FFmpeg
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
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 FFmpeg.
6  *
7  * FFmpeg 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  * FFmpeg 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 FFmpeg; if not, write to the Free Software
19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20  */
21 
22 /**
23  * @file
24  * Sorenson Vector Quantizer #1 (SVQ1) video codec.
25  * For more information of the SVQ1 algorithm, visit:
26  * http://www.pcisys.net/~melanson/codecs/
27  */
28 
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 "mpegutils.h"
36 #include "svq1.h"
37 #include "svq1enc.h"
38 #include "svq1enc_cb.h"
39 #include "libavutil/avassert.h"
40 
41 
42 static void svq1_write_header(SVQ1EncContext *s, int frame_type)
43 {
44  int i;
45 
46  /* frame code */
47  put_bits(&s->pb, 22, 0x20);
48 
49  /* temporal reference (sure hope this is a "don't care") */
50  put_bits(&s->pb, 8, 0x00);
51 
52  /* frame type */
53  put_bits(&s->pb, 2, frame_type - 1);
54 
55  if (frame_type == AV_PICTURE_TYPE_I) {
56  /* no checksum since frame code is 0x20 */
57  /* no embedded string either */
58  /* output 5 unknown bits (2 + 2 + 1) */
59  put_bits(&s->pb, 5, 2); /* 2 needed by quicktime decoder */
60 
63  s->frame_width, s->frame_height);
64  put_bits(&s->pb, 3, i);
65 
66  if (i == 7) {
67  put_bits(&s->pb, 12, s->frame_width);
68  put_bits(&s->pb, 12, s->frame_height);
69  }
70  }
71 
72  /* no checksum or extra data (next 2 bits get 0) */
73  put_bits(&s->pb, 2, 0);
74 }
75 
76 #define QUALITY_THRESHOLD 100
77 #define THRESHOLD_MULTIPLIER 0.6
78 
79 static int ssd_int8_vs_int16_c(const int8_t *pix1, const int16_t *pix2,
80  intptr_t size)
81 {
82  int score = 0, i;
83 
84  for (i = 0; i < size; i++)
85  score += (pix1[i] - pix2[i]) * (pix1[i] - pix2[i]);
86  return score;
87 }
88 
90  uint8_t *decoded, int stride, int level,
91  int threshold, int lambda, int intra)
92 {
93  int count, y, x, i, j, split, best_mean, best_score, best_count;
94  int best_vector[6];
95  int block_sum[7] = { 0, 0, 0, 0, 0, 0 };
96  int w = 2 << (level + 2 >> 1);
97  int h = 2 << (level + 1 >> 1);
98  int size = w * h;
99  int16_t block[7][256];
100  const int8_t *codebook_sum, *codebook;
101  const uint16_t(*mean_vlc)[2];
102  const uint8_t(*multistage_vlc)[2];
103 
104  best_score = 0;
105  // FIXME: Optimize, this does not need to be done multiple times.
106  if (intra) {
107  codebook_sum = svq1_intra_codebook_sum[level];
108  codebook = ff_svq1_intra_codebooks[level];
109  mean_vlc = ff_svq1_intra_mean_vlc;
110  multistage_vlc = ff_svq1_intra_multistage_vlc[level];
111  for (y = 0; y < h; y++) {
112  for (x = 0; x < w; x++) {
113  int v = src[x + y * stride];
114  block[0][x + w * y] = v;
115  best_score += v * v;
116  block_sum[0] += v;
117  }
118  }
119  } else {
120  codebook_sum = svq1_inter_codebook_sum[level];
121  codebook = ff_svq1_inter_codebooks[level];
122  mean_vlc = ff_svq1_inter_mean_vlc + 256;
123  multistage_vlc = ff_svq1_inter_multistage_vlc[level];
124  for (y = 0; y < h; y++) {
125  for (x = 0; x < w; x++) {
126  int v = src[x + y * stride] - ref[x + y * stride];
127  block[0][x + w * y] = v;
128  best_score += v * v;
129  block_sum[0] += v;
130  }
131  }
132  }
133 
134  best_count = 0;
135  best_score -= (int)((unsigned)block_sum[0] * block_sum[0] >> (level + 3));
136  best_mean = block_sum[0] + (size >> 1) >> (level + 3);
137 
138  if (level < 4) {
139  for (count = 1; count < 7; count++) {
140  int best_vector_score = INT_MAX;
141  int best_vector_sum = -999, best_vector_mean = -999;
142  const int stage = count - 1;
143  const int8_t *vector;
144 
145  for (i = 0; i < 16; i++) {
146  int sum = codebook_sum[stage * 16 + i];
147  int sqr, diff, score;
148 
149  vector = codebook + stage * size * 16 + i * size;
150  sqr = s->ssd_int8_vs_int16(vector, block[stage], size);
151  diff = block_sum[stage] - sum;
152  score = sqr - (diff * (int64_t)diff >> (level + 3)); // FIXME: 64bit slooow
153  if (score < best_vector_score) {
154  int mean = diff + (size >> 1) >> (level + 3);
155  av_assert2(mean > -300 && mean < 300);
156  mean = av_clip(mean, intra ? 0 : -256, 255);
157  best_vector_score = score;
158  best_vector[stage] = i;
159  best_vector_sum = sum;
160  best_vector_mean = mean;
161  }
162  }
163  av_assert0(best_vector_mean != -999);
164  vector = codebook + stage * size * 16 + best_vector[stage] * size;
165  for (j = 0; j < size; j++)
166  block[stage + 1][j] = block[stage][j] - vector[j];
167  block_sum[stage + 1] = block_sum[stage] - best_vector_sum;
168  best_vector_score += lambda *
169  (+1 + 4 * count +
170  multistage_vlc[1 + count][1]
171  + mean_vlc[best_vector_mean][1]);
172 
173  if (best_vector_score < best_score) {
174  best_score = best_vector_score;
175  best_count = count;
176  best_mean = best_vector_mean;
177  }
178  }
179  }
180 
181  split = 0;
182  if (best_score > threshold && level) {
183  int score = 0;
184  int offset = level & 1 ? stride * h / 2 : w / 2;
185  PutBitContext backup[6];
186 
187  for (i = level - 1; i >= 0; i--)
188  backup[i] = s->reorder_pb[i];
189  score += encode_block(s, src, ref, decoded, stride, level - 1,
190  threshold >> 1, lambda, intra);
191  score += encode_block(s, src + offset, ref + offset, decoded + offset,
192  stride, level - 1, threshold >> 1, lambda, intra);
193  score += lambda;
194 
195  if (score < best_score) {
196  best_score = score;
197  split = 1;
198  } else {
199  for (i = level - 1; i >= 0; i--)
200  s->reorder_pb[i] = backup[i];
201  }
202  }
203  if (level > 0)
204  put_bits(&s->reorder_pb[level], 1, split);
205 
206  if (!split) {
207  av_assert1(best_mean >= 0 && best_mean < 256 || !intra);
208  av_assert1(best_mean >= -256 && best_mean < 256);
209  av_assert1(best_count >= 0 && best_count < 7);
210  av_assert1(level < 4 || best_count == 0);
211 
212  /* output the encoding */
213  put_bits(&s->reorder_pb[level],
214  multistage_vlc[1 + best_count][1],
215  multistage_vlc[1 + best_count][0]);
216  put_bits(&s->reorder_pb[level], mean_vlc[best_mean][1],
217  mean_vlc[best_mean][0]);
218 
219  for (i = 0; i < best_count; i++) {
220  av_assert2(best_vector[i] >= 0 && best_vector[i] < 16);
221  put_bits(&s->reorder_pb[level], 4, best_vector[i]);
222  }
223 
224  for (y = 0; y < h; y++)
225  for (x = 0; x < w; x++)
226  decoded[x + y * stride] = src[x + y * stride] -
227  block[best_count][x + w * y] +
228  best_mean;
229  }
230 
231  return best_score;
232 }
233 
235  s->block_index[0]= s->b8_stride*(s->mb_y*2 ) + s->mb_x*2;
236  s->block_index[1]= s->b8_stride*(s->mb_y*2 ) + 1 + s->mb_x*2;
237  s->block_index[2]= s->b8_stride*(s->mb_y*2 + 1) + s->mb_x*2;
238  s->block_index[3]= s->b8_stride*(s->mb_y*2 + 1) + 1 + s->mb_x*2;
239  s->block_index[4]= s->mb_stride*(s->mb_y + 1) + s->b8_stride*s->mb_height*2 + s->mb_x;
240  s->block_index[5]= s->mb_stride*(s->mb_y + s->mb_height + 2) + s->b8_stride*s->mb_height*2 + s->mb_x;
241 }
242 
243 static int svq1_encode_plane(SVQ1EncContext *s, int plane,
244  unsigned char *src_plane,
245  unsigned char *ref_plane,
246  unsigned char *decoded_plane,
247  int width, int height, int src_stride, int stride)
248 {
249  const AVFrame *f = s->avctx->coded_frame;
250  int x, y;
251  int i;
252  int block_width, block_height;
253  int level;
254  int threshold[6];
255  uint8_t *src = s->scratchbuf + stride * 32;
256  const int lambda = (f->quality * f->quality) >>
257  (2 * FF_LAMBDA_SHIFT);
258 
259  /* figure out the acceptable level thresholds in advance */
260  threshold[5] = QUALITY_THRESHOLD;
261  for (level = 4; level >= 0; level--)
262  threshold[level] = threshold[level + 1] * THRESHOLD_MULTIPLIER;
263 
264  block_width = (width + 15) / 16;
265  block_height = (height + 15) / 16;
266 
267  if (f->pict_type == AV_PICTURE_TYPE_P) {
268  s->m.avctx = s->avctx;
270  s->m.last_picture_ptr = &s->m.last_picture;
271  s->m.last_picture.f->data[0] = ref_plane;
272  s->m.linesize =
273  s->m.last_picture.f->linesize[0] =
274  s->m.new_picture.f->linesize[0] =
275  s->m.current_picture.f->linesize[0] = stride;
276  s->m.width = width;
277  s->m.height = height;
278  s->m.mb_width = block_width;
279  s->m.mb_height = block_height;
280  s->m.mb_stride = s->m.mb_width + 1;
281  s->m.b8_stride = 2 * s->m.mb_width + 1;
282  s->m.f_code = 1;
283  s->m.pict_type = f->pict_type;
284  s->m.me_method = s->avctx->me_method;
285  s->m.me.scene_change_score = 0;
286  s->m.flags = s->avctx->flags;
287  // s->m.out_format = FMT_H263;
288  // s->m.unrestricted_mv = 1;
289  s->m.lambda = f->quality;
290  s->m.qscale = s->m.lambda * 139 +
291  FF_LAMBDA_SCALE * 64 >>
292  FF_LAMBDA_SHIFT + 7;
293  s->m.lambda2 = s->m.lambda * s->m.lambda +
294  FF_LAMBDA_SCALE / 2 >>
296 
297  if (!s->motion_val8[plane]) {
298  s->motion_val8[plane] = av_mallocz((s->m.b8_stride *
299  block_height * 2 + 2) *
300  2 * sizeof(int16_t));
301  s->motion_val16[plane] = av_mallocz((s->m.mb_stride *
302  (block_height + 2) + 1) *
303  2 * sizeof(int16_t));
304  }
305 
306  s->m.mb_type = s->mb_type;
307 
308  // dummies, to avoid segfaults
310  s->m.current_picture.mb_var = (uint16_t *)s->dummy;
311  s->m.current_picture.mc_mb_var = (uint16_t *)s->dummy;
312  s->m.current_picture.mb_type = s->dummy;
313 
314  s->m.current_picture.motion_val[0] = s->motion_val8[plane] + 2;
315  s->m.p_mv_table = s->motion_val16[plane] +
316  s->m.mb_stride + 1;
317  s->m.dsp = s->dsp; // move
318  ff_init_me(&s->m);
319 
320  s->m.me.dia_size = s->avctx->dia_size;
321  s->m.first_slice_line = 1;
322  for (y = 0; y < block_height; y++) {
323  s->m.new_picture.f->data[0] = src - y * 16 * stride; // ugly
324  s->m.mb_y = y;
325 
326  for (i = 0; i < 16 && i + 16 * y < height; i++) {
327  memcpy(&src[i * stride], &src_plane[(i + 16 * y) * src_stride],
328  width);
329  for (x = width; x < 16 * block_width; x++)
330  src[i * stride + x] = src[i * stride + x - 1];
331  }
332  for (; i < 16 && i + 16 * y < 16 * block_height; i++)
333  memcpy(&src[i * stride], &src[(i - 1) * stride],
334  16 * block_width);
335 
336  for (x = 0; x < block_width; x++) {
337  s->m.mb_x = x;
338  init_block_index(&s->m);
339 
340  ff_estimate_p_frame_motion(&s->m, x, y);
341  }
342  s->m.first_slice_line = 0;
343  }
344 
345  ff_fix_long_p_mvs(&s->m);
346  ff_fix_long_mvs(&s->m, NULL, 0, s->m.p_mv_table, s->m.f_code,
348  }
349 
350  s->m.first_slice_line = 1;
351  for (y = 0; y < block_height; y++) {
352  for (i = 0; i < 16 && i + 16 * y < height; i++) {
353  memcpy(&src[i * stride], &src_plane[(i + 16 * y) * src_stride],
354  width);
355  for (x = width; x < 16 * block_width; x++)
356  src[i * stride + x] = src[i * stride + x - 1];
357  }
358  for (; i < 16 && i + 16 * y < 16 * block_height; i++)
359  memcpy(&src[i * stride], &src[(i - 1) * stride], 16 * block_width);
360 
361  s->m.mb_y = y;
362  for (x = 0; x < block_width; x++) {
363  uint8_t reorder_buffer[3][6][7 * 32];
364  int count[3][6];
365  int offset = y * 16 * stride + x * 16;
366  uint8_t *decoded = decoded_plane + offset;
367  uint8_t *ref = ref_plane + offset;
368  int score[4] = { 0, 0, 0, 0 }, best;
369  uint8_t *temp = s->scratchbuf;
370 
371  if (s->pb.buf_end - s->pb.buf -
372  (put_bits_count(&s->pb) >> 3) < 3000) { // FIXME: check size
373  av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
374  return -1;
375  }
376 
377  s->m.mb_x = x;
378  init_block_index(&s->m);
379 
380  if (f->pict_type == AV_PICTURE_TYPE_I ||
381  (s->m.mb_type[x + y * s->m.mb_stride] &
383  for (i = 0; i < 6; i++)
384  init_put_bits(&s->reorder_pb[i], reorder_buffer[0][i],
385  7 * 32);
386  if (f->pict_type == AV_PICTURE_TYPE_P) {
388  put_bits(&s->reorder_pb[5], vlc[1], vlc[0]);
389  score[0] = vlc[1] * lambda;
390  }
391  score[0] += encode_block(s, src + 16 * x, NULL, temp, stride,
392  5, 64, lambda, 1);
393  for (i = 0; i < 6; i++) {
394  count[0][i] = put_bits_count(&s->reorder_pb[i]);
395  flush_put_bits(&s->reorder_pb[i]);
396  }
397  } else
398  score[0] = INT_MAX;
399 
400  best = 0;
401 
402  if (f->pict_type == AV_PICTURE_TYPE_P) {
404  int mx, my, pred_x, pred_y, dxy;
405  int16_t *motion_ptr;
406 
407  motion_ptr = ff_h263_pred_motion(&s->m, 0, 0, &pred_x, &pred_y);
408  if (s->m.mb_type[x + y * s->m.mb_stride] &
410  for (i = 0; i < 6; i++)
411  init_put_bits(&s->reorder_pb[i], reorder_buffer[1][i],
412  7 * 32);
413 
414  put_bits(&s->reorder_pb[5], vlc[1], vlc[0]);
415 
416  s->m.pb = s->reorder_pb[5];
417  mx = motion_ptr[0];
418  my = motion_ptr[1];
419  av_assert1(mx >= -32 && mx <= 31);
420  av_assert1(my >= -32 && my <= 31);
421  av_assert1(pred_x >= -32 && pred_x <= 31);
422  av_assert1(pred_y >= -32 && pred_y <= 31);
423  ff_h263_encode_motion(&s->m, mx - pred_x, 1);
424  ff_h263_encode_motion(&s->m, my - pred_y, 1);
425  s->reorder_pb[5] = s->m.pb;
426  score[1] += lambda * put_bits_count(&s->reorder_pb[5]);
427 
428  dxy = (mx & 1) + 2 * (my & 1);
429 
430  s->hdsp.put_pixels_tab[0][dxy](temp + 16*stride,
431  ref + (mx >> 1) +
432  stride * (my >> 1),
433  stride, 16);
434 
435  score[1] += encode_block(s, src + 16 * x, temp + 16*stride,
436  decoded, stride, 5, 64, lambda, 0);
437  best = score[1] <= score[0];
438 
440  score[2] = s->dsp.sse[0](NULL, src + 16 * x, ref,
441  stride, 16);
442  score[2] += vlc[1] * lambda;
443  if (score[2] < score[best] && mx == 0 && my == 0) {
444  best = 2;
445  s->hdsp.put_pixels_tab[0][0](decoded, ref, stride, 16);
446  for (i = 0; i < 6; i++)
447  count[2][i] = 0;
448  put_bits(&s->pb, vlc[1], vlc[0]);
449  }
450  }
451 
452  if (best == 1) {
453  for (i = 0; i < 6; i++) {
454  count[1][i] = put_bits_count(&s->reorder_pb[i]);
455  flush_put_bits(&s->reorder_pb[i]);
456  }
457  } else {
458  motion_ptr[0] =
459  motion_ptr[1] =
460  motion_ptr[2] =
461  motion_ptr[3] =
462  motion_ptr[0 + 2 * s->m.b8_stride] =
463  motion_ptr[1 + 2 * s->m.b8_stride] =
464  motion_ptr[2 + 2 * s->m.b8_stride] =
465  motion_ptr[3 + 2 * s->m.b8_stride] = 0;
466  }
467  }
468 
469  s->rd_total += score[best];
470 
471  for (i = 5; i >= 0; i--)
472  avpriv_copy_bits(&s->pb, reorder_buffer[best][i],
473  count[best][i]);
474  if (best == 0)
475  s->hdsp.put_pixels_tab[0][0](decoded, temp, stride, 16);
476  }
477  s->m.first_slice_line = 0;
478  }
479  return 0;
480 }
481 
483 {
484  SVQ1EncContext *const s = avctx->priv_data;
485  int i;
486 
487  av_log(avctx, AV_LOG_DEBUG, "RD: %f\n",
488  s->rd_total / (double)(avctx->width * avctx->height *
489  avctx->frame_number));
490 
491  s->m.mb_type = NULL;
492  ff_MPV_common_end(&s->m);
493 
494  av_freep(&s->m.me.scratchpad);
495  av_freep(&s->m.me.map);
496  av_freep(&s->m.me.score_map);
497  av_freep(&s->mb_type);
498  av_freep(&s->dummy);
499  av_freep(&s->scratchbuf);
500 
501  for (i = 0; i < 3; i++) {
502  av_freep(&s->motion_val8[i]);
503  av_freep(&s->motion_val16[i]);
504  }
505 
508  av_frame_free(&avctx->coded_frame);
509 
510  return 0;
511 }
512 
514 {
515  SVQ1EncContext *const s = avctx->priv_data;
516  int ret;
517 
518  ff_dsputil_init(&s->dsp, avctx);
519  ff_hpeldsp_init(&s->hdsp, avctx->flags);
520  ff_mpegvideoencdsp_init(&s->m.mpvencdsp, avctx);
521 
522  avctx->coded_frame = av_frame_alloc();
525  if (!avctx->coded_frame || !s->current_picture || !s->last_picture) {
526  svq1_encode_end(avctx);
527  return AVERROR(ENOMEM);
528  }
529 
530  s->frame_width = avctx->width;
531  s->frame_height = avctx->height;
532 
533  s->y_block_width = (s->frame_width + 15) / 16;
534  s->y_block_height = (s->frame_height + 15) / 16;
535 
536  s->c_block_width = (s->frame_width / 4 + 15) / 16;
537  s->c_block_height = (s->frame_height / 4 + 15) / 16;
538 
539  s->avctx = avctx;
540  s->m.avctx = avctx;
541 
542  if ((ret = ff_MPV_common_init(&s->m)) < 0) {
543  svq1_encode_end(avctx);
544  return ret;
545  }
546 
548  s->m.me.temp =
549  s->m.me.scratchpad = av_mallocz((avctx->width + 64) *
550  2 * 16 * 2 * sizeof(uint8_t));
551  s->m.me.map = av_mallocz(ME_MAP_SIZE * sizeof(uint32_t));
552  s->m.me.score_map = av_mallocz(ME_MAP_SIZE * sizeof(uint32_t));
553  s->mb_type = av_mallocz((s->y_block_width + 1) *
554  s->y_block_height * sizeof(int16_t));
555  s->dummy = av_mallocz((s->y_block_width + 1) *
556  s->y_block_height * sizeof(int32_t));
558 
559  if (ARCH_PPC)
561  if (ARCH_X86)
563 
564  ff_h263_encode_init(&s->m); // mv_penalty
565 
566  return 0;
567 }
568 
570  const AVFrame *pict, int *got_packet)
571 {
572  SVQ1EncContext *const s = avctx->priv_data;
573  AVFrame *const p = avctx->coded_frame;
574  int i, ret;
575 
576  if ((ret = ff_alloc_packet2(avctx, pkt, s->y_block_width * s->y_block_height *
578  return ret;
579 
580  if (avctx->pix_fmt != AV_PIX_FMT_YUV410P) {
581  av_log(avctx, AV_LOG_ERROR, "unsupported pixel format\n");
582  return -1;
583  }
584 
585  if (!s->current_picture->data[0]) {
586  if ((ret = ff_get_buffer(avctx, s->current_picture, 0))< 0 ||
587  (ret = ff_get_buffer(avctx, s->last_picture, 0)) < 0) {
588  return ret;
589  }
590  s->scratchbuf = av_malloc(s->current_picture->linesize[0] * 16 * 3);
591  }
592 
594 
595  init_put_bits(&s->pb, pkt->data, pkt->size);
596 
597  p->pict_type = avctx->gop_size && avctx->frame_number % avctx->gop_size ?
600  p->quality = pict->quality;
601 
603  for (i = 0; i < 3; i++)
604  if (svq1_encode_plane(s, i,
605  pict->data[i],
606  s->last_picture->data[i],
607  s->current_picture->data[i],
608  s->frame_width / (i ? 4 : 1),
609  s->frame_height / (i ? 4 : 1),
610  pict->linesize[i],
611  s->current_picture->linesize[i]) < 0)
612  return -1;
613 
614  // avpriv_align_put_bits(&s->pb);
615  while (put_bits_count(&s->pb) & 31)
616  put_bits(&s->pb, 1, 0);
617 
618  flush_put_bits(&s->pb);
619 
620  pkt->size = put_bits_count(&s->pb) / 8;
621  if (p->pict_type == AV_PICTURE_TYPE_I)
622  pkt->flags |= AV_PKT_FLAG_KEY;
623  *got_packet = 1;
624 
625  return 0;
626 }
627 
629  .name = "svq1",
630  .long_name = NULL_IF_CONFIG_SMALL("Sorenson Vector Quantizer 1 / Sorenson Video 1 / SVQ1"),
631  .type = AVMEDIA_TYPE_VIDEO,
632  .id = AV_CODEC_ID_SVQ1,
633  .priv_data_size = sizeof(SVQ1EncContext),
635  .encode2 = svq1_encode_frame,
637  .pix_fmts = (const enum AVPixelFormat[]) { AV_PIX_FMT_YUV410P,
638  AV_PIX_FMT_NONE },
639 };