dnxhdenc.c
Go to the documentation of this file.
1 /*
2  * VC3/DNxHD encoder
3  * Copyright (c) 2007 Baptiste Coudurier <baptiste dot coudurier at smartjog dot com>
4  * Copyright (c) 2011 MirriAd Ltd
5  *
6  * VC-3 encoder funded by the British Broadcasting Corporation
7  * 10 bit support added by MirriAd Ltd, Joseph Artsimovich <joseph@mirriad.com>
8  *
9  * This file is part of Libav.
10  *
11  * Libav is free software; you can redistribute it and/or
12  * modify it under the terms of the GNU Lesser General Public
13  * License as published by the Free Software Foundation; either
14  * version 2.1 of the License, or (at your option) any later version.
15  *
16  * Libav is distributed in the hope that it will be useful,
17  * but WITHOUT ANY WARRANTY; without even the implied warranty of
18  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
19  * Lesser General Public License for more details.
20  *
21  * You should have received a copy of the GNU Lesser General Public
22  * License along with Libav; if not, write to the Free Software
23  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
24  */
25 
26 //#define DEBUG
27 #define RC_VARIANCE 1 // use variance or ssd for fast rc
28 
29 #include "libavutil/opt.h"
30 #include "avcodec.h"
31 #include "dsputil.h"
32 #include "mpegvideo.h"
33 #include "mpegvideo_common.h"
34 #include "dnxhdenc.h"
35 
36 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
37 #define DNX10BIT_QMAT_SHIFT 18 // The largest value that will not lead to overflow for 10bit samples.
38 
39 static const AVOption options[]={
40  {"nitris_compat", "encode with Avid Nitris compatibility", offsetof(DNXHDEncContext, nitris_compat), AV_OPT_TYPE_INT, {.dbl = 0}, 0, 1, VE},
41 {NULL}
42 };
43 static const AVClass class = { "dnxhd", av_default_item_name, options, LIBAVUTIL_VERSION_INT };
44 
45 #define LAMBDA_FRAC_BITS 10
46 
47 static void dnxhd_8bit_get_pixels_8x4_sym(DCTELEM *restrict block, const uint8_t *pixels, int line_size)
48 {
49  int i;
50  for (i = 0; i < 4; i++) {
51  block[0] = pixels[0]; block[1] = pixels[1];
52  block[2] = pixels[2]; block[3] = pixels[3];
53  block[4] = pixels[4]; block[5] = pixels[5];
54  block[6] = pixels[6]; block[7] = pixels[7];
55  pixels += line_size;
56  block += 8;
57  }
58  memcpy(block, block - 8, sizeof(*block) * 8);
59  memcpy(block + 8, block - 16, sizeof(*block) * 8);
60  memcpy(block + 16, block - 24, sizeof(*block) * 8);
61  memcpy(block + 24, block - 32, sizeof(*block) * 8);
62 }
63 
64 static av_always_inline void dnxhd_10bit_get_pixels_8x4_sym(DCTELEM *restrict block, const uint8_t *pixels, int line_size)
65 {
66  int i;
67 
68  block += 32;
69 
70  for (i = 0; i < 4; i++) {
71  memcpy(block + i * 8, pixels + i * line_size, 8 * sizeof(*block));
72  memcpy(block - (i+1) * 8, pixels + i * line_size, 8 * sizeof(*block));
73  }
74 }
75 
77  int n, int qscale, int *overflow)
78 {
79  const uint8_t *scantable= ctx->intra_scantable.scantable;
80  const int *qmat = ctx->q_intra_matrix[qscale];
81  int last_non_zero = 0;
82  int i;
83 
84  ctx->dsp.fdct(block);
85 
86  // Divide by 4 with rounding, to compensate scaling of DCT coefficients
87  block[0] = (block[0] + 2) >> 2;
88 
89  for (i = 1; i < 64; ++i) {
90  int j = scantable[i];
91  int sign = block[j] >> 31;
92  int level = (block[j] ^ sign) - sign;
93  level = level * qmat[j] >> DNX10BIT_QMAT_SHIFT;
94  block[j] = (level ^ sign) - sign;
95  if (level)
96  last_non_zero = i;
97  }
98 
99  return last_non_zero;
100 }
101 
103 {
104  int i, j, level, run;
105  int max_level = 1<<(ctx->cid_table->bit_depth+2);
106 
107  FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->vlc_codes, max_level*4*sizeof(*ctx->vlc_codes), fail);
108  FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->vlc_bits, max_level*4*sizeof(*ctx->vlc_bits) , fail);
109  FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_codes, 63*2, fail);
110  FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_bits, 63, fail);
111 
112  ctx->vlc_codes += max_level*2;
113  ctx->vlc_bits += max_level*2;
114  for (level = -max_level; level < max_level; level++) {
115  for (run = 0; run < 2; run++) {
116  int index = (level<<1)|run;
117  int sign, offset = 0, alevel = level;
118 
119  MASK_ABS(sign, alevel);
120  if (alevel > 64) {
121  offset = (alevel-1)>>6;
122  alevel -= offset<<6;
123  }
124  for (j = 0; j < 257; j++) {
125  if (ctx->cid_table->ac_level[j] == alevel &&
126  (!offset || (ctx->cid_table->ac_index_flag[j] && offset)) &&
127  (!run || (ctx->cid_table->ac_run_flag [j] && run))) {
128  assert(!ctx->vlc_codes[index]);
129  if (alevel) {
130  ctx->vlc_codes[index] = (ctx->cid_table->ac_codes[j]<<1)|(sign&1);
131  ctx->vlc_bits [index] = ctx->cid_table->ac_bits[j]+1;
132  } else {
133  ctx->vlc_codes[index] = ctx->cid_table->ac_codes[j];
134  ctx->vlc_bits [index] = ctx->cid_table->ac_bits [j];
135  }
136  break;
137  }
138  }
139  assert(!alevel || j < 257);
140  if (offset) {
141  ctx->vlc_codes[index] = (ctx->vlc_codes[index]<<ctx->cid_table->index_bits)|offset;
142  ctx->vlc_bits [index]+= ctx->cid_table->index_bits;
143  }
144  }
145  }
146  for (i = 0; i < 62; i++) {
147  int run = ctx->cid_table->run[i];
148  assert(run < 63);
149  ctx->run_codes[run] = ctx->cid_table->run_codes[i];
150  ctx->run_bits [run] = ctx->cid_table->run_bits[i];
151  }
152  return 0;
153  fail:
154  return -1;
155 }
156 
157 static int dnxhd_init_qmat(DNXHDEncContext *ctx, int lbias, int cbias)
158 {
159  // init first elem to 1 to avoid div by 0 in convert_matrix
160  uint16_t weight_matrix[64] = {1,}; // convert_matrix needs uint16_t*
161  int qscale, i;
162  const uint8_t *luma_weight_table = ctx->cid_table->luma_weight;
163  const uint8_t *chroma_weight_table = ctx->cid_table->chroma_weight;
164 
165  FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l, (ctx->m.avctx->qmax+1) * 64 * sizeof(int), fail);
166  FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c, (ctx->m.avctx->qmax+1) * 64 * sizeof(int), fail);
167  FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l16, (ctx->m.avctx->qmax+1) * 64 * 2 * sizeof(uint16_t), fail);
168  FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c16, (ctx->m.avctx->qmax+1) * 64 * 2 * sizeof(uint16_t), fail);
169 
170  if (ctx->cid_table->bit_depth == 8) {
171  for (i = 1; i < 64; i++) {
172  int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]];
173  weight_matrix[j] = ctx->cid_table->luma_weight[i];
174  }
175  ff_convert_matrix(&ctx->m.dsp, ctx->qmatrix_l, ctx->qmatrix_l16, weight_matrix,
176  ctx->m.intra_quant_bias, 1, ctx->m.avctx->qmax, 1);
177  for (i = 1; i < 64; i++) {
178  int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]];
179  weight_matrix[j] = ctx->cid_table->chroma_weight[i];
180  }
181  ff_convert_matrix(&ctx->m.dsp, ctx->qmatrix_c, ctx->qmatrix_c16, weight_matrix,
182  ctx->m.intra_quant_bias, 1, ctx->m.avctx->qmax, 1);
183 
184  for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
185  for (i = 0; i < 64; i++) {
186  ctx->qmatrix_l [qscale] [i] <<= 2; ctx->qmatrix_c [qscale] [i] <<= 2;
187  ctx->qmatrix_l16[qscale][0][i] <<= 2; ctx->qmatrix_l16[qscale][1][i] <<= 2;
188  ctx->qmatrix_c16[qscale][0][i] <<= 2; ctx->qmatrix_c16[qscale][1][i] <<= 2;
189  }
190  }
191  } else {
192  // 10-bit
193  for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
194  for (i = 1; i < 64; i++) {
195  int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]];
196 
197  // The quantization formula from the VC-3 standard is:
198  // quantized = sign(block[i]) * floor(abs(block[i]/s) * p / (qscale * weight_table[i]))
199  // Where p is 32 for 8-bit samples and 8 for 10-bit ones.
200  // The s factor compensates scaling of DCT coefficients done by the DCT routines,
201  // and therefore is not present in standard. It's 8 for 8-bit samples and 4 for 10-bit ones.
202  // We want values of ctx->qtmatrix_l and ctx->qtmatrix_r to be:
203  // ((1 << DNX10BIT_QMAT_SHIFT) * (p / s)) / (qscale * weight_table[i])
204  // For 10-bit samples, p / s == 2
205  ctx->qmatrix_l[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) / (qscale * luma_weight_table[i]);
206  ctx->qmatrix_c[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) / (qscale * chroma_weight_table[i]);
207  }
208  }
209  }
210 
211  return 0;
212  fail:
213  return -1;
214 }
215 
217 {
218  FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_rc, 8160*ctx->m.avctx->qmax*sizeof(RCEntry), fail);
219  if (ctx->m.avctx->mb_decision != FF_MB_DECISION_RD)
220  FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_cmp, ctx->m.mb_num*sizeof(RCCMPEntry), fail);
221 
222  ctx->frame_bits = (ctx->cid_table->coding_unit_size - 640 - 4 - ctx->min_padding) * 8;
223  ctx->qscale = 1;
224  ctx->lambda = 2<<LAMBDA_FRAC_BITS; // qscale 2
225  return 0;
226  fail:
227  return -1;
228 }
229 
231 {
232  DNXHDEncContext *ctx = avctx->priv_data;
233  int i, index, bit_depth;
234 
235  switch (avctx->pix_fmt) {
236  case PIX_FMT_YUV422P:
237  bit_depth = 8;
238  break;
239  case PIX_FMT_YUV422P10:
240  bit_depth = 10;
241  break;
242  default:
243  av_log(avctx, AV_LOG_ERROR, "pixel format is incompatible with DNxHD\n");
244  return -1;
245  }
246 
247  ctx->cid = ff_dnxhd_find_cid(avctx, bit_depth);
248  if (!ctx->cid) {
249  av_log(avctx, AV_LOG_ERROR, "video parameters incompatible with DNxHD\n");
250  return -1;
251  }
252  av_log(avctx, AV_LOG_DEBUG, "cid %d\n", ctx->cid);
253 
254  index = ff_dnxhd_get_cid_table(ctx->cid);
256 
257  ctx->m.avctx = avctx;
258  ctx->m.mb_intra = 1;
259  ctx->m.h263_aic = 1;
260 
261  avctx->bits_per_raw_sample = ctx->cid_table->bit_depth;
262 
263  dsputil_init(&ctx->m.dsp, avctx);
264  ff_dct_common_init(&ctx->m);
265  if (!ctx->m.dct_quantize)
267 
268  if (ctx->cid_table->bit_depth == 10) {
271  ctx->block_width_l2 = 4;
272  } else {
274  ctx->block_width_l2 = 3;
275  }
276 
277 #if HAVE_MMX
278  ff_dnxhd_init_mmx(ctx);
279 #endif
280 
281  ctx->m.mb_height = (avctx->height + 15) / 16;
282  ctx->m.mb_width = (avctx->width + 15) / 16;
283 
284  if (avctx->flags & CODEC_FLAG_INTERLACED_DCT) {
285  ctx->interlaced = 1;
286  ctx->m.mb_height /= 2;
287  }
288 
289  ctx->m.mb_num = ctx->m.mb_height * ctx->m.mb_width;
290 
292  ctx->m.intra_quant_bias = avctx->intra_quant_bias;
293  if (dnxhd_init_qmat(ctx, ctx->m.intra_quant_bias, 0) < 0) // XXX tune lbias/cbias
294  return -1;
295 
296  // Avid Nitris hardware decoder requires a minimum amount of padding in the coding unit payload
297  if (ctx->nitris_compat)
298  ctx->min_padding = 1600;
299 
300  if (dnxhd_init_vlc(ctx) < 0)
301  return -1;
302  if (dnxhd_init_rc(ctx) < 0)
303  return -1;
304 
305  FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_size, ctx->m.mb_height*sizeof(uint32_t), fail);
306  FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_offs, ctx->m.mb_height*sizeof(uint32_t), fail);
307  FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_bits, ctx->m.mb_num *sizeof(uint16_t), fail);
308  FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_qscale, ctx->m.mb_num *sizeof(uint8_t), fail);
309 
310  ctx->frame.key_frame = 1;
312  ctx->m.avctx->coded_frame = &ctx->frame;
313 
314  if (avctx->thread_count > MAX_THREADS) {
315  av_log(avctx, AV_LOG_ERROR, "too many threads\n");
316  return -1;
317  }
318 
319  ctx->thread[0] = ctx;
320  for (i = 1; i < avctx->thread_count; i++) {
321  ctx->thread[i] = av_malloc(sizeof(DNXHDEncContext));
322  memcpy(ctx->thread[i], ctx, sizeof(DNXHDEncContext));
323  }
324 
325  return 0;
326  fail: //for FF_ALLOCZ_OR_GOTO
327  return -1;
328 }
329 
330 static int dnxhd_write_header(AVCodecContext *avctx, uint8_t *buf)
331 {
332  DNXHDEncContext *ctx = avctx->priv_data;
333  const uint8_t header_prefix[5] = { 0x00,0x00,0x02,0x80,0x01 };
334 
335  memset(buf, 0, 640);
336 
337  memcpy(buf, header_prefix, 5);
338  buf[5] = ctx->interlaced ? ctx->cur_field+2 : 0x01;
339  buf[6] = 0x80; // crc flag off
340  buf[7] = 0xa0; // reserved
341  AV_WB16(buf + 0x18, avctx->height>>ctx->interlaced); // ALPF
342  AV_WB16(buf + 0x1a, avctx->width); // SPL
343  AV_WB16(buf + 0x1d, avctx->height>>ctx->interlaced); // NAL
344 
345  buf[0x21] = ctx->cid_table->bit_depth == 10 ? 0x58 : 0x38;
346  buf[0x22] = 0x88 + (ctx->interlaced<<2);
347  AV_WB32(buf + 0x28, ctx->cid); // CID
348  buf[0x2c] = ctx->interlaced ? 0 : 0x80;
349 
350  buf[0x5f] = 0x01; // UDL
351 
352  buf[0x167] = 0x02; // reserved
353  AV_WB16(buf + 0x16a, ctx->m.mb_height * 4 + 4); // MSIPS
354  buf[0x16d] = ctx->m.mb_height; // Ns
355  buf[0x16f] = 0x10; // reserved
356 
357  ctx->msip = buf + 0x170;
358  return 0;
359 }
360 
362 {
363  int nbits;
364  if (diff < 0) {
365  nbits = av_log2_16bit(-2*diff);
366  diff--;
367  } else {
368  nbits = av_log2_16bit(2*diff);
369  }
370  put_bits(&ctx->m.pb, ctx->cid_table->dc_bits[nbits] + nbits,
371  (ctx->cid_table->dc_codes[nbits]<<nbits) + (diff & ((1 << nbits) - 1)));
372 }
373 
374 static av_always_inline void dnxhd_encode_block(DNXHDEncContext *ctx, DCTELEM *block, int last_index, int n)
375 {
376  int last_non_zero = 0;
377  int slevel, i, j;
378 
379  dnxhd_encode_dc(ctx, block[0] - ctx->m.last_dc[n]);
380  ctx->m.last_dc[n] = block[0];
381 
382  for (i = 1; i <= last_index; i++) {
383  j = ctx->m.intra_scantable.permutated[i];
384  slevel = block[j];
385  if (slevel) {
386  int run_level = i - last_non_zero - 1;
387  int rlevel = (slevel<<1)|!!run_level;
388  put_bits(&ctx->m.pb, ctx->vlc_bits[rlevel], ctx->vlc_codes[rlevel]);
389  if (run_level)
390  put_bits(&ctx->m.pb, ctx->run_bits[run_level], ctx->run_codes[run_level]);
391  last_non_zero = i;
392  }
393  }
394  put_bits(&ctx->m.pb, ctx->vlc_bits[0], ctx->vlc_codes[0]); // EOB
395 }
396 
397 static av_always_inline void dnxhd_unquantize_c(DNXHDEncContext *ctx, DCTELEM *block, int n, int qscale, int last_index)
398 {
399  const uint8_t *weight_matrix;
400  int level;
401  int i;
402 
403  weight_matrix = (n&2) ? ctx->cid_table->chroma_weight : ctx->cid_table->luma_weight;
404 
405  for (i = 1; i <= last_index; i++) {
406  int j = ctx->m.intra_scantable.permutated[i];
407  level = block[j];
408  if (level) {
409  if (level < 0) {
410  level = (1-2*level) * qscale * weight_matrix[i];
411  if (ctx->cid_table->bit_depth == 10) {
412  if (weight_matrix[i] != 8)
413  level += 8;
414  level >>= 4;
415  } else {
416  if (weight_matrix[i] != 32)
417  level += 32;
418  level >>= 6;
419  }
420  level = -level;
421  } else {
422  level = (2*level+1) * qscale * weight_matrix[i];
423  if (ctx->cid_table->bit_depth == 10) {
424  if (weight_matrix[i] != 8)
425  level += 8;
426  level >>= 4;
427  } else {
428  if (weight_matrix[i] != 32)
429  level += 32;
430  level >>= 6;
431  }
432  }
433  block[j] = level;
434  }
435  }
436 }
437 
439 {
440  int score = 0;
441  int i;
442  for (i = 0; i < 64; i++)
443  score += (block[i] - qblock[i]) * (block[i] - qblock[i]);
444  return score;
445 }
446 
448 {
449  int last_non_zero = 0;
450  int bits = 0;
451  int i, j, level;
452  for (i = 1; i <= last_index; i++) {
453  j = ctx->m.intra_scantable.permutated[i];
454  level = block[j];
455  if (level) {
456  int run_level = i - last_non_zero - 1;
457  bits += ctx->vlc_bits[(level<<1)|!!run_level]+ctx->run_bits[run_level];
458  last_non_zero = i;
459  }
460  }
461  return bits;
462 }
463 
464 static av_always_inline void dnxhd_get_blocks(DNXHDEncContext *ctx, int mb_x, int mb_y)
465 {
466  const int bs = ctx->block_width_l2;
467  const int bw = 1 << bs;
468  const uint8_t *ptr_y = ctx->thread[0]->src[0] + ((mb_y << 4) * ctx->m.linesize) + (mb_x << bs+1);
469  const uint8_t *ptr_u = ctx->thread[0]->src[1] + ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs);
470  const uint8_t *ptr_v = ctx->thread[0]->src[2] + ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs);
471  DSPContext *dsp = &ctx->m.dsp;
472 
473  dsp->get_pixels(ctx->blocks[0], ptr_y, ctx->m.linesize);
474  dsp->get_pixels(ctx->blocks[1], ptr_y + bw, ctx->m.linesize);
475  dsp->get_pixels(ctx->blocks[2], ptr_u, ctx->m.uvlinesize);
476  dsp->get_pixels(ctx->blocks[3], ptr_v, ctx->m.uvlinesize);
477 
478  if (mb_y+1 == ctx->m.mb_height && ctx->m.avctx->height == 1080) {
479  if (ctx->interlaced) {
480  ctx->get_pixels_8x4_sym(ctx->blocks[4], ptr_y + ctx->dct_y_offset, ctx->m.linesize);
481  ctx->get_pixels_8x4_sym(ctx->blocks[5], ptr_y + ctx->dct_y_offset + bw, ctx->m.linesize);
482  ctx->get_pixels_8x4_sym(ctx->blocks[6], ptr_u + ctx->dct_uv_offset, ctx->m.uvlinesize);
483  ctx->get_pixels_8x4_sym(ctx->blocks[7], ptr_v + ctx->dct_uv_offset, ctx->m.uvlinesize);
484  } else {
485  dsp->clear_block(ctx->blocks[4]);
486  dsp->clear_block(ctx->blocks[5]);
487  dsp->clear_block(ctx->blocks[6]);
488  dsp->clear_block(ctx->blocks[7]);
489  }
490  } else {
491  dsp->get_pixels(ctx->blocks[4], ptr_y + ctx->dct_y_offset, ctx->m.linesize);
492  dsp->get_pixels(ctx->blocks[5], ptr_y + ctx->dct_y_offset + bw, ctx->m.linesize);
493  dsp->get_pixels(ctx->blocks[6], ptr_u + ctx->dct_uv_offset, ctx->m.uvlinesize);
494  dsp->get_pixels(ctx->blocks[7], ptr_v + ctx->dct_uv_offset, ctx->m.uvlinesize);
495  }
496 }
497 
499 {
500  if (i&2) {
501  ctx->m.q_intra_matrix16 = ctx->qmatrix_c16;
502  ctx->m.q_intra_matrix = ctx->qmatrix_c;
503  return 1 + (i&1);
504  } else {
505  ctx->m.q_intra_matrix16 = ctx->qmatrix_l16;
506  ctx->m.q_intra_matrix = ctx->qmatrix_l;
507  return 0;
508  }
509 }
510 
511 static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg, int jobnr, int threadnr)
512 {
513  DNXHDEncContext *ctx = avctx->priv_data;
514  int mb_y = jobnr, mb_x;
515  int qscale = ctx->qscale;
517  ctx = ctx->thread[threadnr];
518 
519  ctx->m.last_dc[0] =
520  ctx->m.last_dc[1] =
521  ctx->m.last_dc[2] = 1 << (ctx->cid_table->bit_depth + 2);
522 
523  for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
524  unsigned mb = mb_y * ctx->m.mb_width + mb_x;
525  int ssd = 0;
526  int ac_bits = 0;
527  int dc_bits = 0;
528  int i;
529 
530  dnxhd_get_blocks(ctx, mb_x, mb_y);
531 
532  for (i = 0; i < 8; i++) {
533  DCTELEM *src_block = ctx->blocks[i];
534  int overflow, nbits, diff, last_index;
535  int n = dnxhd_switch_matrix(ctx, i);
536 
537  memcpy(block, src_block, 64*sizeof(*block));
538  last_index = ctx->m.dct_quantize(&ctx->m, block, i, qscale, &overflow);
539  ac_bits += dnxhd_calc_ac_bits(ctx, block, last_index);
540 
541  diff = block[0] - ctx->m.last_dc[n];
542  if (diff < 0) nbits = av_log2_16bit(-2*diff);
543  else nbits = av_log2_16bit( 2*diff);
544 
545  assert(nbits < ctx->cid_table->bit_depth + 4);
546  dc_bits += ctx->cid_table->dc_bits[nbits] + nbits;
547 
548  ctx->m.last_dc[n] = block[0];
549 
550  if (avctx->mb_decision == FF_MB_DECISION_RD || !RC_VARIANCE) {
551  dnxhd_unquantize_c(ctx, block, i, qscale, last_index);
552  ctx->m.dsp.idct(block);
553  ssd += dnxhd_ssd_block(block, src_block);
554  }
555  }
556  ctx->mb_rc[qscale][mb].ssd = ssd;
557  ctx->mb_rc[qscale][mb].bits = ac_bits+dc_bits+12+8*ctx->vlc_bits[0];
558  }
559  return 0;
560 }
561 
562 static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg, int jobnr, int threadnr)
563 {
564  DNXHDEncContext *ctx = avctx->priv_data;
565  int mb_y = jobnr, mb_x;
566  ctx = ctx->thread[threadnr];
567  init_put_bits(&ctx->m.pb, (uint8_t *)arg + 640 + ctx->slice_offs[jobnr], ctx->slice_size[jobnr]);
568 
569  ctx->m.last_dc[0] =
570  ctx->m.last_dc[1] =
571  ctx->m.last_dc[2] = 1 << (ctx->cid_table->bit_depth + 2);
572  for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
573  unsigned mb = mb_y * ctx->m.mb_width + mb_x;
574  int qscale = ctx->mb_qscale[mb];
575  int i;
576 
577  put_bits(&ctx->m.pb, 12, qscale<<1);
578 
579  dnxhd_get_blocks(ctx, mb_x, mb_y);
580 
581  for (i = 0; i < 8; i++) {
582  DCTELEM *block = ctx->blocks[i];
583  int overflow, n = dnxhd_switch_matrix(ctx, i);
584  int last_index = ctx->m.dct_quantize(&ctx->m, block, i,
585  qscale, &overflow);
586  //START_TIMER;
587  dnxhd_encode_block(ctx, block, last_index, n);
588  //STOP_TIMER("encode_block");
589  }
590  }
591  if (put_bits_count(&ctx->m.pb)&31)
592  put_bits(&ctx->m.pb, 32-(put_bits_count(&ctx->m.pb)&31), 0);
593  flush_put_bits(&ctx->m.pb);
594  return 0;
595 }
596 
598 {
599  int mb_y, mb_x;
600  int offset = 0;
601  for (mb_y = 0; mb_y < ctx->m.mb_height; mb_y++) {
602  int thread_size;
603  ctx->slice_offs[mb_y] = offset;
604  ctx->slice_size[mb_y] = 0;
605  for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
606  unsigned mb = mb_y * ctx->m.mb_width + mb_x;
607  ctx->slice_size[mb_y] += ctx->mb_bits[mb];
608  }
609  ctx->slice_size[mb_y] = (ctx->slice_size[mb_y]+31)&~31;
610  ctx->slice_size[mb_y] >>= 3;
611  thread_size = ctx->slice_size[mb_y];
612  offset += thread_size;
613  }
614 }
615 
616 static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg, int jobnr, int threadnr)
617 {
618  DNXHDEncContext *ctx = avctx->priv_data;
619  int mb_y = jobnr, mb_x;
620  ctx = ctx->thread[threadnr];
621  if (ctx->cid_table->bit_depth == 8) {
622  uint8_t *pix = ctx->thread[0]->src[0] + ((mb_y<<4) * ctx->m.linesize);
623  for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x, pix += 16) {
624  unsigned mb = mb_y * ctx->m.mb_width + mb_x;
625  int sum = ctx->m.dsp.pix_sum(pix, ctx->m.linesize);
626  int varc = (ctx->m.dsp.pix_norm1(pix, ctx->m.linesize) - (((unsigned)sum*sum)>>8)+128)>>8;
627  ctx->mb_cmp[mb].value = varc;
628  ctx->mb_cmp[mb].mb = mb;
629  }
630  } else { // 10-bit
631  int const linesize = ctx->m.linesize >> 1;
632  for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x) {
633  uint16_t *pix = (uint16_t*)ctx->thread[0]->src[0] + ((mb_y << 4) * linesize) + (mb_x << 4);
634  unsigned mb = mb_y * ctx->m.mb_width + mb_x;
635  int sum = 0;
636  int sqsum = 0;
637  int mean, sqmean;
638  int i, j;
639  // Macroblocks are 16x16 pixels, unlike DCT blocks which are 8x8.
640  for (i = 0; i < 16; ++i) {
641  for (j = 0; j < 16; ++j) {
642  // Turn 16-bit pixels into 10-bit ones.
643  int const sample = (unsigned)pix[j] >> 6;
644  sum += sample;
645  sqsum += sample * sample;
646  // 2^10 * 2^10 * 16 * 16 = 2^28, which is less than INT_MAX
647  }
648  pix += linesize;
649  }
650  mean = sum >> 8; // 16*16 == 2^8
651  sqmean = sqsum >> 8;
652  ctx->mb_cmp[mb].value = sqmean - mean * mean;
653  ctx->mb_cmp[mb].mb = mb;
654  }
655  }
656  return 0;
657 }
658 
660 {
661  int lambda, up_step, down_step;
662  int last_lower = INT_MAX, last_higher = 0;
663  int x, y, q;
664 
665  for (q = 1; q < avctx->qmax; q++) {
666  ctx->qscale = q;
667  avctx->execute2(avctx, dnxhd_calc_bits_thread, NULL, NULL, ctx->m.mb_height);
668  }
669  up_step = down_step = 2<<LAMBDA_FRAC_BITS;
670  lambda = ctx->lambda;
671 
672  for (;;) {
673  int bits = 0;
674  int end = 0;
675  if (lambda == last_higher) {
676  lambda++;
677  end = 1; // need to set final qscales/bits
678  }
679  for (y = 0; y < ctx->m.mb_height; y++) {
680  for (x = 0; x < ctx->m.mb_width; x++) {
681  unsigned min = UINT_MAX;
682  int qscale = 1;
683  int mb = y*ctx->m.mb_width+x;
684  for (q = 1; q < avctx->qmax; q++) {
685  unsigned score = ctx->mb_rc[q][mb].bits*lambda+
686  ((unsigned)ctx->mb_rc[q][mb].ssd<<LAMBDA_FRAC_BITS);
687  if (score < min) {
688  min = score;
689  qscale = q;
690  }
691  }
692  bits += ctx->mb_rc[qscale][mb].bits;
693  ctx->mb_qscale[mb] = qscale;
694  ctx->mb_bits[mb] = ctx->mb_rc[qscale][mb].bits;
695  }
696  bits = (bits+31)&~31; // padding
697  if (bits > ctx->frame_bits)
698  break;
699  }
700  //av_dlog(ctx->m.avctx, "lambda %d, up %u, down %u, bits %d, frame %d\n",
701  // lambda, last_higher, last_lower, bits, ctx->frame_bits);
702  if (end) {
703  if (bits > ctx->frame_bits)
704  return -1;
705  break;
706  }
707  if (bits < ctx->frame_bits) {
708  last_lower = FFMIN(lambda, last_lower);
709  if (last_higher != 0)
710  lambda = (lambda+last_higher)>>1;
711  else
712  lambda -= down_step;
713  down_step = FFMIN((int64_t)down_step*5, INT_MAX);
714  up_step = 1<<LAMBDA_FRAC_BITS;
715  lambda = FFMAX(1, lambda);
716  if (lambda == last_lower)
717  break;
718  } else {
719  last_higher = FFMAX(lambda, last_higher);
720  if (last_lower != INT_MAX)
721  lambda = (lambda+last_lower)>>1;
722  else if ((int64_t)lambda + up_step > INT_MAX)
723  return -1;
724  else
725  lambda += up_step;
726  up_step = FFMIN((int64_t)up_step*5, INT_MAX);
727  down_step = 1<<LAMBDA_FRAC_BITS;
728  }
729  }
730  //av_dlog(ctx->m.avctx, "out lambda %d\n", lambda);
731  ctx->lambda = lambda;
732  return 0;
733 }
734 
736 {
737  int bits = 0;
738  int up_step = 1;
739  int down_step = 1;
740  int last_higher = 0;
741  int last_lower = INT_MAX;
742  int qscale;
743  int x, y;
744 
745  qscale = ctx->qscale;
746  for (;;) {
747  bits = 0;
748  ctx->qscale = qscale;
749  // XXX avoid recalculating bits
751  for (y = 0; y < ctx->m.mb_height; y++) {
752  for (x = 0; x < ctx->m.mb_width; x++)
753  bits += ctx->mb_rc[qscale][y*ctx->m.mb_width+x].bits;
754  bits = (bits+31)&~31; // padding
755  if (bits > ctx->frame_bits)
756  break;
757  }
758  //av_dlog(ctx->m.avctx, "%d, qscale %d, bits %d, frame %d, higher %d, lower %d\n",
759  // ctx->m.avctx->frame_number, qscale, bits, ctx->frame_bits, last_higher, last_lower);
760  if (bits < ctx->frame_bits) {
761  if (qscale == 1)
762  return 1;
763  if (last_higher == qscale - 1) {
764  qscale = last_higher;
765  break;
766  }
767  last_lower = FFMIN(qscale, last_lower);
768  if (last_higher != 0)
769  qscale = (qscale+last_higher)>>1;
770  else
771  qscale -= down_step++;
772  if (qscale < 1)
773  qscale = 1;
774  up_step = 1;
775  } else {
776  if (last_lower == qscale + 1)
777  break;
778  last_higher = FFMAX(qscale, last_higher);
779  if (last_lower != INT_MAX)
780  qscale = (qscale+last_lower)>>1;
781  else
782  qscale += up_step++;
783  down_step = 1;
784  if (qscale >= ctx->m.avctx->qmax)
785  return -1;
786  }
787  }
788  //av_dlog(ctx->m.avctx, "out qscale %d\n", qscale);
789  ctx->qscale = qscale;
790  return 0;
791 }
792 
793 #define BUCKET_BITS 8
794 #define RADIX_PASSES 4
795 #define NBUCKETS (1 << BUCKET_BITS)
796 
797 static inline int get_bucket(int value, int shift)
798 {
799  value >>= shift;
800  value &= NBUCKETS - 1;
801  return NBUCKETS - 1 - value;
802 }
803 
804 static void radix_count(const RCCMPEntry *data, int size, int buckets[RADIX_PASSES][NBUCKETS])
805 {
806  int i, j;
807  memset(buckets, 0, sizeof(buckets[0][0]) * RADIX_PASSES * NBUCKETS);
808  for (i = 0; i < size; i++) {
809  int v = data[i].value;
810  for (j = 0; j < RADIX_PASSES; j++) {
811  buckets[j][get_bucket(v, 0)]++;
812  v >>= BUCKET_BITS;
813  }
814  assert(!v);
815  }
816  for (j = 0; j < RADIX_PASSES; j++) {
817  int offset = size;
818  for (i = NBUCKETS - 1; i >= 0; i--)
819  buckets[j][i] = offset -= buckets[j][i];
820  assert(!buckets[j][0]);
821  }
822 }
823 
824 static void radix_sort_pass(RCCMPEntry *dst, const RCCMPEntry *data, int size, int buckets[NBUCKETS], int pass)
825 {
826  int shift = pass * BUCKET_BITS;
827  int i;
828  for (i = 0; i < size; i++) {
829  int v = get_bucket(data[i].value, shift);
830  int pos = buckets[v]++;
831  dst[pos] = data[i];
832  }
833 }
834 
835 static void radix_sort(RCCMPEntry *data, int size)
836 {
837  int buckets[RADIX_PASSES][NBUCKETS];
838  RCCMPEntry *tmp = av_malloc(sizeof(*tmp) * size);
839  radix_count(data, size, buckets);
840  radix_sort_pass(tmp, data, size, buckets[0], 0);
841  radix_sort_pass(data, tmp, size, buckets[1], 1);
842  if (buckets[2][NBUCKETS - 1] || buckets[3][NBUCKETS - 1]) {
843  radix_sort_pass(tmp, data, size, buckets[2], 2);
844  radix_sort_pass(data, tmp, size, buckets[3], 3);
845  }
846  av_free(tmp);
847 }
848 
850 {
851  int max_bits = 0;
852  int ret, x, y;
853  if ((ret = dnxhd_find_qscale(ctx)) < 0)
854  return -1;
855  for (y = 0; y < ctx->m.mb_height; y++) {
856  for (x = 0; x < ctx->m.mb_width; x++) {
857  int mb = y*ctx->m.mb_width+x;
858  int delta_bits;
859  ctx->mb_qscale[mb] = ctx->qscale;
860  ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale][mb].bits;
861  max_bits += ctx->mb_rc[ctx->qscale][mb].bits;
862  if (!RC_VARIANCE) {
863  delta_bits = ctx->mb_rc[ctx->qscale][mb].bits-ctx->mb_rc[ctx->qscale+1][mb].bits;
864  ctx->mb_cmp[mb].mb = mb;
865  ctx->mb_cmp[mb].value = delta_bits ?
866  ((ctx->mb_rc[ctx->qscale][mb].ssd-ctx->mb_rc[ctx->qscale+1][mb].ssd)*100)/delta_bits
867  : INT_MIN; //avoid increasing qscale
868  }
869  }
870  max_bits += 31; //worst padding
871  }
872  if (!ret) {
873  if (RC_VARIANCE)
874  avctx->execute2(avctx, dnxhd_mb_var_thread, NULL, NULL, ctx->m.mb_height);
875  radix_sort(ctx->mb_cmp, ctx->m.mb_num);
876  for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) {
877  int mb = ctx->mb_cmp[x].mb;
878  max_bits -= ctx->mb_rc[ctx->qscale][mb].bits - ctx->mb_rc[ctx->qscale+1][mb].bits;
879  ctx->mb_qscale[mb] = ctx->qscale+1;
880  ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale+1][mb].bits;
881  }
882  }
883  return 0;
884 }
885 
886 static void dnxhd_load_picture(DNXHDEncContext *ctx, const AVFrame *frame)
887 {
888  int i;
889 
890  for (i = 0; i < 3; i++) {
891  ctx->frame.data[i] = frame->data[i];
892  ctx->frame.linesize[i] = frame->linesize[i];
893  }
894 
895  for (i = 0; i < ctx->m.avctx->thread_count; i++) {
896  ctx->thread[i]->m.linesize = ctx->frame.linesize[0]<<ctx->interlaced;
897  ctx->thread[i]->m.uvlinesize = ctx->frame.linesize[1]<<ctx->interlaced;
898  ctx->thread[i]->dct_y_offset = ctx->m.linesize *8;
899  ctx->thread[i]->dct_uv_offset = ctx->m.uvlinesize*8;
900  }
901 
903  ctx->cur_field = frame->interlaced_frame && !frame->top_field_first;
904 }
905 
906 static int dnxhd_encode_picture(AVCodecContext *avctx, unsigned char *buf, int buf_size, void *data)
907 {
908  DNXHDEncContext *ctx = avctx->priv_data;
909  int first_field = 1;
910  int offset, i, ret;
911 
912  if (buf_size < ctx->cid_table->frame_size) {
913  av_log(avctx, AV_LOG_ERROR, "output buffer is too small to compress picture\n");
914  return -1;
915  }
916 
917  dnxhd_load_picture(ctx, data);
918 
919  encode_coding_unit:
920  for (i = 0; i < 3; i++) {
921  ctx->src[i] = ctx->frame.data[i];
922  if (ctx->interlaced && ctx->cur_field)
923  ctx->src[i] += ctx->frame.linesize[i];
924  }
925 
926  dnxhd_write_header(avctx, buf);
927 
928  if (avctx->mb_decision == FF_MB_DECISION_RD)
929  ret = dnxhd_encode_rdo(avctx, ctx);
930  else
931  ret = dnxhd_encode_fast(avctx, ctx);
932  if (ret < 0) {
933  av_log(avctx, AV_LOG_ERROR,
934  "picture could not fit ratecontrol constraints, increase qmax\n");
935  return -1;
936  }
937 
939 
940  offset = 0;
941  for (i = 0; i < ctx->m.mb_height; i++) {
942  AV_WB32(ctx->msip + i * 4, offset);
943  offset += ctx->slice_size[i];
944  assert(!(ctx->slice_size[i] & 3));
945  }
946 
947  avctx->execute2(avctx, dnxhd_encode_thread, buf, NULL, ctx->m.mb_height);
948 
949  assert(640 + offset + 4 <= ctx->cid_table->coding_unit_size);
950  memset(buf + 640 + offset, 0, ctx->cid_table->coding_unit_size - 4 - offset - 640);
951 
952  AV_WB32(buf + ctx->cid_table->coding_unit_size - 4, 0x600DC0DE); // EOF
953 
954  if (ctx->interlaced && first_field) {
955  first_field = 0;
956  ctx->cur_field ^= 1;
957  buf += ctx->cid_table->coding_unit_size;
958  buf_size -= ctx->cid_table->coding_unit_size;
959  goto encode_coding_unit;
960  }
961 
962  ctx->frame.quality = ctx->qscale*FF_QP2LAMBDA;
963 
964  return ctx->cid_table->frame_size;
965 }
966 
968 {
969  DNXHDEncContext *ctx = avctx->priv_data;
970  int max_level = 1<<(ctx->cid_table->bit_depth+2);
971  int i;
972 
973  av_free(ctx->vlc_codes-max_level*2);
974  av_free(ctx->vlc_bits -max_level*2);
975  av_freep(&ctx->run_codes);
976  av_freep(&ctx->run_bits);
977 
978  av_freep(&ctx->mb_bits);
979  av_freep(&ctx->mb_qscale);
980  av_freep(&ctx->mb_rc);
981  av_freep(&ctx->mb_cmp);
982  av_freep(&ctx->slice_size);
983  av_freep(&ctx->slice_offs);
984 
985  av_freep(&ctx->qmatrix_c);
986  av_freep(&ctx->qmatrix_l);
987  av_freep(&ctx->qmatrix_c16);
988  av_freep(&ctx->qmatrix_l16);
989 
990  for (i = 1; i < avctx->thread_count; i++)
991  av_freep(&ctx->thread[i]);
992 
993  return 0;
994 }
995 
997  .name = "dnxhd",
998  .type = AVMEDIA_TYPE_VIDEO,
999  .id = CODEC_ID_DNXHD,
1000  .priv_data_size = sizeof(DNXHDEncContext),
1002  .encode = dnxhd_encode_picture,
1004  .capabilities = CODEC_CAP_SLICE_THREADS,
1005  .pix_fmts = (const enum PixelFormat[]){PIX_FMT_YUV422P, PIX_FMT_YUV422P10, PIX_FMT_NONE},
1006  .long_name = NULL_IF_CONFIG_SMALL("VC3/DNxHD"),
1007  .priv_class = &class,
1008 };