Libav
svq1enc.c
Go to the documentation of this file.
1 /*
2  * SVQ1 Encoder
3  * Copyright (C) 2004 Mike Melanson <melanson@pcisys.net>
4  *
5  * This file is part of Libav.
6  *
7  * Libav is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU Lesser General Public
9  * License as published by the Free Software Foundation; either
10  * version 2.1 of the License, or (at your option) any later version.
11  *
12  * Libav is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15  * Lesser General Public License for more details.
16  *
17  * You should have received a copy of the GNU Lesser General Public
18  * License along with Libav; if not, write to the Free Software
19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20  */
21 
29 #include "avcodec.h"
30 #include "dsputil.h"
31 #include "hpeldsp.h"
32 #include "mpegvideo.h"
33 #include "h263.h"
34 #include "internal.h"
35 #include "svq1.h"
36 #include "svq1enc_cb.h"
37 
38 #undef NDEBUG
39 #include <assert.h>
40 
41 typedef struct SVQ1Context {
42  /* FIXME: Needed for motion estimation, should not be used for anything
43  * else, the idea is to make the motion estimation eventually independent
44  * of MpegEncContext, so this will be removed then. */
53 
54  /* why ooh why this sick breadth first order,
55  * everything is slower and more complex */
57 
60 
61  /* Y plane block dimensions */
64 
65  /* U & V plane (C planes) block dimensions */
68 
69  uint16_t *mb_type;
70  uint32_t *dummy;
71  int16_t (*motion_val8[3])[2];
72  int16_t (*motion_val16[3])[2];
73 
74  int64_t rd_total;
75 
77 } SVQ1Context;
78 
79 static void svq1_write_header(SVQ1Context *s, int frame_type)
80 {
81  int i;
82 
83  /* frame code */
84  put_bits(&s->pb, 22, 0x20);
85 
86  /* temporal reference (sure hope this is a "don't care") */
87  put_bits(&s->pb, 8, 0x00);
88 
89  /* frame type */
90  put_bits(&s->pb, 2, frame_type - 1);
91 
92  if (frame_type == AV_PICTURE_TYPE_I) {
93  /* no checksum since frame code is 0x20 */
94  /* no embedded string either */
95  /* output 5 unknown bits (2 + 2 + 1) */
96  put_bits(&s->pb, 5, 2); /* 2 needed by quicktime decoder */
97 
100  s->frame_width, s->frame_height);
101  put_bits(&s->pb, 3, i);
102 
103  if (i == 7) {
104  put_bits(&s->pb, 12, s->frame_width);
105  put_bits(&s->pb, 12, s->frame_height);
106  }
107  }
108 
109  /* no checksum or extra data (next 2 bits get 0) */
110  put_bits(&s->pb, 2, 0);
111 }
112 
113 #define QUALITY_THRESHOLD 100
114 #define THRESHOLD_MULTIPLIER 0.6
115 
116 static int encode_block(SVQ1Context *s, uint8_t *src, uint8_t *ref,
117  uint8_t *decoded, int stride, int level,
118  int threshold, int lambda, int intra)
119 {
120  int count, y, x, i, j, split, best_mean, best_score, best_count;
121  int best_vector[6];
122  int block_sum[7] = { 0, 0, 0, 0, 0, 0 };
123  int w = 2 << (level + 2 >> 1);
124  int h = 2 << (level + 1 >> 1);
125  int size = w * h;
126  int16_t block[7][256];
127  const int8_t *codebook_sum, *codebook;
128  const uint16_t(*mean_vlc)[2];
129  const uint8_t(*multistage_vlc)[2];
130 
131  best_score = 0;
132  // FIXME: Optimize, this does not need to be done multiple times.
133  if (intra) {
134  codebook_sum = svq1_intra_codebook_sum[level];
135  codebook = ff_svq1_intra_codebooks[level];
136  mean_vlc = ff_svq1_intra_mean_vlc;
137  multistage_vlc = ff_svq1_intra_multistage_vlc[level];
138  for (y = 0; y < h; y++) {
139  for (x = 0; x < w; x++) {
140  int v = src[x + y * stride];
141  block[0][x + w * y] = v;
142  best_score += v * v;
143  block_sum[0] += v;
144  }
145  }
146  } else {
147  codebook_sum = svq1_inter_codebook_sum[level];
148  codebook = ff_svq1_inter_codebooks[level];
149  mean_vlc = ff_svq1_inter_mean_vlc + 256;
150  multistage_vlc = ff_svq1_inter_multistage_vlc[level];
151  for (y = 0; y < h; y++) {
152  for (x = 0; x < w; x++) {
153  int v = src[x + y * stride] - ref[x + y * stride];
154  block[0][x + w * y] = v;
155  best_score += v * v;
156  block_sum[0] += v;
157  }
158  }
159  }
160 
161  best_count = 0;
162  best_score -= (int)((unsigned)block_sum[0] * block_sum[0] >> (level + 3));
163  best_mean = block_sum[0] + (size >> 1) >> (level + 3);
164 
165  if (level < 4) {
166  for (count = 1; count < 7; count++) {
167  int best_vector_score = INT_MAX;
168  int best_vector_sum = -999, best_vector_mean = -999;
169  const int stage = count - 1;
170  const int8_t *vector;
171 
172  for (i = 0; i < 16; i++) {
173  int sum = codebook_sum[stage * 16 + i];
174  int sqr, diff, score;
175 
176  vector = codebook + stage * size * 16 + i * size;
177  sqr = s->dsp.ssd_int8_vs_int16(vector, block[stage], size);
178  diff = block_sum[stage] - sum;
179  score = sqr - (diff * (int64_t)diff >> (level + 3)); // FIXME: 64bit slooow
180  if (score < best_vector_score) {
181  int mean = diff + (size >> 1) >> (level + 3);
182  assert(mean > -300 && mean < 300);
183  mean = av_clip(mean, intra ? 0 : -256, 255);
184  best_vector_score = score;
185  best_vector[stage] = i;
186  best_vector_sum = sum;
187  best_vector_mean = mean;
188  }
189  }
190  assert(best_vector_mean != -999);
191  vector = codebook + stage * size * 16 + best_vector[stage] * size;
192  for (j = 0; j < size; j++)
193  block[stage + 1][j] = block[stage][j] - vector[j];
194  block_sum[stage + 1] = block_sum[stage] - best_vector_sum;
195  best_vector_score += lambda *
196  (+1 + 4 * count +
197  multistage_vlc[1 + count][1]
198  + mean_vlc[best_vector_mean][1]);
199 
200  if (best_vector_score < best_score) {
201  best_score = best_vector_score;
202  best_count = count;
203  best_mean = best_vector_mean;
204  }
205  }
206  }
207 
208  split = 0;
209  if (best_score > threshold && level) {
210  int score = 0;
211  int offset = level & 1 ? stride * h / 2 : w / 2;
212  PutBitContext backup[6];
213 
214  for (i = level - 1; i >= 0; i--)
215  backup[i] = s->reorder_pb[i];
216  score += encode_block(s, src, ref, decoded, stride, level - 1,
217  threshold >> 1, lambda, intra);
218  score += encode_block(s, src + offset, ref + offset, decoded + offset,
219  stride, level - 1, threshold >> 1, lambda, intra);
220  score += lambda;
221 
222  if (score < best_score) {
223  best_score = score;
224  split = 1;
225  } else {
226  for (i = level - 1; i >= 0; i--)
227  s->reorder_pb[i] = backup[i];
228  }
229  }
230  if (level > 0)
231  put_bits(&s->reorder_pb[level], 1, split);
232 
233  if (!split) {
234  assert(best_mean >= 0 && best_mean < 256 || !intra);
235  assert(best_mean >= -256 && best_mean < 256);
236  assert(best_count >= 0 && best_count < 7);
237  assert(level < 4 || best_count == 0);
238 
239  /* output the encoding */
240  put_bits(&s->reorder_pb[level],
241  multistage_vlc[1 + best_count][1],
242  multistage_vlc[1 + best_count][0]);
243  put_bits(&s->reorder_pb[level], mean_vlc[best_mean][1],
244  mean_vlc[best_mean][0]);
245 
246  for (i = 0; i < best_count; i++) {
247  assert(best_vector[i] >= 0 && best_vector[i] < 16);
248  put_bits(&s->reorder_pb[level], 4, best_vector[i]);
249  }
250 
251  for (y = 0; y < h; y++)
252  for (x = 0; x < w; x++)
253  decoded[x + y * stride] = src[x + y * stride] -
254  block[best_count][x + w * y] +
255  best_mean;
256  }
257 
258  return best_score;
259 }
260 
261 static int svq1_encode_plane(SVQ1Context *s, int plane,
262  unsigned char *src_plane,
263  unsigned char *ref_plane,
264  unsigned char *decoded_plane,
265  int width, int height, int src_stride, int stride)
266 {
267  const AVFrame *f = s->avctx->coded_frame;
268  int x, y;
269  int i;
270  int block_width, block_height;
271  int level;
272  int threshold[6];
273  uint8_t *src = s->scratchbuf + stride * 16;
274  const int lambda = (f->quality * f->quality) >>
275  (2 * FF_LAMBDA_SHIFT);
276 
277  /* figure out the acceptable level thresholds in advance */
278  threshold[5] = QUALITY_THRESHOLD;
279  for (level = 4; level >= 0; level--)
280  threshold[level] = threshold[level + 1] * THRESHOLD_MULTIPLIER;
281 
282  block_width = (width + 15) / 16;
283  block_height = (height + 15) / 16;
284 
285  if (f->pict_type == AV_PICTURE_TYPE_P) {
286  s->m.avctx = s->avctx;
288  s->m.last_picture_ptr = &s->m.last_picture;
289  s->m.last_picture.f.data[0] = ref_plane;
290  s->m.linesize =
291  s->m.last_picture.f.linesize[0] =
292  s->m.new_picture.f.linesize[0] =
294  s->m.width = width;
295  s->m.height = height;
296  s->m.mb_width = block_width;
297  s->m.mb_height = block_height;
298  s->m.mb_stride = s->m.mb_width + 1;
299  s->m.b8_stride = 2 * s->m.mb_width + 1;
300  s->m.f_code = 1;
301  s->m.pict_type = f->pict_type;
302  s->m.me_method = s->avctx->me_method;
303  s->m.me.scene_change_score = 0;
304  s->m.flags = s->avctx->flags;
305  // s->m.out_format = FMT_H263;
306  // s->m.unrestricted_mv = 1;
307  s->m.lambda = f->quality;
308  s->m.qscale = s->m.lambda * 139 +
309  FF_LAMBDA_SCALE * 64 >>
310  FF_LAMBDA_SHIFT + 7;
311  s->m.lambda2 = s->m.lambda * s->m.lambda +
312  FF_LAMBDA_SCALE / 2 >>
314 
315  if (!s->motion_val8[plane]) {
316  s->motion_val8[plane] = av_mallocz((s->m.b8_stride *
317  block_height * 2 + 2) *
318  2 * sizeof(int16_t));
319  s->motion_val16[plane] = av_mallocz((s->m.mb_stride *
320  (block_height + 2) + 1) *
321  2 * sizeof(int16_t));
322  }
323 
324  s->m.mb_type = s->mb_type;
325 
326  // dummies, to avoid segfaults
328  s->m.current_picture.mb_var = (uint16_t *)s->dummy;
329  s->m.current_picture.mc_mb_var = (uint16_t *)s->dummy;
330  s->m.current_picture.mb_type = s->dummy;
331 
332  s->m.current_picture.motion_val[0] = s->motion_val8[plane] + 2;
333  s->m.p_mv_table = s->motion_val16[plane] +
334  s->m.mb_stride + 1;
335  s->m.dsp = s->dsp; // move
336  ff_init_me(&s->m);
337 
338  s->m.me.dia_size = s->avctx->dia_size;
339  s->m.first_slice_line = 1;
340  for (y = 0; y < block_height; y++) {
341  s->m.new_picture.f.data[0] = src - y * 16 * stride; // ugly
342  s->m.mb_y = y;
343 
344  for (i = 0; i < 16 && i + 16 * y < height; i++) {
345  memcpy(&src[i * stride], &src_plane[(i + 16 * y) * src_stride],
346  width);
347  for (x = width; x < 16 * block_width; x++)
348  src[i * stride + x] = src[i * stride + x - 1];
349  }
350  for (; i < 16 && i + 16 * y < 16 * block_height; i++)
351  memcpy(&src[i * stride], &src[(i - 1) * stride],
352  16 * block_width);
353 
354  for (x = 0; x < block_width; x++) {
355  s->m.mb_x = x;
356  ff_init_block_index(&s->m);
358 
359  ff_estimate_p_frame_motion(&s->m, x, y);
360  }
361  s->m.first_slice_line = 0;
362  }
363 
364  ff_fix_long_p_mvs(&s->m);
365  ff_fix_long_mvs(&s->m, NULL, 0, s->m.p_mv_table, s->m.f_code,
367  }
368 
369  s->m.first_slice_line = 1;
370  for (y = 0; y < block_height; y++) {
371  for (i = 0; i < 16 && i + 16 * y < height; i++) {
372  memcpy(&src[i * stride], &src_plane[(i + 16 * y) * src_stride],
373  width);
374  for (x = width; x < 16 * block_width; x++)
375  src[i * stride + x] = src[i * stride + x - 1];
376  }
377  for (; i < 16 && i + 16 * y < 16 * block_height; i++)
378  memcpy(&src[i * stride], &src[(i - 1) * stride], 16 * block_width);
379 
380  s->m.mb_y = y;
381  for (x = 0; x < block_width; x++) {
382  uint8_t reorder_buffer[3][6][7 * 32];
383  int count[3][6];
384  int offset = y * 16 * stride + x * 16;
385  uint8_t *decoded = decoded_plane + offset;
386  uint8_t *ref = ref_plane + offset;
387  int score[4] = { 0, 0, 0, 0 }, best;
388  uint8_t *temp = s->scratchbuf;
389 
390  if (s->pb.buf_end - s->pb.buf -
391  (put_bits_count(&s->pb) >> 3) < 3000) { // FIXME: check size
392  av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
393  return -1;
394  }
395 
396  s->m.mb_x = x;
397  ff_init_block_index(&s->m);
399 
400  if (f->pict_type == AV_PICTURE_TYPE_I ||
401  (s->m.mb_type[x + y * s->m.mb_stride] &
403  for (i = 0; i < 6; i++)
404  init_put_bits(&s->reorder_pb[i], reorder_buffer[0][i],
405  7 * 32);
406  if (f->pict_type == AV_PICTURE_TYPE_P) {
408  put_bits(&s->reorder_pb[5], vlc[1], vlc[0]);
409  score[0] = vlc[1] * lambda;
410  }
411  score[0] += encode_block(s, src + 16 * x, NULL, temp, stride,
412  5, 64, lambda, 1);
413  for (i = 0; i < 6; i++) {
414  count[0][i] = put_bits_count(&s->reorder_pb[i]);
415  flush_put_bits(&s->reorder_pb[i]);
416  }
417  } else
418  score[0] = INT_MAX;
419 
420  best = 0;
421 
422  if (f->pict_type == AV_PICTURE_TYPE_P) {
424  int mx, my, pred_x, pred_y, dxy;
425  int16_t *motion_ptr;
426 
427  motion_ptr = ff_h263_pred_motion(&s->m, 0, 0, &pred_x, &pred_y);
428  if (s->m.mb_type[x + y * s->m.mb_stride] &
430  for (i = 0; i < 6; i++)
431  init_put_bits(&s->reorder_pb[i], reorder_buffer[1][i],
432  7 * 32);
433 
434  put_bits(&s->reorder_pb[5], vlc[1], vlc[0]);
435 
436  s->m.pb = s->reorder_pb[5];
437  mx = motion_ptr[0];
438  my = motion_ptr[1];
439  assert(mx >= -32 && mx <= 31);
440  assert(my >= -32 && my <= 31);
441  assert(pred_x >= -32 && pred_x <= 31);
442  assert(pred_y >= -32 && pred_y <= 31);
443  ff_h263_encode_motion(&s->m, mx - pred_x, 1);
444  ff_h263_encode_motion(&s->m, my - pred_y, 1);
445  s->reorder_pb[5] = s->m.pb;
446  score[1] += lambda * put_bits_count(&s->reorder_pb[5]);
447 
448  dxy = (mx & 1) + 2 * (my & 1);
449 
450  s->hdsp.put_pixels_tab[0][dxy](temp + 16,
451  ref + (mx >> 1) +
452  stride * (my >> 1),
453  stride, 16);
454 
455  score[1] += encode_block(s, src + 16 * x, temp + 16,
456  decoded, stride, 5, 64, lambda, 0);
457  best = score[1] <= score[0];
458 
460  score[2] = s->dsp.sse[0](NULL, src + 16 * x, ref,
461  stride, 16);
462  score[2] += vlc[1] * lambda;
463  if (score[2] < score[best] && mx == 0 && my == 0) {
464  best = 2;
465  s->hdsp.put_pixels_tab[0][0](decoded, ref, stride, 16);
466  for (i = 0; i < 6; i++)
467  count[2][i] = 0;
468  put_bits(&s->pb, vlc[1], vlc[0]);
469  }
470  }
471 
472  if (best == 1) {
473  for (i = 0; i < 6; i++) {
474  count[1][i] = put_bits_count(&s->reorder_pb[i]);
475  flush_put_bits(&s->reorder_pb[i]);
476  }
477  } else {
478  motion_ptr[0] =
479  motion_ptr[1] =
480  motion_ptr[2] =
481  motion_ptr[3] =
482  motion_ptr[0 + 2 * s->m.b8_stride] =
483  motion_ptr[1 + 2 * s->m.b8_stride] =
484  motion_ptr[2 + 2 * s->m.b8_stride] =
485  motion_ptr[3 + 2 * s->m.b8_stride] = 0;
486  }
487  }
488 
489  s->rd_total += score[best];
490 
491  for (i = 5; i >= 0; i--)
492  avpriv_copy_bits(&s->pb, reorder_buffer[best][i],
493  count[best][i]);
494  if (best == 0)
495  s->hdsp.put_pixels_tab[0][0](decoded, temp, stride, 16);
496  }
497  s->m.first_slice_line = 0;
498  }
499  return 0;
500 }
501 
503 {
504  SVQ1Context *const s = avctx->priv_data;
505  int i;
506 
507  av_log(avctx, AV_LOG_DEBUG, "RD: %f\n",
508  s->rd_total / (double)(avctx->width * avctx->height *
509  avctx->frame_number));
510 
511  av_freep(&s->m.me.scratchpad);
512  av_freep(&s->m.me.map);
513  av_freep(&s->m.me.score_map);
514  av_freep(&s->mb_type);
515  av_freep(&s->dummy);
516  av_freep(&s->scratchbuf);
517 
518  for (i = 0; i < 3; i++) {
519  av_freep(&s->motion_val8[i]);
520  av_freep(&s->motion_val16[i]);
521  }
522 
525  av_frame_free(&avctx->coded_frame);
526 
527  return 0;
528 }
529 
531 {
532  SVQ1Context *const s = avctx->priv_data;
533 
534  ff_dsputil_init(&s->dsp, avctx);
535  ff_hpeldsp_init(&s->hdsp, avctx->flags);
536 
537  avctx->coded_frame = av_frame_alloc();
540  if (!avctx->coded_frame || !s->current_picture || !s->last_picture) {
541  svq1_encode_end(avctx);
542  return AVERROR(ENOMEM);
543  }
544 
545  s->frame_width = avctx->width;
546  s->frame_height = avctx->height;
547 
548  s->y_block_width = (s->frame_width + 15) / 16;
549  s->y_block_height = (s->frame_height + 15) / 16;
550 
551  s->c_block_width = (s->frame_width / 4 + 15) / 16;
552  s->c_block_height = (s->frame_height / 4 + 15) / 16;
553 
554  s->avctx = avctx;
555  s->m.avctx = avctx;
557  s->m.me.temp =
558  s->m.me.scratchpad = av_mallocz((avctx->width + 64) *
559  2 * 16 * 2 * sizeof(uint8_t));
560  s->m.me.map = av_mallocz(ME_MAP_SIZE * sizeof(uint32_t));
561  s->m.me.score_map = av_mallocz(ME_MAP_SIZE * sizeof(uint32_t));
562  s->mb_type = av_mallocz((s->y_block_width + 1) *
563  s->y_block_height * sizeof(int16_t));
564  s->dummy = av_mallocz((s->y_block_width + 1) *
565  s->y_block_height * sizeof(int32_t));
566  ff_h263_encode_init(&s->m); // mv_penalty
567 
568  return 0;
569 }
570 
571 static int svq1_encode_frame(AVCodecContext *avctx, AVPacket *pkt,
572  const AVFrame *pict, int *got_packet)
573 {
574  SVQ1Context *const s = avctx->priv_data;
575  AVFrame *const p = avctx->coded_frame;
576  int i, ret;
577 
578  if (!pkt->data &&
579  (ret = av_new_packet(pkt, s->y_block_width * s->y_block_height *
580  MAX_MB_BYTES * 3 + FF_MIN_BUFFER_SIZE)) < 0) {
581  av_log(avctx, AV_LOG_ERROR, "Error getting output packet.\n");
582  return ret;
583  }
584 
585  if (avctx->pix_fmt != AV_PIX_FMT_YUV410P) {
586  av_log(avctx, AV_LOG_ERROR, "unsupported pixel format\n");
587  return -1;
588  }
589 
590  if (!s->current_picture->data[0]) {
591  ff_get_buffer(avctx, s->current_picture, 0);
592  ff_get_buffer(avctx, s->last_picture, 0);
593  s->scratchbuf = av_malloc(s->current_picture->linesize[0] * 16 * 2);
594  }
595 
597 
598  init_put_bits(&s->pb, pkt->data, pkt->size);
599 
600  p->pict_type = avctx->gop_size && avctx->frame_number % avctx->gop_size ?
603  p->quality = pict->quality;
604 
606  for (i = 0; i < 3; i++)
607  if (svq1_encode_plane(s, i,
608  pict->data[i],
609  s->last_picture->data[i],
610  s->current_picture->data[i],
611  s->frame_width / (i ? 4 : 1),
612  s->frame_height / (i ? 4 : 1),
613  pict->linesize[i],
614  s->current_picture->linesize[i]) < 0)
615  return -1;
616 
617  // avpriv_align_put_bits(&s->pb);
618  while (put_bits_count(&s->pb) & 31)
619  put_bits(&s->pb, 1, 0);
620 
621  flush_put_bits(&s->pb);
622 
623  pkt->size = put_bits_count(&s->pb) / 8;
624  if (p->pict_type == AV_PICTURE_TYPE_I)
625  pkt->flags |= AV_PKT_FLAG_KEY;
626  *got_packet = 1;
627 
628  return 0;
629 }
630 
632  .name = "svq1",
633  .long_name = NULL_IF_CONFIG_SMALL("Sorenson Vector Quantizer 1 / Sorenson Video 1 / SVQ1"),
634  .type = AVMEDIA_TYPE_VIDEO,
635  .id = AV_CODEC_ID_SVQ1,
636  .priv_data_size = sizeof(SVQ1Context),
638  .encode2 = svq1_encode_frame,
640  .pix_fmts = (const enum AVPixelFormat[]) { AV_PIX_FMT_YUV410P,
641  AV_PIX_FMT_NONE },
642 };