
866 | 1177 decode_bgr_bitstream(s, width); |
2967 | 1178 |
903 | 1179 add_left_prediction_bgr32(p->data[0] + p->linesize[0]*y, s->temp[0], width, &leftr, &leftg, &leftb); |
866 | 1180 if(s->predictor == PLANE){ |
2351 | 1181 if((y&s->interlaced)==0 && y<s->height-1-s->interlaced){ |
2967 | 1182 s->dsp.add_bytes(p->data[0] + p->linesize[0]*y, |
903 | 1183 p->data[0] + p->linesize[0]*y + fake_ystride, fake_ystride); |
866 | 1184 } |
1185 } | |
1186 } | |
2628
511e3afc43e1
Ministry of English Composition, reporting for duty (and the word is "skipped", not "skiped"; "skiped" would rhyme with "hyped")
melanson
parents:
2552
diff
changeset
|
1187 draw_slice(s, height); // just 1 large slice as this is not possible in reverse order |
866 | 1188 break; |
1189 default: | |
1598
932d306bf1dc
av_log() patch by (Michel Bardiaux <mbardiaux at peaktime dot be>)
michael
parents:
1529
diff
changeset
|
1190 av_log(avctx, AV_LOG_ERROR, "prediction type not supported!\n"); |
866 | 1191 } |
1192 }else{ | |
1193 | |
2628
511e3afc43e1
Ministry of English Composition, reporting for duty (and the word is "skipped", not "skiped"; "skiped" would rhyme with "hyped")
melanson
parents:
2552
diff
changeset
|
1194 av_log(avctx, AV_LOG_ERROR, "BGR24 output is not implemented yet\n"); |
866 | 1195 return -1; |
1196 } | |
1197 } | |
1198 emms_c(); | |
2967 | 1199 |
903 | 1200 *picture= *p; |
925 | 1201 *data_size = sizeof(AVFrame); |
2967 | 1202 |
3234
823272bdb4f7
dont forget table_size in the decode_frame return value
michael
parents:
3201
diff
changeset
|
1203 return (get_bits_count(&s->gb)+31)/32*4 + table_size; |
866 | 1204 } |
8590 | 1205 #endif /* CONFIG_HUFFYUV_DECODER || CONFIG_FFVHUFF_DECODER */ |
866 | 1206 |
2422 | 1207 static int common_end(HYuvContext *s){ |
1208 int i; | |
2967 | 1209 |
2422 | 1210 for(i=0; i<3; i++){ |
1211 av_freep(&s->temp[i]); | |
1212 } | |
1213 return 0; | |
1214 } | |
1215 | |
8590 | 1216 #if CONFIG_HUFFYUV_DECODER || CONFIG_FFVHUFF_DECODER |
6517
48759bfbd073
Apply 'cold' attribute to init/uninit functions in libavcodec
zuxy
parents:
6234
diff
changeset
|
1217 static av_cold int decode_end(AVCodecContext *avctx) |
866 | 1218 { |
1219 HYuvContext *s = avctx->priv_data; | |
1220 int i; | |
2967 | 1221 |
2422 | 1222 common_end(s); |
1223 av_freep(&s->bitstream_buffer); | |
2967 | 1224 |
5221 | 1225 for(i=0; i<6; i++){ |
903 | 1226 free_vlc(&s->vlc[i]); |
1227 } | |
866 | 1228 |
1229 return 0; | |
1230 } | |
8590 | 1231 #endif /* CONFIG_HUFFYUV_DECODER || CONFIG_FFVHUFF_DECODER */ |
866 | 1232 |
8590 | 1233 #if CONFIG_HUFFYUV_ENCODER || CONFIG_FFVHUFF_ENCODER |
866 | 1234 static int encode_frame(AVCodecContext *avctx, unsigned char *buf, int buf_size, void *data){ |
1235 HYuvContext *s = avctx->priv_data; | |
925 | 1236 AVFrame *pict = data; |
866 | 1237 const int width= s->width; |
1238 const int width2= s->width>>1; | |
1239 const int height= s->height; | |
1240 const int fake_ystride= s->interlaced ? pict->linesize[0]*2 : pict->linesize[0]; | |
1241 const int fake_ustride= s->interlaced ? pict->linesize[1]*2 : pict->linesize[1]; | |
1242 const int fake_vstride= s->interlaced ? pict->linesize[2]*2 : pict->linesize[2]; | |
925 | 1243 AVFrame * const p= &s->picture; |
2369 | 1244 int i, j, size=0; |
866 | 1245 |
903 | 1246 *p = *pict; |
1006 | 1247 p->pict_type= FF_I_TYPE; |
1248 p->key_frame= 1; | |
2967 | 1249 |
2369 | 1250 if(s->context){ |
1251 for(i=0; i<3; i++){ | |
1252 generate_len_table(s->len[i], s->stats[i], 256); | |
1253 if(generate_bits_table(s->bits[i], s->len[i])<0) | |
1254 return -1; | |
1255 size+= store_table(s, s->len[i], &buf[size]); | |
1256 } | |
1257 | |
1258 for(i=0; i<3; i++) | |
1259 for(j=0; j<256; j++) | |
1260 s->stats[i][j] >>= 1; | |
1261 } | |
1262 | |
1263 init_put_bits(&s->pb, buf+size, buf_size-size); | |
1264 | |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1265 if(avctx->pix_fmt == PIX_FMT_YUV422P || avctx->pix_fmt == PIX_FMT_YUV420P){ |
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1266 int lefty, leftu, leftv, y, cy; |
866 | 1267 |
903 | 1268 put_bits(&s->pb, 8, leftv= p->data[2][0]); |
1269 put_bits(&s->pb, 8, lefty= p->data[0][1]); | |
1270 put_bits(&s->pb, 8, leftu= p->data[1][0]); | |
1271 put_bits(&s->pb, 8, p->data[0][0]); | |
2967 | 1272 |
9695
f1f5e78cf0aa
Fix passing of unaligned addresses to dsputil diff_bytes() by
heydowns
parents:
9694
diff
changeset
|
1273 lefty= sub_left_prediction(s, s->temp[0], p->data[0], width , 0); |
f1f5e78cf0aa
Fix passing of unaligned addresses to dsputil diff_bytes() by
heydowns
parents:
9694
diff
changeset
|
1274 leftu= sub_left_prediction(s, s->temp[1], p->data[1], width2, 0); |
f1f5e78cf0aa
Fix passing of unaligned addresses to dsputil diff_bytes() by
heydowns
parents:
9694
diff
changeset
|
1275 leftv= sub_left_prediction(s, s->temp[2], p->data[2], width2, 0); |
2967 | 1276 |
9695
f1f5e78cf0aa
Fix passing of unaligned addresses to dsputil diff_bytes() by
heydowns
parents:
9694
diff
changeset
|
1277 encode_422_bitstream(s, 2, width-2); |
2967 | 1278 |
866 | 1279 if(s->predictor==MEDIAN){ |
1280 int lefttopy, lefttopu, lefttopv; | |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1281 cy=y=1; |
866 | 1282 if(s->interlaced){ |
903 | 1283 lefty= sub_left_prediction(s, s->temp[0], p->data[0]+p->linesize[0], width , lefty); |
1284 leftu= sub_left_prediction(s, s->temp[1], p->data[1]+p->linesize[1], width2, leftu); | |
1285 leftv= sub_left_prediction(s, s->temp[2], p->data[2]+p->linesize[2], width2, leftv); | |
2967 | 1286 |
9694
39f6d1f21ef8
Add offset argument to encode_422_bitstream() to control where
heydowns
parents:
9428
diff
changeset
|
1287 encode_422_bitstream(s, 0, width); |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1288 y++; cy++; |
866 | 1289 } |
2967 | 1290 |
903 | 1291 lefty= sub_left_prediction(s, s->temp[0], p->data[0]+fake_ystride, 4, lefty); |
2190 | 1292 leftu= sub_left_prediction(s, s->temp[1], p->data[1]+fake_ustride, 2, leftu); |
1293 leftv= sub_left_prediction(s, s->temp[2], p->data[2]+fake_vstride, 2, leftv); | |
2967 | 1294 |
9694
39f6d1f21ef8
Add offset argument to encode_422_bitstream() to control where
heydowns
parents:
9428
diff
changeset
|
1295 encode_422_bitstream(s, 0, 4); |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1296 |
903 | 1297 lefttopy= p->data[0][3]; |
1298 lefttopu= p->data[1][1]; | |
1299 lefttopv= p->data[2][1]; | |
1527 | 1300 s->dsp.sub_hfyu_median_prediction(s->temp[0], p->data[0]+4, p->data[0] + fake_ystride+4, width-4 , &lefty, &lefttopy); |
1301 s->dsp.sub_hfyu_median_prediction(s->temp[1], p->data[1]+2, p->data[1] + fake_ustride+2, width2-2, &leftu, &lefttopu); | |
1302 s->dsp.sub_hfyu_median_prediction(s->temp[2], p->data[2]+2, p->data[2] + fake_vstride+2, width2-2, &leftv, &lefttopv); | |
9694
39f6d1f21ef8
Add offset argument to encode_422_bitstream() to control where
heydowns
parents:
9428
diff
changeset
|
1303 encode_422_bitstream(s, 0, width-4); |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1304 y++; cy++; |
866 | 1305 |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1306 for(; y<height; y++,cy++){ |
866 | 1307 uint8_t *ydst, *udst, *vdst; |
2967 | 1308 |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1309 if(s->bitstream_bpp==12){ |
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1310 while(2*cy > y){ |
903 | 1311 ydst= p->data[0] + p->linesize[0]*y; |
1527 | 1312 s->dsp.sub_hfyu_median_prediction(s->temp[0], ydst - fake_ystride, ydst, width , &lefty, &lefttopy); |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1313 encode_gray_bitstream(s, width); |
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1314 y++; |
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1315 } |
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1316 if(y>=height) break; |
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1317 } |
903 | 1318 ydst= p->data[0] + p->linesize[0]*y; |
1319 udst= p->data[1] + p->linesize[1]*cy; | |
1320 vdst= p->data[2] + p->linesize[2]*cy; | |
866 | 1321 |
1527 | 1322 s->dsp.sub_hfyu_median_prediction(s->temp[0], ydst - fake_ystride, ydst, width , &lefty, &lefttopy); |
1323 s->dsp.sub_hfyu_median_prediction(s->temp[1], udst - fake_ustride, udst, width2, &leftu, &lefttopu); | |
1324 s->dsp.sub_hfyu_median_prediction(s->temp[2], vdst - fake_vstride, vdst, width2, &leftv, &lefttopv); | |
866 | 1325 |
9694
39f6d1f21ef8
Add offset argument to encode_422_bitstream() to control where
heydowns
parents:
9428
diff
changeset
|
1326 encode_422_bitstream(s, 0, width); |
866 | 1327 } |
1328 }else{ | |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1329 for(cy=y=1; y<height; y++,cy++){ |
866 | 1330 uint8_t *ydst, *udst, *vdst; |
2967 | 1331 |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1332 /* encode a luma only line & y++ */ |
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1333 if(s->bitstream_bpp==12){ |
903 | 1334 ydst= p->data[0] + p->linesize[0]*y; |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1335 |
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1336 if(s->predictor == PLANE && s->interlaced < y){ |
871 | 1337 s->dsp.diff_bytes(s->temp[1], ydst, ydst - fake_ystride, width); |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1338 |
871 | 1339 lefty= sub_left_prediction(s, s->temp[0], s->temp[1], width , lefty); |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1340 }else{ |
871 | 1341 lefty= sub_left_prediction(s, s->temp[0], ydst, width , lefty); |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1342 } |
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1343 encode_gray_bitstream(s, width); |
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1344 y++; |
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1345 if(y>=height) break; |
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1346 } |
2967 | 1347 |
903 | 1348 ydst= p->data[0] + p->linesize[0]*y; |
1349 udst= p->data[1] + p->linesize[1]*cy; | |
1350 vdst= p->data[2] + p->linesize[2]*cy; | |
866 | 1351 |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1352 if(s->predictor == PLANE && s->interlaced < cy){ |
871 | 1353 s->dsp.diff_bytes(s->temp[1], ydst, ydst - fake_ystride, width); |
1354 s->dsp.diff_bytes(s->temp[2], udst, udst - fake_ustride, width2); | |
2511 | 1355 s->dsp.diff_bytes(s->temp[2] + width2, vdst, vdst - fake_vstride, width2); |
866 | 1356 |
871 | 1357 lefty= sub_left_prediction(s, s->temp[0], s->temp[1], width , lefty); |
1358 leftu= sub_left_prediction(s, s->temp[1], s->temp[2], width2, leftu); | |
2511 | 1359 leftv= sub_left_prediction(s, s->temp[2], s->temp[2] + width2, width2, leftv); |
866 | 1360 }else{ |
871 | 1361 lefty= sub_left_prediction(s, s->temp[0], ydst, width , lefty); |
1362 leftu= sub_left_prediction(s, s->temp[1], udst, width2, leftu); | |
1363 leftv= sub_left_prediction(s, s->temp[2], vdst, width2, leftv); | |
866 | 1364 } |
1365 | |
9694
39f6d1f21ef8
Add offset argument to encode_422_bitstream() to control where
heydowns
parents:
9428
diff
changeset
|
1366 encode_422_bitstream(s, 0, width); |
866 | 1367 } |
2967 | 1368 } |
4682 | 1369 }else if(avctx->pix_fmt == PIX_FMT_RGB32){ |
1370 uint8_t *data = p->data[0] + (height-1)*p->linesize[0]; | |
1371 const int stride = -p->linesize[0]; | |
1372 const int fake_stride = -fake_ystride; | |
1373 int y; | |
1374 int leftr, leftg, leftb; | |
1375 | |
1376 put_bits(&s->pb, 8, leftr= data[R]); | |
1377 put_bits(&s->pb, 8, leftg= data[G]); | |
1378 put_bits(&s->pb, 8, leftb= data[B]); | |
1379 put_bits(&s->pb, 8, 0); | |
1380 | |
1381 sub_left_prediction_bgr32(s, s->temp[0], data+4, width-1, &leftr, &leftg, &leftb); | |
1382 encode_bgr_bitstream(s, width-1); | |
1383 | |
1384 for(y=1; y<s->height; y++){ | |
1385 uint8_t *dst = data + y*stride; | |
1386 if(s->predictor == PLANE && s->interlaced < y){ | |
1387 s->dsp.diff_bytes(s->temp[1], dst, dst - fake_stride, width*4); | |
1388 sub_left_prediction_bgr32(s, s->temp[0], s->temp[1], width, &leftr, &leftg, &leftb); | |
1389 }else{ | |
1390 sub_left_prediction_bgr32(s, s->temp[0], dst, width, &leftr, &leftg, &leftb); | |
1391 } | |
1392 encode_bgr_bitstream(s, width); | |
1393 } | |
866 | 1394 }else{ |
1598
932d306bf1dc
av_log() patch by (Michel Bardiaux <mbardiaux at peaktime dot be>)
michael
parents:
1529
diff
changeset
|
1395 av_log(avctx, AV_LOG_ERROR, "Format not supported!\n"); |
866 | 1396 } |
1397 emms_c(); | |
2967 | 1398 |
2369 | 1399 size+= (put_bits_count(&s->pb)+31)/8; |
9346
ee59f4f22dce
Make sure the last bits of huffyuv frames are zeroed.
michael
parents:
9007
diff
changeset
|
1400 put_bits(&s->pb, 16, 0); |
ee59f4f22dce
Make sure the last bits of huffyuv frames are zeroed.
michael
parents:
9007
diff
changeset
|
1401 put_bits(&s->pb, 15, 0); |
2369 | 1402 size/= 4; |
2967 | 1403 |
866 | 1404 if((s->flags&CODEC_FLAG_PASS1) && (s->picture_number&31)==0){ |
1405 int j; | |
1406 char *p= avctx->stats_out; | |
2423 | 1407 char *end= p + 1024*30; |
866 | 1408 for(i=0; i<3; i++){ |
1409 for(j=0; j<256; j++){ | |
2962 | 1410 snprintf(p, end-p, "%"PRIu64" ", s->stats[i][j]); |
866 | 1411 p+= strlen(p); |
1412 s->stats[i][j]= 0; | |
1413 } | |
2423 | 1414 snprintf(p, end-p, "\n"); |
866 | 1415 p++; |
1416 } | |
5025
65dd8127ca46
r3938 broke 2pass huffyuv (not that anyone uses it)
lorenm
parents:
4962
diff
changeset
|
1417 } else |
65dd8127ca46
r3938 broke 2pass huffyuv (not that anyone uses it)
lorenm
parents:
4962
diff
changeset
|
1418 avctx->stats_out[0] = '\0'; |
2501 | 1419 if(!(s->avctx->flags2 & CODEC_FLAG2_NO_OUTPUT)){ |
868
dfe7d66da6f0
YV12 support (warning this is experimental & wont work with offical huffyuv but there is a approx. 20% speed & compression gain)
michaelni
parents:
866
diff
changeset
|
1420 flush_put_bits(&s->pb); |
1273 | 1421 s->dsp.bswap_buf((uint32_t*)buf, (uint32_t*)buf, size); |
866 | 1422 } |
2967 | 1423 |
866 | 1424 s->picture_number++; |
903 | 1425 |
866 | 1426 return size*4; |
1427 } | |
1428 | |
6517
48759bfbd073
Apply 'cold' attribute to init/uninit functions in libavcodec
zuxy
parents:
6234
diff
changeset
|
1429 static av_cold int encode_end(AVCodecContext *avctx) |
866 | 1430 { |
2422 | 1431 HYuvContext *s = avctx->priv_data; |
2967 | 1432 |
2422 | 1433 common_end(s); |
866 | 1434 |
1435 av_freep(&avctx->extradata); | |
1436 av_freep(&avctx->stats_out); | |
2967 | 1437 |
866 | 1438 return 0; |
1439 } | |
8590 | 1440 #endif /* CONFIG_HUFFYUV_ENCODER || CONFIG_FFVHUFF_ENCODER */ |
866 | 1441 |
8590 | 1442 #if CONFIG_HUFFYUV_DECODER |
866 | 1443 AVCodec huffyuv_decoder = { |
1444 "huffyuv", | |
1445 CODEC_TYPE_VIDEO, | |
1446 CODEC_ID_HUFFYUV, | |
1447 sizeof(HYuvContext), | |
1448 decode_init, | |
1449 NULL, | |
1450 decode_end, | |
1451 decode_frame, | |
871 | 1452 CODEC_CAP_DR1 | CODEC_CAP_DRAW_HORIZ_BAND, |
6722 | 1453 NULL, |
7040
e943e1409077
Make AVCodec long_names definition conditional depending on CONFIG_SMALL.
stefano
parents:
6788
diff
changeset
|
1454 .long_name = NULL_IF_CONFIG_SMALL("Huffyuv / HuffYUV"), |
866 | 1455 }; |
7776
dbcdd0165e55
Replace generic CONFIG_DECODERS preprocessor conditionals by more specific
diego
parents:
7040
diff
changeset
|
1456 #endif |
866 | 1457 |
8590 | 1458 #if CONFIG_FFVHUFF_DECODER |
2373 | 1459 AVCodec ffvhuff_decoder = { |
1460 "ffvhuff", | |
1461 CODEC_TYPE_VIDEO, | |
1462 CODEC_ID_FFVHUFF, | |
1463 sizeof(HYuvContext), | |
1464 decode_init, | |
1465 NULL, | |
1466 decode_end, | |
1467 decode_frame, | |
1468 CODEC_CAP_DR1 | CODEC_CAP_DRAW_HORIZ_BAND, | |
6722 | 1469 NULL, |
7040
e943e1409077
Make AVCodec long_names definition conditional depending on CONFIG_SMALL.
stefano
parents:
6788
diff
changeset
|
1470 .long_name = NULL_IF_CONFIG_SMALL("Huffyuv FFmpeg variant"), |
2373 | 1471 }; |
3777 | 1472 #endif |
2373 | 1473 |
8590 | 1474 #if CONFIG_HUFFYUV_ENCODER |
866 | 1475 AVCodec huffyuv_encoder = { |
1476 "huffyuv", | |
1477 CODEC_TYPE_VIDEO, | |
1478 CODEC_ID_HUFFYUV, | |
1479 sizeof(HYuvContext), | |
1480 encode_init, | |
1481 encode_frame, | |
1482 encode_end, | |
6788 | 1483 .pix_fmts= (enum PixelFormat[]){PIX_FMT_YUV422P, PIX_FMT_RGB32, PIX_FMT_NONE}, |
7040
e943e1409077
Make AVCodec long_names definition conditional depending on CONFIG_SMALL.
stefano
parents:
6788
diff
changeset
|
1484 .long_name = NULL_IF_CONFIG_SMALL("Huffyuv / HuffYUV"), |
866 | 1485 }; |
7782
6efb15a24e91
Replace generic CONFIG_ENCODERS preprocessor conditionals by more specific
diego
parents:
7776
diff
changeset
|
1486 #endif |
1232
e88d3b1fb2a1
more #ifdef CONFIG_ENCODERS by (Wolfgang Hesseler <qv at multimediaware dot com>)
michaelni
parents:
1228
diff
changeset
|
1487 |
8590 | 1488 #if CONFIG_FFVHUFF_ENCODER |
2373 | 1489 AVCodec ffvhuff_encoder = { |
1490 "ffvhuff", | |
1491 CODEC_TYPE_VIDEO, | |
1492 CODEC_ID_FFVHUFF, | |
1493 sizeof(HYuvContext), | |
1494 encode_init, | |
1495 encode_frame, | |
1496 encode_end, | |
6788 | 1497 .pix_fmts= (enum PixelFormat[]){PIX_FMT_YUV420P, PIX_FMT_YUV422P, PIX_FMT_RGB32, PIX_FMT_NONE}, |
7040
e943e1409077
Make AVCodec long_names definition conditional depending on CONFIG_SMALL.
stefano
parents:
6788
diff
changeset
|
1498 .long_name = NULL_IF_CONFIG_SMALL("Huffyuv FFmpeg variant"), |
2373 | 1499 }; |
7782
6efb15a24e91
Replace generic CONFIG_ENCODERS preprocessor conditionals by more specific
diego
parents:
7776
diff
changeset
|
1500 #endif |