Mercurial > libavcodec.hg
annotate mpegvideo.c @ 112:a2c063b6ecf9 libavcodec
fixed a bug in the tmp buffer
fixed the color range for yuv
fixed the width %8!=0 bug (another 1% speed loss)
author | michael |
---|---|
date | Fri, 19 Oct 2001 13:41:38 +0000 |
parents | 1e4a4af694d1 |
children | de80712db90b |
rev | line source |
---|---|
0 | 1 /* |
2 * The simplest mpeg encoder (well, it was the simplest!) | |
3 * Copyright (c) 2000,2001 Gerard Lantau. | |
4 * | |
5 * This program is free software; you can redistribute it and/or modify | |
6 * it under the terms of the GNU General Public License as published by | |
7 * the Free Software Foundation; either version 2 of the License, or | |
8 * (at your option) any later version. | |
9 * | |
10 * This program is distributed in the hope that it will be useful, | |
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
13 * GNU General Public License for more details. | |
14 * | |
15 * You should have received a copy of the GNU General Public License | |
16 * along with this program; if not, write to the Free Software | |
17 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. | |
18 */ | |
19 #include <stdlib.h> | |
20 #include <stdio.h> | |
21 #include <math.h> | |
22 #include <string.h> | |
23 #include "avcodec.h" | |
24 #include "dsputil.h" | |
25 #include "mpegvideo.h" | |
26 | |
17
b69fe46fd708
Adding fastmemcpy stuff to speedup mplayer project
nickols_k
parents:
13
diff
changeset
|
27 #ifdef USE_FASTMEMCPY |
b69fe46fd708
Adding fastmemcpy stuff to speedup mplayer project
nickols_k
parents:
13
diff
changeset
|
28 #include "fastmemcpy.h" |
b69fe46fd708
Adding fastmemcpy stuff to speedup mplayer project
nickols_k
parents:
13
diff
changeset
|
29 #endif |
b69fe46fd708
Adding fastmemcpy stuff to speedup mplayer project
nickols_k
parents:
13
diff
changeset
|
30 |
13
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
31 static void encode_picture(MpegEncContext *s, int picture_number); |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
32 static void rate_control_init(MpegEncContext *s); |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
33 static int rate_estimate_qscale(MpegEncContext *s); |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
34 static void dct_unquantize_mpeg1_c(MpegEncContext *s, |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
35 DCTELEM *block, int n, int qscale); |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
36 static void dct_unquantize_h263_c(MpegEncContext *s, |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
37 DCTELEM *block, int n, int qscale); |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
38 static int dct_quantize(MpegEncContext *s, DCTELEM *block, int n, int qscale); |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
39 static int dct_quantize_mmx(MpegEncContext *s, |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
40 DCTELEM *block, int n, |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
41 int qscale); |
0 | 42 #define EDGE_WIDTH 16 |
43 | |
44 /* enable all paranoid tests for rounding, overflows, etc... */ | |
45 //#define PARANOID | |
46 | |
47 //#define DEBUG | |
48 | |
49 /* for jpeg fast DCT */ | |
50 #define CONST_BITS 14 | |
51 | |
52 static const unsigned short aanscales[64] = { | |
53 /* precomputed values scaled up by 14 bits */ | |
54 16384, 22725, 21407, 19266, 16384, 12873, 8867, 4520, | |
55 22725, 31521, 29692, 26722, 22725, 17855, 12299, 6270, | |
56 21407, 29692, 27969, 25172, 21407, 16819, 11585, 5906, | |
57 19266, 26722, 25172, 22654, 19266, 15137, 10426, 5315, | |
58 16384, 22725, 21407, 19266, 16384, 12873, 8867, 4520, | |
59 12873, 17855, 16819, 15137, 12873, 10114, 6967, 3552, | |
60 8867, 12299, 11585, 10426, 8867, 6967, 4799, 2446, | |
61 4520, 6270, 5906, 5315, 4520, 3552, 2446, 1247 | |
62 }; | |
63 | |
64 static UINT8 h263_chroma_roundtab[16] = { | |
65 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, | |
66 }; | |
67 | |
68 /* default motion estimation */ | |
69 int motion_estimation_method = ME_LOG; | |
70 | |
71 /* XXX: should use variable shift ? */ | |
72 #define QMAT_SHIFT_MMX 19 | |
73 #define QMAT_SHIFT 25 | |
74 | |
75 static void convert_matrix(int *qmat, const UINT16 *quant_matrix, int qscale) | |
76 { | |
77 int i; | |
78 | |
79 if (av_fdct == jpeg_fdct_ifast) { | |
80 for(i=0;i<64;i++) { | |
81 /* 16 <= qscale * quant_matrix[i] <= 7905 */ | |
82 /* 19952 <= aanscales[i] * qscale * quant_matrix[i] <= 249205026 */ | |
83 | |
64 | 84 qmat[i] = (int)((UINT64_C(1) << (QMAT_SHIFT + 11)) / |
85 (aanscales[i] * qscale * quant_matrix[i])); | |
0 | 86 } |
87 } else { | |
88 for(i=0;i<64;i++) { | |
89 /* We can safely suppose that 16 <= quant_matrix[i] <= 255 | |
90 So 16 <= qscale * quant_matrix[i] <= 7905 | |
91 so (1 << QMAT_SHIFT) / 16 >= qmat[i] >= (1 << QMAT_SHIFT) / 7905 | |
92 */ | |
93 qmat[i] = (1 << QMAT_SHIFT_MMX) / (qscale * quant_matrix[i]); | |
94 } | |
95 } | |
96 } | |
97 | |
98 /* init common structure for both encoder and decoder */ | |
99 int MPV_common_init(MpegEncContext *s) | |
100 { | |
101 int c_size, i; | |
102 UINT8 *pict; | |
103 | |
13
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
104 if (s->out_format == FMT_H263) |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
105 s->dct_unquantize = dct_unquantize_h263_c; |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
106 else |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
107 s->dct_unquantize = dct_unquantize_mpeg1_c; |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
108 |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
109 #ifdef HAVE_MMX |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
110 MPV_common_init_mmx(s); |
8 | 111 #endif |
0 | 112 s->mb_width = (s->width + 15) / 16; |
113 s->mb_height = (s->height + 15) / 16; | |
114 s->linesize = s->mb_width * 16 + 2 * EDGE_WIDTH; | |
115 | |
116 for(i=0;i<3;i++) { | |
117 int w, h, shift, pict_start; | |
118 | |
119 w = s->linesize; | |
120 h = s->mb_height * 16 + 2 * EDGE_WIDTH; | |
121 shift = (i == 0) ? 0 : 1; | |
122 c_size = (w >> shift) * (h >> shift); | |
123 pict_start = (w >> shift) * (EDGE_WIDTH >> shift) + (EDGE_WIDTH >> shift); | |
124 | |
125 pict = av_mallocz(c_size); | |
126 if (pict == NULL) | |
127 goto fail; | |
128 s->last_picture_base[i] = pict; | |
129 s->last_picture[i] = pict + pict_start; | |
130 | |
131 pict = av_mallocz(c_size); | |
132 if (pict == NULL) | |
133 goto fail; | |
134 s->next_picture_base[i] = pict; | |
135 s->next_picture[i] = pict + pict_start; | |
136 | |
137 if (s->has_b_frames) { | |
138 pict = av_mallocz(c_size); | |
139 if (pict == NULL) | |
140 goto fail; | |
141 s->aux_picture_base[i] = pict; | |
142 s->aux_picture[i] = pict + pict_start; | |
143 } | |
144 } | |
145 | |
146 if (s->out_format == FMT_H263) { | |
147 int size; | |
148 /* MV prediction */ | |
149 size = (2 * s->mb_width + 2) * (2 * s->mb_height + 2); | |
150 s->motion_val = malloc(size * 2 * sizeof(INT16)); | |
151 if (s->motion_val == NULL) | |
152 goto fail; | |
153 memset(s->motion_val, 0, size * 2 * sizeof(INT16)); | |
154 } | |
155 | |
156 if (s->h263_pred) { | |
157 int y_size, c_size, i, size; | |
158 | |
159 /* dc values */ | |
160 | |
161 y_size = (2 * s->mb_width + 2) * (2 * s->mb_height + 2); | |
162 c_size = (s->mb_width + 2) * (s->mb_height + 2); | |
163 size = y_size + 2 * c_size; | |
164 s->dc_val[0] = malloc(size * sizeof(INT16)); | |
165 if (s->dc_val[0] == NULL) | |
166 goto fail; | |
167 s->dc_val[1] = s->dc_val[0] + y_size; | |
168 s->dc_val[2] = s->dc_val[1] + c_size; | |
169 for(i=0;i<size;i++) | |
170 s->dc_val[0][i] = 1024; | |
171 | |
172 /* ac values */ | |
173 s->ac_val[0] = av_mallocz(size * sizeof(INT16) * 16); | |
174 if (s->ac_val[0] == NULL) | |
175 goto fail; | |
176 s->ac_val[1] = s->ac_val[0] + y_size; | |
177 s->ac_val[2] = s->ac_val[1] + c_size; | |
178 | |
179 /* cbp values */ | |
180 s->coded_block = av_mallocz(y_size); | |
181 if (!s->coded_block) | |
182 goto fail; | |
183 } | |
184 /* default structure is frame */ | |
185 s->picture_structure = PICT_FRAME; | |
186 | |
7
1d3ac9654178
added skip macroblock optimization (big perf win on black regions for example)
glantau
parents:
0
diff
changeset
|
187 /* init macroblock skip table */ |
1d3ac9654178
added skip macroblock optimization (big perf win on black regions for example)
glantau
parents:
0
diff
changeset
|
188 if (!s->encoding) { |
1d3ac9654178
added skip macroblock optimization (big perf win on black regions for example)
glantau
parents:
0
diff
changeset
|
189 s->mbskip_table = av_mallocz(s->mb_width * s->mb_height); |
1d3ac9654178
added skip macroblock optimization (big perf win on black regions for example)
glantau
parents:
0
diff
changeset
|
190 if (!s->mbskip_table) |
1d3ac9654178
added skip macroblock optimization (big perf win on black regions for example)
glantau
parents:
0
diff
changeset
|
191 goto fail; |
1d3ac9654178
added skip macroblock optimization (big perf win on black regions for example)
glantau
parents:
0
diff
changeset
|
192 } |
1d3ac9654178
added skip macroblock optimization (big perf win on black regions for example)
glantau
parents:
0
diff
changeset
|
193 |
0 | 194 s->context_initialized = 1; |
195 return 0; | |
196 fail: | |
197 if (s->motion_val) | |
198 free(s->motion_val); | |
199 if (s->dc_val[0]) | |
200 free(s->dc_val[0]); | |
201 if (s->ac_val[0]) | |
202 free(s->ac_val[0]); | |
203 if (s->coded_block) | |
204 free(s->coded_block); | |
7
1d3ac9654178
added skip macroblock optimization (big perf win on black regions for example)
glantau
parents:
0
diff
changeset
|
205 if (s->mbskip_table) |
1d3ac9654178
added skip macroblock optimization (big perf win on black regions for example)
glantau
parents:
0
diff
changeset
|
206 free(s->mbskip_table); |
0 | 207 for(i=0;i<3;i++) { |
208 if (s->last_picture_base[i]) | |
209 free(s->last_picture_base[i]); | |
210 if (s->next_picture_base[i]) | |
211 free(s->next_picture_base[i]); | |
212 if (s->aux_picture_base[i]) | |
213 free(s->aux_picture_base[i]); | |
214 } | |
215 return -1; | |
216 } | |
217 | |
218 /* init common structure for both encoder and decoder */ | |
219 void MPV_common_end(MpegEncContext *s) | |
220 { | |
221 int i; | |
222 | |
223 if (s->motion_val) | |
224 free(s->motion_val); | |
225 if (s->h263_pred) { | |
226 free(s->dc_val[0]); | |
227 free(s->ac_val[0]); | |
228 free(s->coded_block); | |
229 } | |
7
1d3ac9654178
added skip macroblock optimization (big perf win on black regions for example)
glantau
parents:
0
diff
changeset
|
230 if (s->mbskip_table) |
1d3ac9654178
added skip macroblock optimization (big perf win on black regions for example)
glantau
parents:
0
diff
changeset
|
231 free(s->mbskip_table); |
0 | 232 for(i=0;i<3;i++) { |
233 free(s->last_picture_base[i]); | |
234 free(s->next_picture_base[i]); | |
235 if (s->has_b_frames) | |
236 free(s->aux_picture_base[i]); | |
237 } | |
238 s->context_initialized = 0; | |
239 } | |
240 | |
241 /* init video encoder */ | |
242 int MPV_encode_init(AVCodecContext *avctx) | |
243 { | |
244 MpegEncContext *s = avctx->priv_data; | |
60 | 245 int i; |
0 | 246 |
247 s->bit_rate = avctx->bit_rate; | |
248 s->frame_rate = avctx->frame_rate; | |
249 s->width = avctx->width; | |
250 s->height = avctx->height; | |
251 s->gop_size = avctx->gop_size; | |
252 if (s->gop_size <= 1) { | |
253 s->intra_only = 1; | |
254 s->gop_size = 12; | |
255 } else { | |
256 s->intra_only = 0; | |
257 } | |
258 s->full_search = motion_estimation_method; | |
259 | |
260 s->fixed_qscale = (avctx->flags & CODEC_FLAG_QSCALE); | |
261 | |
262 switch(avctx->codec->id) { | |
263 case CODEC_ID_MPEG1VIDEO: | |
264 s->out_format = FMT_MPEG1; | |
265 break; | |
266 case CODEC_ID_MJPEG: | |
267 s->out_format = FMT_MJPEG; | |
268 s->intra_only = 1; /* force intra only for jpeg */ | |
269 if (mjpeg_init(s) < 0) | |
270 return -1; | |
271 break; | |
272 case CODEC_ID_H263: | |
273 if (h263_get_picture_format(s->width, s->height) == 7) | |
274 return -1; | |
275 s->out_format = FMT_H263; | |
276 break; | |
277 case CODEC_ID_H263P: | |
278 s->out_format = FMT_H263; | |
279 s->h263_plus = 1; | |
78 | 280 s->unrestricted_mv = 1; |
79
82e579c37bc3
Moved some H.263+ variables to MpegEncContext to be thread-safe.
pulento
parents:
78
diff
changeset
|
281 |
82e579c37bc3
Moved some H.263+ variables to MpegEncContext to be thread-safe.
pulento
parents:
78
diff
changeset
|
282 /* These are just to be sure */ |
82e579c37bc3
Moved some H.263+ variables to MpegEncContext to be thread-safe.
pulento
parents:
78
diff
changeset
|
283 s->umvplus = 0; |
82e579c37bc3
Moved some H.263+ variables to MpegEncContext to be thread-safe.
pulento
parents:
78
diff
changeset
|
284 s->umvplus_dec = 0; |
0 | 285 break; |
286 case CODEC_ID_RV10: | |
287 s->out_format = FMT_H263; | |
288 s->h263_rv10 = 1; | |
289 break; | |
71 | 290 case CODEC_ID_MPEG4: |
0 | 291 s->out_format = FMT_H263; |
292 s->h263_pred = 1; | |
293 s->unrestricted_mv = 1; | |
294 break; | |
295 case CODEC_ID_MSMPEG4: | |
296 s->out_format = FMT_H263; | |
297 s->h263_msmpeg4 = 1; | |
298 s->h263_pred = 1; | |
299 s->unrestricted_mv = 1; | |
300 break; | |
301 default: | |
302 return -1; | |
303 } | |
304 | |
305 if (s->out_format == FMT_H263) | |
306 h263_encode_init_vlc(s); | |
307 | |
7
1d3ac9654178
added skip macroblock optimization (big perf win on black regions for example)
glantau
parents:
0
diff
changeset
|
308 s->encoding = 1; |
1d3ac9654178
added skip macroblock optimization (big perf win on black regions for example)
glantau
parents:
0
diff
changeset
|
309 |
0 | 310 /* init */ |
311 if (MPV_common_init(s) < 0) | |
312 return -1; | |
313 | |
60 | 314 /* init default q matrix */ |
315 for(i=0;i<64;i++) { | |
316 s->intra_matrix[i] = default_intra_matrix[i]; | |
317 s->non_intra_matrix[i] = default_non_intra_matrix[i]; | |
318 } | |
319 | |
0 | 320 /* rate control init */ |
321 rate_control_init(s); | |
322 | |
323 s->picture_number = 0; | |
324 s->fake_picture_number = 0; | |
325 /* motion detector init */ | |
326 s->f_code = 1; | |
327 | |
328 return 0; | |
329 } | |
330 | |
331 int MPV_encode_end(AVCodecContext *avctx) | |
332 { | |
333 MpegEncContext *s = avctx->priv_data; | |
334 | |
335 #ifdef STATS | |
336 print_stats(); | |
337 #endif | |
338 MPV_common_end(s); | |
339 if (s->out_format == FMT_MJPEG) | |
340 mjpeg_close(s); | |
341 return 0; | |
342 } | |
343 | |
344 /* draw the edges of width 'w' of an image of size width, height */ | |
345 static void draw_edges(UINT8 *buf, int wrap, int width, int height, int w) | |
346 { | |
347 UINT8 *ptr, *last_line; | |
348 int i; | |
349 | |
350 last_line = buf + (height - 1) * wrap; | |
351 for(i=0;i<w;i++) { | |
352 /* top and bottom */ | |
353 memcpy(buf - (i + 1) * wrap, buf, width); | |
354 memcpy(last_line + (i + 1) * wrap, last_line, width); | |
355 } | |
356 /* left and right */ | |
357 ptr = buf; | |
358 for(i=0;i<height;i++) { | |
359 memset(ptr - w, ptr[0], w); | |
360 memset(ptr + width, ptr[width-1], w); | |
361 ptr += wrap; | |
362 } | |
363 /* corners */ | |
364 for(i=0;i<w;i++) { | |
365 memset(buf - (i + 1) * wrap - w, buf[0], w); /* top left */ | |
366 memset(buf - (i + 1) * wrap + width, buf[width-1], w); /* top right */ | |
367 memset(last_line + (i + 1) * wrap - w, last_line[0], w); /* top left */ | |
368 memset(last_line + (i + 1) * wrap + width, last_line[width-1], w); /* top right */ | |
369 } | |
370 } | |
371 | |
372 /* generic function for encode/decode called before a frame is coded/decoded */ | |
373 void MPV_frame_start(MpegEncContext *s) | |
374 { | |
375 int i; | |
376 UINT8 *tmp; | |
377 | |
46
931417475f5b
fixed mpeg1 first block bug (pb with black picture optimisation for B frames)
glantau
parents:
40
diff
changeset
|
378 s->mb_skiped = 0; |
0 | 379 if (s->pict_type == B_TYPE) { |
380 for(i=0;i<3;i++) { | |
381 s->current_picture[i] = s->aux_picture[i]; | |
382 } | |
383 } else { | |
384 for(i=0;i<3;i++) { | |
385 /* swap next and last */ | |
386 tmp = s->last_picture[i]; | |
387 s->last_picture[i] = s->next_picture[i]; | |
388 s->next_picture[i] = tmp; | |
389 s->current_picture[i] = tmp; | |
390 } | |
391 } | |
392 } | |
13
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
393 |
0 | 394 /* generic function for encode/decode called after a frame has been coded/decoded */ |
395 void MPV_frame_end(MpegEncContext *s) | |
396 { | |
397 /* draw edge for correct motion prediction if outside */ | |
398 if (s->pict_type != B_TYPE) { | |
399 draw_edges(s->current_picture[0], s->linesize, s->width, s->height, EDGE_WIDTH); | |
400 draw_edges(s->current_picture[1], s->linesize/2, s->width/2, s->height/2, EDGE_WIDTH/2); | |
401 draw_edges(s->current_picture[2], s->linesize/2, s->width/2, s->height/2, EDGE_WIDTH/2); | |
402 } | |
403 } | |
404 | |
405 int MPV_encode_picture(AVCodecContext *avctx, | |
406 unsigned char *buf, int buf_size, void *data) | |
407 { | |
408 MpegEncContext *s = avctx->priv_data; | |
409 AVPicture *pict = data; | |
410 int i, j; | |
411 | |
412 if (s->fixed_qscale) | |
413 s->qscale = avctx->quality; | |
414 | |
415 init_put_bits(&s->pb, buf, buf_size, NULL, NULL); | |
416 | |
417 if (!s->intra_only) { | |
418 /* first picture of GOP is intra */ | |
419 if ((s->picture_number % s->gop_size) == 0) | |
420 s->pict_type = I_TYPE; | |
421 else | |
422 s->pict_type = P_TYPE; | |
423 } else { | |
424 s->pict_type = I_TYPE; | |
425 } | |
426 avctx->key_frame = (s->pict_type == I_TYPE); | |
427 | |
428 MPV_frame_start(s); | |
429 | |
430 for(i=0;i<3;i++) { | |
431 UINT8 *src = pict->data[i]; | |
432 UINT8 *dest = s->current_picture[i]; | |
433 int src_wrap = pict->linesize[i]; | |
434 int dest_wrap = s->linesize; | |
435 int w = s->width; | |
436 int h = s->height; | |
437 | |
438 if (i >= 1) { | |
439 dest_wrap >>= 1; | |
440 w >>= 1; | |
441 h >>= 1; | |
442 } | |
443 | |
444 for(j=0;j<h;j++) { | |
445 memcpy(dest, src, w); | |
446 dest += dest_wrap; | |
447 src += src_wrap; | |
448 } | |
449 s->new_picture[i] = s->current_picture[i]; | |
450 } | |
451 | |
452 encode_picture(s, s->picture_number); | |
453 | |
454 MPV_frame_end(s); | |
455 s->picture_number++; | |
456 | |
457 if (s->out_format == FMT_MJPEG) | |
458 mjpeg_picture_trailer(s); | |
459 | |
460 flush_put_bits(&s->pb); | |
461 s->total_bits += (s->pb.buf_ptr - s->pb.buf) * 8; | |
462 avctx->quality = s->qscale; | |
463 return s->pb.buf_ptr - s->pb.buf; | |
464 } | |
465 | |
466 static inline int clip(int a, int amin, int amax) | |
467 { | |
468 if (a < amin) | |
469 return amin; | |
470 else if (a > amax) | |
471 return amax; | |
472 else | |
473 return a; | |
474 } | |
475 | |
476 /* apply one mpeg motion vector to the three components */ | |
477 static inline void mpeg_motion(MpegEncContext *s, | |
478 UINT8 *dest_y, UINT8 *dest_cb, UINT8 *dest_cr, | |
479 int dest_offset, | |
480 UINT8 **ref_picture, int src_offset, | |
481 int field_based, op_pixels_func *pix_op, | |
482 int motion_x, int motion_y, int h) | |
483 { | |
484 UINT8 *ptr; | |
485 int dxy, offset, mx, my, src_x, src_y, height, linesize; | |
486 | |
487 dxy = ((motion_y & 1) << 1) | (motion_x & 1); | |
488 src_x = s->mb_x * 16 + (motion_x >> 1); | |
489 src_y = s->mb_y * (16 >> field_based) + (motion_y >> 1); | |
490 | |
491 /* WARNING: do no forget half pels */ | |
492 height = s->height >> field_based; | |
493 src_x = clip(src_x, -16, s->width); | |
494 if (src_x == s->width) | |
495 dxy &= ~1; | |
496 src_y = clip(src_y, -16, height); | |
497 if (src_y == height) | |
498 dxy &= ~2; | |
499 linesize = s->linesize << field_based; | |
500 ptr = ref_picture[0] + (src_y * linesize) + (src_x) + src_offset; | |
501 dest_y += dest_offset; | |
502 pix_op[dxy](dest_y, ptr, linesize, h); | |
503 pix_op[dxy](dest_y + 8, ptr + 8, linesize, h); | |
504 | |
505 if (s->out_format == FMT_H263) { | |
506 dxy = 0; | |
507 if ((motion_x & 3) != 0) | |
508 dxy |= 1; | |
509 if ((motion_y & 3) != 0) | |
510 dxy |= 2; | |
511 mx = motion_x >> 2; | |
512 my = motion_y >> 2; | |
513 } else { | |
514 mx = motion_x / 2; | |
515 my = motion_y / 2; | |
516 dxy = ((my & 1) << 1) | (mx & 1); | |
517 mx >>= 1; | |
518 my >>= 1; | |
519 } | |
520 | |
521 src_x = s->mb_x * 8 + mx; | |
522 src_y = s->mb_y * (8 >> field_based) + my; | |
523 src_x = clip(src_x, -8, s->width >> 1); | |
524 if (src_x == (s->width >> 1)) | |
525 dxy &= ~1; | |
526 src_y = clip(src_y, -8, height >> 1); | |
527 if (src_y == (height >> 1)) | |
528 dxy &= ~2; | |
529 | |
530 offset = (src_y * (linesize >> 1)) + src_x + (src_offset >> 1); | |
531 ptr = ref_picture[1] + offset; | |
532 pix_op[dxy](dest_cb + (dest_offset >> 1), ptr, linesize >> 1, h >> 1); | |
533 ptr = ref_picture[2] + offset; | |
534 pix_op[dxy](dest_cr + (dest_offset >> 1), ptr, linesize >> 1, h >> 1); | |
535 } | |
536 | |
537 static inline void MPV_motion(MpegEncContext *s, | |
538 UINT8 *dest_y, UINT8 *dest_cb, UINT8 *dest_cr, | |
539 int dir, UINT8 **ref_picture, | |
540 op_pixels_func *pix_op) | |
541 { | |
542 int dxy, offset, mx, my, src_x, src_y, motion_x, motion_y; | |
543 int mb_x, mb_y, i; | |
544 UINT8 *ptr, *dest; | |
545 | |
546 mb_x = s->mb_x; | |
547 mb_y = s->mb_y; | |
548 | |
549 switch(s->mv_type) { | |
550 case MV_TYPE_16X16: | |
551 mpeg_motion(s, dest_y, dest_cb, dest_cr, 0, | |
552 ref_picture, 0, | |
553 0, pix_op, | |
554 s->mv[dir][0][0], s->mv[dir][0][1], 16); | |
555 break; | |
556 case MV_TYPE_8X8: | |
557 for(i=0;i<4;i++) { | |
558 motion_x = s->mv[dir][i][0]; | |
559 motion_y = s->mv[dir][i][1]; | |
560 | |
561 dxy = ((motion_y & 1) << 1) | (motion_x & 1); | |
562 src_x = mb_x * 16 + (motion_x >> 1) + (i & 1) * 8; | |
563 src_y = mb_y * 16 + (motion_y >> 1) + ((i >> 1) & 1) * 8; | |
564 | |
565 /* WARNING: do no forget half pels */ | |
566 src_x = clip(src_x, -16, s->width); | |
567 if (src_x == s->width) | |
568 dxy &= ~1; | |
569 src_y = clip(src_y, -16, s->height); | |
570 if (src_y == s->height) | |
571 dxy &= ~2; | |
572 | |
573 ptr = ref_picture[0] + (src_y * s->linesize) + (src_x); | |
574 dest = dest_y + ((i & 1) * 8) + (i >> 1) * 8 * s->linesize; | |
575 pix_op[dxy](dest, ptr, s->linesize, 8); | |
576 } | |
577 /* In case of 8X8, we construct a single chroma motion vector | |
578 with a special rounding */ | |
579 mx = 0; | |
580 my = 0; | |
581 for(i=0;i<4;i++) { | |
582 mx += s->mv[dir][i][0]; | |
583 my += s->mv[dir][i][1]; | |
584 } | |
585 if (mx >= 0) | |
586 mx = (h263_chroma_roundtab[mx & 0xf] + ((mx >> 3) & ~1)); | |
587 else { | |
588 mx = -mx; | |
589 mx = -(h263_chroma_roundtab[mx & 0xf] + ((mx >> 3) & ~1)); | |
590 } | |
591 if (my >= 0) | |
592 my = (h263_chroma_roundtab[my & 0xf] + ((my >> 3) & ~1)); | |
593 else { | |
594 my = -my; | |
595 my = -(h263_chroma_roundtab[my & 0xf] + ((my >> 3) & ~1)); | |
596 } | |
597 dxy = ((my & 1) << 1) | (mx & 1); | |
598 mx >>= 1; | |
599 my >>= 1; | |
600 | |
601 src_x = mb_x * 8 + mx; | |
602 src_y = mb_y * 8 + my; | |
603 src_x = clip(src_x, -8, s->width/2); | |
604 if (src_x == s->width/2) | |
605 dxy &= ~1; | |
606 src_y = clip(src_y, -8, s->height/2); | |
607 if (src_y == s->height/2) | |
608 dxy &= ~2; | |
609 | |
610 offset = (src_y * (s->linesize >> 1)) + src_x; | |
611 ptr = ref_picture[1] + offset; | |
612 pix_op[dxy](dest_cb, ptr, s->linesize >> 1, 8); | |
613 ptr = ref_picture[2] + offset; | |
614 pix_op[dxy](dest_cr, ptr, s->linesize >> 1, 8); | |
615 break; | |
616 case MV_TYPE_FIELD: | |
617 if (s->picture_structure == PICT_FRAME) { | |
618 /* top field */ | |
619 mpeg_motion(s, dest_y, dest_cb, dest_cr, 0, | |
620 ref_picture, s->field_select[dir][0] ? s->linesize : 0, | |
621 1, pix_op, | |
622 s->mv[dir][0][0], s->mv[dir][0][1], 8); | |
623 /* bottom field */ | |
624 mpeg_motion(s, dest_y, dest_cb, dest_cr, s->linesize, | |
625 ref_picture, s->field_select[dir][1] ? s->linesize : 0, | |
626 1, pix_op, | |
627 s->mv[dir][1][0], s->mv[dir][1][1], 8); | |
628 } else { | |
629 | |
630 | |
631 } | |
632 break; | |
633 } | |
634 } | |
635 | |
636 | |
637 /* put block[] to dest[] */ | |
638 static inline void put_dct(MpegEncContext *s, | |
639 DCTELEM *block, int i, UINT8 *dest, int line_size) | |
640 { | |
641 if (!s->mpeg2) | |
13
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
642 s->dct_unquantize(s, block, i, s->qscale); |
19
82d4c9be9873
MMX/MMXEXT iDCT support, using external functions currently defined in libmpeg2
arpi_esp
parents:
18
diff
changeset
|
643 ff_idct (block); |
0 | 644 put_pixels_clamped(block, dest, line_size); |
645 } | |
646 | |
647 /* add block[] to dest[] */ | |
648 static inline void add_dct(MpegEncContext *s, | |
649 DCTELEM *block, int i, UINT8 *dest, int line_size) | |
650 { | |
651 if (s->block_last_index[i] >= 0) { | |
652 if (!s->mpeg2) | |
13
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
653 s->dct_unquantize(s, block, i, s->qscale); |
19
82d4c9be9873
MMX/MMXEXT iDCT support, using external functions currently defined in libmpeg2
arpi_esp
parents:
18
diff
changeset
|
654 ff_idct (block); |
0 | 655 add_pixels_clamped(block, dest, line_size); |
656 } | |
657 } | |
658 | |
659 /* generic function called after a macroblock has been parsed by the | |
660 decoder or after it has been encoded by the encoder. | |
661 | |
662 Important variables used: | |
663 s->mb_intra : true if intra macroblock | |
664 s->mv_dir : motion vector direction | |
665 s->mv_type : motion vector type | |
666 s->mv : motion vector | |
667 s->interlaced_dct : true if interlaced dct used (mpeg2) | |
668 */ | |
669 void MPV_decode_mb(MpegEncContext *s, DCTELEM block[6][64]) | |
670 { | |
671 int mb_x, mb_y, motion_x, motion_y; | |
672 int dct_linesize, dct_offset; | |
673 op_pixels_func *op_pix; | |
674 | |
675 mb_x = s->mb_x; | |
676 mb_y = s->mb_y; | |
677 | |
108
1e4a4af694d1
exporting qscale data for postprocessing (for MPlayer)
arpi_esp
parents:
79
diff
changeset
|
678 #ifdef FF_POSTPROCESS |
1e4a4af694d1
exporting qscale data for postprocessing (for MPlayer)
arpi_esp
parents:
79
diff
changeset
|
679 quant_store[mb_y][mb_x]=s->qscale; |
1e4a4af694d1
exporting qscale data for postprocessing (for MPlayer)
arpi_esp
parents:
79
diff
changeset
|
680 //printf("[%02d][%02d] %d\n",mb_x,mb_y,s->qscale); |
1e4a4af694d1
exporting qscale data for postprocessing (for MPlayer)
arpi_esp
parents:
79
diff
changeset
|
681 #endif |
1e4a4af694d1
exporting qscale data for postprocessing (for MPlayer)
arpi_esp
parents:
79
diff
changeset
|
682 |
0 | 683 /* update DC predictors for P macroblocks */ |
684 if (!s->mb_intra) { | |
685 if (s->h263_pred) { | |
686 int wrap, x, y, v; | |
687 wrap = 2 * s->mb_width + 2; | |
688 v = 1024; | |
689 x = 2 * mb_x + 1; | |
690 y = 2 * mb_y + 1; | |
691 s->dc_val[0][(x) + (y) * wrap] = v; | |
692 s->dc_val[0][(x + 1) + (y) * wrap] = v; | |
693 s->dc_val[0][(x) + (y + 1) * wrap] = v; | |
694 s->dc_val[0][(x + 1) + (y + 1) * wrap] = v; | |
695 /* ac pred */ | |
696 memset(s->ac_val[0][(x) + (y) * wrap], 0, 16 * sizeof(INT16)); | |
697 memset(s->ac_val[0][(x + 1) + (y) * wrap], 0, 16 * sizeof(INT16)); | |
698 memset(s->ac_val[0][(x) + (y + 1) * wrap], 0, 16 * sizeof(INT16)); | |
699 memset(s->ac_val[0][(x + 1) + (y + 1) * wrap], 0, 16 * sizeof(INT16)); | |
700 if (s->h263_msmpeg4) { | |
701 s->coded_block[(x) + (y) * wrap] = 0; | |
702 s->coded_block[(x + 1) + (y) * wrap] = 0; | |
703 s->coded_block[(x) + (y + 1) * wrap] = 0; | |
704 s->coded_block[(x + 1) + (y + 1) * wrap] = 0; | |
705 } | |
706 /* chroma */ | |
707 wrap = s->mb_width + 2; | |
708 x = mb_x + 1; | |
709 y = mb_y + 1; | |
710 s->dc_val[1][(x) + (y) * wrap] = v; | |
711 s->dc_val[2][(x) + (y) * wrap] = v; | |
712 /* ac pred */ | |
713 memset(s->ac_val[1][(x) + (y) * wrap], 0, 16 * sizeof(INT16)); | |
714 memset(s->ac_val[2][(x) + (y) * wrap], 0, 16 * sizeof(INT16)); | |
715 } else { | |
716 s->last_dc[0] = 128 << s->intra_dc_precision; | |
717 s->last_dc[1] = 128 << s->intra_dc_precision; | |
718 s->last_dc[2] = 128 << s->intra_dc_precision; | |
719 } | |
720 } | |
721 | |
722 /* update motion predictor */ | |
723 if (s->out_format == FMT_H263) { | |
724 int x, y, wrap; | |
725 | |
726 x = 2 * mb_x + 1; | |
727 y = 2 * mb_y + 1; | |
728 wrap = 2 * s->mb_width + 2; | |
729 if (s->mb_intra) { | |
730 motion_x = 0; | |
731 motion_y = 0; | |
732 goto motion_init; | |
733 } else if (s->mv_type == MV_TYPE_16X16) { | |
734 motion_x = s->mv[0][0][0]; | |
735 motion_y = s->mv[0][0][1]; | |
736 motion_init: | |
737 /* no update if 8X8 because it has been done during parsing */ | |
738 s->motion_val[(x) + (y) * wrap][0] = motion_x; | |
739 s->motion_val[(x) + (y) * wrap][1] = motion_y; | |
740 s->motion_val[(x + 1) + (y) * wrap][0] = motion_x; | |
741 s->motion_val[(x + 1) + (y) * wrap][1] = motion_y; | |
742 s->motion_val[(x) + (y + 1) * wrap][0] = motion_x; | |
743 s->motion_val[(x) + (y + 1) * wrap][1] = motion_y; | |
744 s->motion_val[(x + 1) + (y + 1) * wrap][0] = motion_x; | |
745 s->motion_val[(x + 1) + (y + 1) * wrap][1] = motion_y; | |
746 } | |
747 } | |
748 | |
749 if (!s->intra_only) { | |
750 UINT8 *dest_y, *dest_cb, *dest_cr; | |
7
1d3ac9654178
added skip macroblock optimization (big perf win on black regions for example)
glantau
parents:
0
diff
changeset
|
751 UINT8 *mbskip_ptr; |
1d3ac9654178
added skip macroblock optimization (big perf win on black regions for example)
glantau
parents:
0
diff
changeset
|
752 |
1d3ac9654178
added skip macroblock optimization (big perf win on black regions for example)
glantau
parents:
0
diff
changeset
|
753 /* avoid copy if macroblock skipped in last frame too */ |
13
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
754 if (!s->encoding && s->pict_type != B_TYPE) { |
7
1d3ac9654178
added skip macroblock optimization (big perf win on black regions for example)
glantau
parents:
0
diff
changeset
|
755 mbskip_ptr = &s->mbskip_table[s->mb_y * s->mb_width + s->mb_x]; |
1d3ac9654178
added skip macroblock optimization (big perf win on black regions for example)
glantau
parents:
0
diff
changeset
|
756 if (s->mb_skiped) { |
1d3ac9654178
added skip macroblock optimization (big perf win on black regions for example)
glantau
parents:
0
diff
changeset
|
757 s->mb_skiped = 0; |
1d3ac9654178
added skip macroblock optimization (big perf win on black regions for example)
glantau
parents:
0
diff
changeset
|
758 /* if previous was skipped too, then nothing to do ! */ |
1d3ac9654178
added skip macroblock optimization (big perf win on black regions for example)
glantau
parents:
0
diff
changeset
|
759 if (*mbskip_ptr != 0) |
1d3ac9654178
added skip macroblock optimization (big perf win on black regions for example)
glantau
parents:
0
diff
changeset
|
760 goto the_end; |
1d3ac9654178
added skip macroblock optimization (big perf win on black regions for example)
glantau
parents:
0
diff
changeset
|
761 *mbskip_ptr = 1; /* indicate that this time we skiped it */ |
1d3ac9654178
added skip macroblock optimization (big perf win on black regions for example)
glantau
parents:
0
diff
changeset
|
762 } else { |
1d3ac9654178
added skip macroblock optimization (big perf win on black regions for example)
glantau
parents:
0
diff
changeset
|
763 *mbskip_ptr = 0; /* not skipped */ |
1d3ac9654178
added skip macroblock optimization (big perf win on black regions for example)
glantau
parents:
0
diff
changeset
|
764 } |
1d3ac9654178
added skip macroblock optimization (big perf win on black regions for example)
glantau
parents:
0
diff
changeset
|
765 } |
0 | 766 |
767 dest_y = s->current_picture[0] + (mb_y * 16 * s->linesize) + mb_x * 16; | |
768 dest_cb = s->current_picture[1] + (mb_y * 8 * (s->linesize >> 1)) + mb_x * 8; | |
769 dest_cr = s->current_picture[2] + (mb_y * 8 * (s->linesize >> 1)) + mb_x * 8; | |
770 | |
771 if (s->interlaced_dct) { | |
772 dct_linesize = s->linesize * 2; | |
773 dct_offset = s->linesize; | |
774 } else { | |
775 dct_linesize = s->linesize; | |
776 dct_offset = s->linesize * 8; | |
777 } | |
778 | |
779 if (!s->mb_intra) { | |
780 /* motion handling */ | |
781 if (!s->no_rounding) | |
782 op_pix = put_pixels_tab; | |
783 else | |
784 op_pix = put_no_rnd_pixels_tab; | |
785 | |
786 if (s->mv_dir & MV_DIR_FORWARD) { | |
787 MPV_motion(s, dest_y, dest_cb, dest_cr, 0, s->last_picture, op_pix); | |
788 if (!s->no_rounding) | |
789 op_pix = avg_pixels_tab; | |
790 else | |
791 op_pix = avg_no_rnd_pixels_tab; | |
792 } | |
793 if (s->mv_dir & MV_DIR_BACKWARD) { | |
794 MPV_motion(s, dest_y, dest_cb, dest_cr, 1, s->next_picture, op_pix); | |
795 } | |
796 | |
797 /* add dct residue */ | |
798 add_dct(s, block[0], 0, dest_y, dct_linesize); | |
799 add_dct(s, block[1], 1, dest_y + 8, dct_linesize); | |
800 add_dct(s, block[2], 2, dest_y + dct_offset, dct_linesize); | |
801 add_dct(s, block[3], 3, dest_y + dct_offset + 8, dct_linesize); | |
802 | |
57 | 803 add_dct(s, block[4], 4, dest_cb, s->linesize >> 1); |
804 add_dct(s, block[5], 5, dest_cr, s->linesize >> 1); | |
0 | 805 } else { |
806 /* dct only in intra block */ | |
807 put_dct(s, block[0], 0, dest_y, dct_linesize); | |
808 put_dct(s, block[1], 1, dest_y + 8, dct_linesize); | |
809 put_dct(s, block[2], 2, dest_y + dct_offset, dct_linesize); | |
810 put_dct(s, block[3], 3, dest_y + dct_offset + 8, dct_linesize); | |
811 | |
57 | 812 put_dct(s, block[4], 4, dest_cb, s->linesize >> 1); |
813 put_dct(s, block[5], 5, dest_cr, s->linesize >> 1); | |
0 | 814 } |
815 } | |
7
1d3ac9654178
added skip macroblock optimization (big perf win on black regions for example)
glantau
parents:
0
diff
changeset
|
816 the_end: |
1d3ac9654178
added skip macroblock optimization (big perf win on black regions for example)
glantau
parents:
0
diff
changeset
|
817 emms_c(); |
0 | 818 } |
819 | |
820 static void encode_picture(MpegEncContext *s, int picture_number) | |
821 { | |
822 int mb_x, mb_y, wrap; | |
823 UINT8 *ptr; | |
824 int i, motion_x, motion_y; | |
825 | |
826 s->picture_number = picture_number; | |
827 if (!s->fixed_qscale) | |
828 s->qscale = rate_estimate_qscale(s); | |
829 | |
830 /* precompute matrix */ | |
831 if (s->out_format == FMT_MJPEG) { | |
832 /* for mjpeg, we do include qscale in the matrix */ | |
833 s->intra_matrix[0] = default_intra_matrix[0]; | |
834 for(i=1;i<64;i++) | |
835 s->intra_matrix[i] = (default_intra_matrix[i] * s->qscale) >> 3; | |
836 convert_matrix(s->q_intra_matrix, s->intra_matrix, 8); | |
837 } else { | |
838 convert_matrix(s->q_intra_matrix, s->intra_matrix, s->qscale); | |
839 convert_matrix(s->q_non_intra_matrix, s->non_intra_matrix, s->qscale); | |
840 } | |
841 | |
842 switch(s->out_format) { | |
843 case FMT_MJPEG: | |
844 mjpeg_picture_header(s); | |
845 break; | |
846 case FMT_H263: | |
847 if (s->h263_msmpeg4) | |
848 msmpeg4_encode_picture_header(s, picture_number); | |
849 else if (s->h263_pred) | |
850 mpeg4_encode_picture_header(s, picture_number); | |
851 else if (s->h263_rv10) | |
852 rv10_encode_picture_header(s, picture_number); | |
853 else | |
854 h263_encode_picture_header(s, picture_number); | |
855 break; | |
856 case FMT_MPEG1: | |
857 mpeg1_encode_picture_header(s, picture_number); | |
858 break; | |
859 } | |
860 | |
861 /* init last dc values */ | |
862 /* note: quant matrix value (8) is implied here */ | |
863 s->last_dc[0] = 128; | |
864 s->last_dc[1] = 128; | |
865 s->last_dc[2] = 128; | |
866 s->mb_incr = 1; | |
867 s->last_mv[0][0][0] = 0; | |
868 s->last_mv[0][0][1] = 0; | |
869 s->mv_type = MV_TYPE_16X16; | |
870 s->mv_dir = MV_DIR_FORWARD; | |
871 | |
872 for(mb_y=0; mb_y < s->mb_height; mb_y++) { | |
873 for(mb_x=0; mb_x < s->mb_width; mb_x++) { | |
874 | |
875 s->mb_x = mb_x; | |
876 s->mb_y = mb_y; | |
877 | |
878 /* compute motion vector and macro block type (intra or non intra) */ | |
879 motion_x = 0; | |
880 motion_y = 0; | |
881 if (s->pict_type == P_TYPE) { | |
882 s->mb_intra = estimate_motion(s, mb_x, mb_y, | |
883 &motion_x, | |
884 &motion_y); | |
885 } else { | |
886 s->mb_intra = 1; | |
887 } | |
888 | |
889 /* get the pixels */ | |
890 wrap = s->linesize; | |
891 ptr = s->new_picture[0] + (mb_y * 16 * wrap) + mb_x * 16; | |
13
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
892 get_pixels(s->block[0], ptr, wrap); |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
893 get_pixels(s->block[1], ptr + 8, wrap); |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
894 get_pixels(s->block[2], ptr + 8 * wrap, wrap); |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
895 get_pixels(s->block[3], ptr + 8 * wrap + 8, wrap); |
0 | 896 wrap = s->linesize >> 1; |
897 ptr = s->new_picture[1] + (mb_y * 8 * wrap) + mb_x * 8; | |
13
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
898 get_pixels(s->block[4], ptr, wrap); |
0 | 899 |
900 wrap = s->linesize >> 1; | |
901 ptr = s->new_picture[2] + (mb_y * 8 * wrap) + mb_x * 8; | |
13
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
902 get_pixels(s->block[5], ptr, wrap); |
0 | 903 |
904 /* subtract previous frame if non intra */ | |
905 if (!s->mb_intra) { | |
906 int dxy, offset, mx, my; | |
907 | |
908 dxy = ((motion_y & 1) << 1) | (motion_x & 1); | |
909 ptr = s->last_picture[0] + | |
910 ((mb_y * 16 + (motion_y >> 1)) * s->linesize) + | |
911 (mb_x * 16 + (motion_x >> 1)); | |
912 | |
13
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
913 sub_pixels_2(s->block[0], ptr, s->linesize, dxy); |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
914 sub_pixels_2(s->block[1], ptr + 8, s->linesize, dxy); |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
915 sub_pixels_2(s->block[2], ptr + s->linesize * 8, s->linesize, dxy); |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
916 sub_pixels_2(s->block[3], ptr + 8 + s->linesize * 8, s->linesize ,dxy); |
0 | 917 |
918 if (s->out_format == FMT_H263) { | |
919 /* special rounding for h263 */ | |
920 dxy = 0; | |
921 if ((motion_x & 3) != 0) | |
922 dxy |= 1; | |
923 if ((motion_y & 3) != 0) | |
924 dxy |= 2; | |
925 mx = motion_x >> 2; | |
926 my = motion_y >> 2; | |
927 } else { | |
928 mx = motion_x / 2; | |
929 my = motion_y / 2; | |
930 dxy = ((my & 1) << 1) | (mx & 1); | |
931 mx >>= 1; | |
932 my >>= 1; | |
933 } | |
934 offset = ((mb_y * 8 + my) * (s->linesize >> 1)) + (mb_x * 8 + mx); | |
935 ptr = s->last_picture[1] + offset; | |
13
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
936 sub_pixels_2(s->block[4], ptr, s->linesize >> 1, dxy); |
0 | 937 ptr = s->last_picture[2] + offset; |
13
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
938 sub_pixels_2(s->block[5], ptr, s->linesize >> 1, dxy); |
0 | 939 } |
7
1d3ac9654178
added skip macroblock optimization (big perf win on black regions for example)
glantau
parents:
0
diff
changeset
|
940 emms_c(); |
0 | 941 |
942 /* DCT & quantize */ | |
943 if (s->h263_msmpeg4) { | |
944 msmpeg4_dc_scale(s); | |
945 } else if (s->h263_pred) { | |
946 h263_dc_scale(s); | |
947 } else { | |
948 /* default quantization values */ | |
949 s->y_dc_scale = 8; | |
950 s->c_dc_scale = 8; | |
951 } | |
952 | |
953 for(i=0;i<6;i++) { | |
954 int last_index; | |
955 if (av_fdct == jpeg_fdct_ifast) | |
13
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
956 last_index = dct_quantize(s, s->block[i], i, s->qscale); |
0 | 957 else |
13
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
958 last_index = dct_quantize_mmx(s, s->block[i], i, s->qscale); |
0 | 959 s->block_last_index[i] = last_index; |
960 } | |
961 | |
962 /* huffman encode */ | |
963 switch(s->out_format) { | |
964 case FMT_MPEG1: | |
13
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
965 mpeg1_encode_mb(s, s->block, motion_x, motion_y); |
0 | 966 break; |
967 case FMT_H263: | |
968 if (s->h263_msmpeg4) | |
13
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
969 msmpeg4_encode_mb(s, s->block, motion_x, motion_y); |
0 | 970 else |
13
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
971 h263_encode_mb(s, s->block, motion_x, motion_y); |
0 | 972 break; |
973 case FMT_MJPEG: | |
13
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
974 mjpeg_encode_mb(s, s->block); |
0 | 975 break; |
976 } | |
977 | |
978 /* decompress blocks so that we keep the state of the decoder */ | |
979 s->mv[0][0][0] = motion_x; | |
980 s->mv[0][0][1] = motion_y; | |
981 | |
13
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
982 MPV_decode_mb(s, s->block); |
0 | 983 } |
984 } | |
985 } | |
986 | |
987 static int dct_quantize(MpegEncContext *s, | |
988 DCTELEM *block, int n, | |
989 int qscale) | |
990 { | |
991 int i, j, level, last_non_zero, q; | |
992 const int *qmat; | |
993 | |
994 av_fdct (block); | |
995 | |
64 | 996 /* we need this permutation so that we correct the IDCT |
997 permutation. will be moved into DCT code */ | |
998 block_permute(block); | |
999 | |
0 | 1000 if (s->mb_intra) { |
1001 if (n < 4) | |
1002 q = s->y_dc_scale; | |
1003 else | |
1004 q = s->c_dc_scale; | |
1005 q = q << 3; | |
1006 | |
1007 /* note: block[0] is assumed to be positive */ | |
1008 block[0] = (block[0] + (q >> 1)) / q; | |
1009 i = 1; | |
1010 last_non_zero = 0; | |
1011 if (s->out_format == FMT_H263) { | |
1012 qmat = s->q_non_intra_matrix; | |
1013 } else { | |
1014 qmat = s->q_intra_matrix; | |
1015 } | |
1016 } else { | |
1017 i = 0; | |
1018 last_non_zero = -1; | |
1019 qmat = s->q_non_intra_matrix; | |
1020 } | |
1021 | |
1022 for(;i<64;i++) { | |
1023 j = zigzag_direct[i]; | |
1024 level = block[j]; | |
1025 level = level * qmat[j]; | |
1026 #ifdef PARANOID | |
1027 { | |
1028 static int count = 0; | |
1029 int level1, level2, qmat1; | |
1030 double val; | |
1031 if (qmat == s->q_non_intra_matrix) { | |
1032 qmat1 = default_non_intra_matrix[j] * s->qscale; | |
1033 } else { | |
1034 qmat1 = default_intra_matrix[j] * s->qscale; | |
1035 } | |
1036 if (av_fdct != jpeg_fdct_ifast) | |
1037 val = ((double)block[j] * 8.0) / (double)qmat1; | |
1038 else | |
1039 val = ((double)block[j] * 8.0 * 2048.0) / | |
1040 ((double)qmat1 * aanscales[j]); | |
1041 level1 = (int)val; | |
1042 level2 = level / (1 << (QMAT_SHIFT - 3)); | |
1043 if (level1 != level2) { | |
1044 fprintf(stderr, "%d: quant error qlevel=%d wanted=%d level=%d qmat1=%d qmat=%d wantedf=%0.6f\n", | |
1045 count, level2, level1, block[j], qmat1, qmat[j], | |
1046 val); | |
1047 count++; | |
1048 } | |
1049 | |
1050 } | |
1051 #endif | |
1052 /* XXX: slight error for the low range. Test should be equivalent to | |
1053 (level <= -(1 << (QMAT_SHIFT - 3)) || level >= (1 << | |
1054 (QMAT_SHIFT - 3))) | |
1055 */ | |
1056 if (((level << (31 - (QMAT_SHIFT - 3))) >> (31 - (QMAT_SHIFT - 3))) != | |
1057 level) { | |
1058 level = level / (1 << (QMAT_SHIFT - 3)); | |
1059 /* XXX: currently, this code is not optimal. the range should be: | |
1060 mpeg1: -255..255 | |
1061 mpeg2: -2048..2047 | |
1062 h263: -128..127 | |
1063 mpeg4: -2048..2047 | |
1064 */ | |
1065 if (level > 127) | |
1066 level = 127; | |
1067 else if (level < -128) | |
1068 level = -128; | |
1069 block[j] = level; | |
1070 last_non_zero = i; | |
1071 } else { | |
1072 block[j] = 0; | |
1073 } | |
1074 } | |
1075 return last_non_zero; | |
1076 } | |
1077 | |
1078 static int dct_quantize_mmx(MpegEncContext *s, | |
1079 DCTELEM *block, int n, | |
1080 int qscale) | |
1081 { | |
1082 int i, j, level, last_non_zero, q; | |
1083 const int *qmat; | |
1084 | |
1085 av_fdct (block); | |
40
f324d9047ca3
fixed block permutation in encoder (not optimal - should move it in forward DCT code)
glantau
parents:
19
diff
changeset
|
1086 |
f324d9047ca3
fixed block permutation in encoder (not optimal - should move it in forward DCT code)
glantau
parents:
19
diff
changeset
|
1087 /* we need this permutation so that we correct the IDCT |
f324d9047ca3
fixed block permutation in encoder (not optimal - should move it in forward DCT code)
glantau
parents:
19
diff
changeset
|
1088 permutation. will be moved into DCT code */ |
f324d9047ca3
fixed block permutation in encoder (not optimal - should move it in forward DCT code)
glantau
parents:
19
diff
changeset
|
1089 block_permute(block); |
0 | 1090 |
1091 if (s->mb_intra) { | |
1092 if (n < 4) | |
1093 q = s->y_dc_scale; | |
1094 else | |
1095 q = s->c_dc_scale; | |
1096 | |
1097 /* note: block[0] is assumed to be positive */ | |
1098 block[0] = (block[0] + (q >> 1)) / q; | |
1099 i = 1; | |
1100 last_non_zero = 0; | |
1101 if (s->out_format == FMT_H263) { | |
1102 qmat = s->q_non_intra_matrix; | |
1103 } else { | |
1104 qmat = s->q_intra_matrix; | |
1105 } | |
1106 } else { | |
1107 i = 0; | |
1108 last_non_zero = -1; | |
1109 qmat = s->q_non_intra_matrix; | |
1110 } | |
1111 | |
1112 for(;i<64;i++) { | |
1113 j = zigzag_direct[i]; | |
1114 level = block[j]; | |
1115 level = level * qmat[j]; | |
1116 /* XXX: slight error for the low range. Test should be equivalent to | |
1117 (level <= -(1 << (QMAT_SHIFT_MMX - 3)) || level >= (1 << | |
1118 (QMAT_SHIFT_MMX - 3))) | |
1119 */ | |
1120 if (((level << (31 - (QMAT_SHIFT_MMX - 3))) >> (31 - (QMAT_SHIFT_MMX - 3))) != | |
1121 level) { | |
1122 level = level / (1 << (QMAT_SHIFT_MMX - 3)); | |
1123 /* XXX: currently, this code is not optimal. the range should be: | |
1124 mpeg1: -255..255 | |
1125 mpeg2: -2048..2047 | |
1126 h263: -128..127 | |
1127 mpeg4: -2048..2047 | |
1128 */ | |
1129 if (level > 127) | |
1130 level = 127; | |
1131 else if (level < -128) | |
1132 level = -128; | |
1133 block[j] = level; | |
1134 last_non_zero = i; | |
1135 } else { | |
1136 block[j] = 0; | |
1137 } | |
1138 } | |
1139 return last_non_zero; | |
1140 } | |
1141 | |
13
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
1142 static void dct_unquantize_mpeg1_c(MpegEncContext *s, |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
1143 DCTELEM *block, int n, int qscale) |
0 | 1144 { |
1145 int i, level; | |
1146 const UINT16 *quant_matrix; | |
1147 | |
1148 if (s->mb_intra) { | |
1149 if (n < 4) | |
1150 block[0] = block[0] * s->y_dc_scale; | |
1151 else | |
1152 block[0] = block[0] * s->c_dc_scale; | |
1153 /* XXX: only mpeg1 */ | |
1154 quant_matrix = s->intra_matrix; | |
1155 for(i=1;i<64;i++) { | |
1156 level = block[i]; | |
1157 if (level) { | |
1158 if (level < 0) { | |
1159 level = -level; | |
1160 level = (int)(level * qscale * quant_matrix[i]) >> 3; | |
1161 level = (level - 1) | 1; | |
1162 level = -level; | |
1163 } else { | |
1164 level = (int)(level * qscale * quant_matrix[i]) >> 3; | |
1165 level = (level - 1) | 1; | |
1166 } | |
1167 #ifdef PARANOID | |
1168 if (level < -2048 || level > 2047) | |
1169 fprintf(stderr, "unquant error %d %d\n", i, level); | |
1170 #endif | |
1171 block[i] = level; | |
1172 } | |
1173 } | |
1174 } else { | |
1175 i = 0; | |
1176 quant_matrix = s->non_intra_matrix; | |
1177 for(;i<64;i++) { | |
1178 level = block[i]; | |
1179 if (level) { | |
1180 if (level < 0) { | |
1181 level = -level; | |
1182 level = (((level << 1) + 1) * qscale * | |
1183 ((int) (quant_matrix[i]))) >> 4; | |
1184 level = (level - 1) | 1; | |
1185 level = -level; | |
1186 } else { | |
1187 level = (((level << 1) + 1) * qscale * | |
1188 ((int) (quant_matrix[i]))) >> 4; | |
1189 level = (level - 1) | 1; | |
1190 } | |
1191 #ifdef PARANOID | |
1192 if (level < -2048 || level > 2047) | |
1193 fprintf(stderr, "unquant error %d %d\n", i, level); | |
1194 #endif | |
1195 block[i] = level; | |
1196 } | |
1197 } | |
1198 } | |
1199 } | |
13
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
1200 |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
1201 static void dct_unquantize_h263_c(MpegEncContext *s, |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
1202 DCTELEM *block, int n, int qscale) |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
1203 { |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
1204 int i, level, qmul, qadd; |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
1205 |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
1206 if (s->mb_intra) { |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
1207 if (n < 4) |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
1208 block[0] = block[0] * s->y_dc_scale; |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
1209 else |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
1210 block[0] = block[0] * s->c_dc_scale; |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
1211 i = 1; |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
1212 } else { |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
1213 i = 0; |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
1214 } |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
1215 |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
1216 qmul = s->qscale << 1; |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
1217 qadd = (s->qscale - 1) | 1; |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
1218 |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
1219 for(;i<64;i++) { |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
1220 level = block[i]; |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
1221 if (level) { |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
1222 if (level < 0) { |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
1223 level = level * qmul - qadd; |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
1224 } else { |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
1225 level = level * qmul + qadd; |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
1226 } |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
1227 #ifdef PARANOID |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
1228 if (level < -2048 || level > 2047) |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
1229 fprintf(stderr, "unquant error %d %d\n", i, level); |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
1230 #endif |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
1231 block[i] = level; |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
1232 } |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
1233 } |
174ef88f619a
use block[] in structure to have it aligned on 8 bytes for mmx optimizations - dct_unquantize is always a function pointer - added specialized dct_unquantize_h263
glantau
parents:
8
diff
changeset
|
1234 } |
0 | 1235 |
1236 /* rate control */ | |
1237 | |
1238 /* an I frame is I_FRAME_SIZE_RATIO bigger than a P frame */ | |
1239 #define I_FRAME_SIZE_RATIO 3.0 | |
1240 #define QSCALE_K 20 | |
1241 | |
1242 static void rate_control_init(MpegEncContext *s) | |
1243 { | |
1244 s->wanted_bits = 0; | |
1245 | |
1246 if (s->intra_only) { | |
1247 s->I_frame_bits = ((INT64)s->bit_rate * FRAME_RATE_BASE) / s->frame_rate; | |
1248 s->P_frame_bits = s->I_frame_bits; | |
1249 } else { | |
1250 s->P_frame_bits = (int) ((float)(s->gop_size * s->bit_rate) / | |
1251 (float)((float)s->frame_rate / FRAME_RATE_BASE * (I_FRAME_SIZE_RATIO + s->gop_size - 1))); | |
1252 s->I_frame_bits = (int)(s->P_frame_bits * I_FRAME_SIZE_RATIO); | |
1253 } | |
1254 | |
1255 #if defined(DEBUG) | |
1256 printf("I_frame_size=%d P_frame_size=%d\n", | |
1257 s->I_frame_bits, s->P_frame_bits); | |
1258 #endif | |
1259 } | |
1260 | |
1261 | |
1262 /* | |
1263 * This heuristic is rather poor, but at least we do not have to | |
1264 * change the qscale at every macroblock. | |
1265 */ | |
1266 static int rate_estimate_qscale(MpegEncContext *s) | |
1267 { | |
64 | 1268 INT64 total_bits = s->total_bits; |
0 | 1269 float q; |
1270 int qscale, diff, qmin; | |
1271 | |
1272 if (s->pict_type == I_TYPE) { | |
1273 s->wanted_bits += s->I_frame_bits; | |
1274 } else { | |
1275 s->wanted_bits += s->P_frame_bits; | |
1276 } | |
1277 diff = s->wanted_bits - total_bits; | |
1278 q = 31.0 - (float)diff / (QSCALE_K * s->mb_height * s->mb_width); | |
1279 /* adjust for I frame */ | |
1280 if (s->pict_type == I_TYPE && !s->intra_only) { | |
1281 q /= I_FRAME_SIZE_RATIO; | |
1282 } | |
1283 | |
1284 /* using a too small Q scale leeds to problems in mpeg1 and h263 | |
1285 because AC coefficients are clamped to 255 or 127 */ | |
1286 qmin = 3; | |
1287 if (q < qmin) | |
1288 q = qmin; | |
1289 else if (q > 31) | |
1290 q = 31; | |
1291 qscale = (int)(q + 0.5); | |
1292 #if defined(DEBUG) | |
64 | 1293 printf("%d: total=%0.0f br=%0.1f diff=%d qest=%0.1f\n", |
0 | 1294 s->picture_number, |
64 | 1295 (double)total_bits, |
0 | 1296 (float)s->frame_rate / FRAME_RATE_BASE * |
1297 total_bits / s->picture_number, | |
1298 diff, q); | |
1299 #endif | |
1300 return qscale; | |
1301 } | |
1302 | |
1303 AVCodec mpeg1video_encoder = { | |
1304 "mpeg1video", | |
1305 CODEC_TYPE_VIDEO, | |
1306 CODEC_ID_MPEG1VIDEO, | |
1307 sizeof(MpegEncContext), | |
1308 MPV_encode_init, | |
1309 MPV_encode_picture, | |
1310 MPV_encode_end, | |
1311 }; | |
1312 | |
1313 AVCodec h263_encoder = { | |
1314 "h263", | |
1315 CODEC_TYPE_VIDEO, | |
1316 CODEC_ID_H263, | |
1317 sizeof(MpegEncContext), | |
1318 MPV_encode_init, | |
1319 MPV_encode_picture, | |
1320 MPV_encode_end, | |
1321 }; | |
1322 | |
1323 AVCodec h263p_encoder = { | |
1324 "h263p", | |
1325 CODEC_TYPE_VIDEO, | |
1326 CODEC_ID_H263P, | |
1327 sizeof(MpegEncContext), | |
1328 MPV_encode_init, | |
1329 MPV_encode_picture, | |
1330 MPV_encode_end, | |
1331 }; | |
1332 | |
1333 AVCodec rv10_encoder = { | |
1334 "rv10", | |
1335 CODEC_TYPE_VIDEO, | |
1336 CODEC_ID_RV10, | |
1337 sizeof(MpegEncContext), | |
1338 MPV_encode_init, | |
1339 MPV_encode_picture, | |
1340 MPV_encode_end, | |
1341 }; | |
1342 | |
1343 AVCodec mjpeg_encoder = { | |
1344 "mjpeg", | |
1345 CODEC_TYPE_VIDEO, | |
1346 CODEC_ID_MJPEG, | |
1347 sizeof(MpegEncContext), | |
1348 MPV_encode_init, | |
1349 MPV_encode_picture, | |
1350 MPV_encode_end, | |
1351 }; | |
1352 | |
71 | 1353 AVCodec mpeg4_encoder = { |
1354 "mpeg4", | |
0 | 1355 CODEC_TYPE_VIDEO, |
71 | 1356 CODEC_ID_MPEG4, |
0 | 1357 sizeof(MpegEncContext), |
1358 MPV_encode_init, | |
1359 MPV_encode_picture, | |
1360 MPV_encode_end, | |
1361 }; | |
1362 | |
1363 AVCodec msmpeg4_encoder = { | |
1364 "msmpeg4", | |
1365 CODEC_TYPE_VIDEO, | |
1366 CODEC_ID_MSMPEG4, | |
1367 sizeof(MpegEncContext), | |
1368 MPV_encode_init, | |
1369 MPV_encode_picture, | |
1370 MPV_encode_end, | |
1371 }; |