libavcodec/dnxhdenc.c

   1 /*
   2  * VC3/DNxHD encoder
   3  * Copyright (c) 2007 Baptiste Coudurier <baptiste dot coudurier at smartjog dot com>
   4  *
   5  * VC-3 encoder funded by the British Broadcasting Corporation
   6  *
   7  * This file is part of FFmpeg.
   8  *
   9  * FFmpeg is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public
  11  * License as published by the Free Software Foundation; either
  12  * version 2.1 of the License, or (at your option) any later version.
  13  *
  14  * FFmpeg is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with FFmpeg; if not, write to the Free Software
  21  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  22  */
  23
  24 //#define DEBUG
  25 #define RC_VARIANCE 1 // use variance or ssd for fast rc
  26
  27 #include "avcodec.h"
  28 #include "dsputil.h"
  29 #include "mpegvideo.h"
  30 #include "dnxhddata.h"
  31
  32 typedef struct {
  33     uint16_t mb;
  34     int value;
  35 } RCCMPEntry;
  36
  37 typedef struct {
  38     int ssd;
  39     int bits;
  40 } RCEntry;
  41
  42 int dct_quantize_c(MpegEncContext *s, DCTELEM *block, int n, int qscale, int *overflow);
  43
  44 typedef struct DNXHDEncContext {
  45     MpegEncContext m; ///< Used for quantization dsp functions
  46
  47     AVFrame frame;
  48     int cid;
  49     const CIDEntry *cid_table;
  50     uint8_t *msip; ///< Macroblock Scan Indices Payload
  51     uint32_t *slice_size;
  52
  53     struct DNXHDEncContext *thread[MAX_THREADS];
  54
  55     unsigned dct_y_offset;
  56     unsigned dct_uv_offset;
  57     int interlaced;
  58     int cur_field;
  59
  60     DECLARE_ALIGNED_16(DCTELEM, blocks[8][64]);
  61
  62     int      (*qmatrix_c)     [64];
  63     int      (*qmatrix_l)     [64];
  64     uint16_t (*qmatrix_l16)[2][64];
  65     uint16_t (*qmatrix_c16)[2][64];
  66
  67     unsigned frame_bits;
  68     uint8_t *src[3];
  69
  70     uint16_t *table_vlc_codes;
  71     uint8_t  *table_vlc_bits;
  72     uint16_t *table_run_codes;
  73     uint8_t  *table_run_bits;
  74
  75     /** Rate control */
  76     unsigned slice_bits;
  77     unsigned qscale;
  78     unsigned lambda;
  79
  80     unsigned thread_size;
  81
  82     uint16_t *mb_bits;
  83     uint8_t  *mb_qscale;
  84
  85     RCCMPEntry *mb_cmp;
  86     RCEntry   (*mb_rc)[8160];
  87 } DNXHDEncContext;
  88
  89 #define LAMBDA_FRAC_BITS 10
  90
  91 static int dnxhd_init_vlc(DNXHDEncContext *ctx)
  92 {
  93     int i;
  94
  95     CHECKED_ALLOCZ(ctx->table_vlc_codes, 449*2);
  96     CHECKED_ALLOCZ(ctx->table_vlc_bits,    449);
  97     CHECKED_ALLOCZ(ctx->table_run_codes,  63*2);
  98     CHECKED_ALLOCZ(ctx->table_run_bits,     63);
  99
 100     for (i = 0; i < 257; i++) {
 101         int level = ctx->cid_table->ac_level[i] +
 102             (ctx->cid_table->ac_run_flag[i] << 7) + (ctx->cid_table->ac_index_flag[i] << 8);
 103         assert(level < 449);
 104         if (ctx->cid_table->ac_level[i] == 64 && ctx->cid_table->ac_index_flag[i])
 105             level -= 64; // use 0+(1<<8) level
 106         ctx->table_vlc_codes[level] = ctx->cid_table->ac_codes[i];
 107         ctx->table_vlc_bits [level] = ctx->cid_table->ac_bits[i];
 108     }
 109     for (i = 0; i < 62; i++) {
 110         int run = ctx->cid_table->run[i];
 111         assert(run < 63);
 112         ctx->table_run_codes[run] = ctx->cid_table->run_codes[i];
 113         ctx->table_run_bits [run] = ctx->cid_table->run_bits[i];
 114     }
 115     return 0;
 116  fail:
 117     return -1;
 118 }
 119
 120 static int dnxhd_init_qmat(DNXHDEncContext *ctx, int lbias, int cbias)
 121 {
 122     // init first elem to 1 to avoid div by 0 in convert_matrix
 123     uint16_t weight_matrix[64] = {1,}; // convert_matrix needs uint16_t*
 124     int qscale, i;
 125
 126     CHECKED_ALLOCZ(ctx->qmatrix_l,   (ctx->m.avctx->qmax+1) * 64 * sizeof(int));
 127     CHECKED_ALLOCZ(ctx->qmatrix_c,   (ctx->m.avctx->qmax+1) * 64 * sizeof(int));
 128     CHECKED_ALLOCZ(ctx->qmatrix_l16, (ctx->m.avctx->qmax+1) * 64 * 2 * sizeof(uint16_t));
 129     CHECKED_ALLOCZ(ctx->qmatrix_c16, (ctx->m.avctx->qmax+1) * 64 * 2 * sizeof(uint16_t));
 130
 131     for (i = 1; i < 64; i++) {
 132         int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]];
 133         weight_matrix[j] = ctx->cid_table->luma_weight[i];
 134     }
 135     ff_convert_matrix(&ctx->m.dsp, ctx->qmatrix_l, ctx->qmatrix_l16, weight_matrix,
 136                       ctx->m.intra_quant_bias, 1, ctx->m.avctx->qmax, 1);
 137     for (i = 1; i < 64; i++) {
 138         int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]];
 139         weight_matrix[j] = ctx->cid_table->chroma_weight[i];
 140     }
 141     ff_convert_matrix(&ctx->m.dsp, ctx->qmatrix_c, ctx->qmatrix_c16, weight_matrix,
 142                       ctx->m.intra_quant_bias, 1, ctx->m.avctx->qmax, 1);
 143     for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
 144         for (i = 0; i < 64; i++) {
 145             ctx->qmatrix_l  [qscale]   [i] <<= 2; ctx->qmatrix_c  [qscale]   [i] <<= 2;
 146             ctx->qmatrix_l16[qscale][0][i] <<= 2; ctx->qmatrix_l16[qscale][1][i] <<= 2;
 147             ctx->qmatrix_c16[qscale][0][i] <<= 2; ctx->qmatrix_c16[qscale][1][i] <<= 2;
 148         }
 149     }
 150     return 0;
 151  fail:
 152     return -1;
 153 }
 154
 155 static int dnxhd_init_rc(DNXHDEncContext *ctx)
 156 {
 157     CHECKED_ALLOCZ(ctx->mb_rc, 8160*ctx->m.avctx->qmax*sizeof(RCEntry));
 158     if (ctx->m.avctx->mb_decision != FF_MB_DECISION_RD)
 159         CHECKED_ALLOCZ(ctx->mb_cmp, ctx->m.mb_num*sizeof(RCCMPEntry));
 160
 161     ctx->frame_bits = (ctx->cid_table->coding_unit_size - 640 - 4) * 8;
 162     ctx->qscale = 1;
 163     ctx->lambda = 2<<LAMBDA_FRAC_BITS; // qscale 2
 164     return 0;
 165  fail:
 166     return -1;
 167 }
 168
 169 static int dnxhd_encode_init(AVCodecContext *avctx)
 170 {
 171     DNXHDEncContext *ctx = avctx->priv_data;
 172     int i, index;
 173
 174     ctx->cid = ff_dnxhd_find_cid(avctx);
 175     if (!ctx->cid || avctx->pix_fmt != PIX_FMT_YUV422P) {
 176         av_log(avctx, AV_LOG_ERROR, "video parameters incompatible with DNxHD\n");
 177         return -1;
 178     }
 179     av_log(avctx, AV_LOG_DEBUG, "cid %d\n", ctx->cid);
 180
 181     index = ff_dnxhd_get_cid_table(ctx->cid);
 182     ctx->cid_table = &ff_dnxhd_cid_table[index];
 183
 184     ctx->m.avctx = avctx;
 185     ctx->m.mb_intra = 1;
 186     ctx->m.h263_aic = 1;
 187
 188     dsputil_init(&ctx->m.dsp, avctx);
 189     ff_dct_common_init(&ctx->m);
 190     if (!ctx->m.dct_quantize)
 191         ctx->m.dct_quantize = dct_quantize_c;
 192
 193     ctx->m.mb_height = (avctx->height + 15) / 16;
 194     ctx->m.mb_width  = (avctx->width  + 15) / 16;
 195
 196     if (avctx->flags & CODEC_FLAG_INTERLACED_DCT) {
 197         ctx->interlaced = 1;
 198         ctx->m.mb_height /= 2;
 199     }
 200
 201     ctx->m.mb_num = ctx->m.mb_height * ctx->m.mb_width;
 202
 203     if (avctx->intra_quant_bias != FF_DEFAULT_QUANT_BIAS)
 204         ctx->m.intra_quant_bias = avctx->intra_quant_bias;
 205     if (dnxhd_init_qmat(ctx, ctx->m.intra_quant_bias, 0) < 0) // XXX tune lbias/cbias
 206         return -1;
 207
 208     if (dnxhd_init_vlc(ctx) < 0)
 209         return -1;
 210     if (dnxhd_init_rc(ctx) < 0)
 211         return -1;
 212
 213     CHECKED_ALLOCZ(ctx->slice_size, ctx->m.mb_height*sizeof(uint32_t));
 214     CHECKED_ALLOCZ(ctx->mb_bits,    ctx->m.mb_num   *sizeof(uint16_t));
 215     CHECKED_ALLOCZ(ctx->mb_qscale,  ctx->m.mb_num   *sizeof(uint8_t));
 216
 217     ctx->frame.key_frame = 1;
 218     ctx->frame.pict_type = FF_I_TYPE;
 219     ctx->m.avctx->coded_frame = &ctx->frame;
 220
 221     if (avctx->thread_count > MAX_THREADS || (avctx->thread_count > ctx->m.mb_height)) {
 222         av_log(avctx, AV_LOG_ERROR, "too many threads\n");
 223         return -1;
 224     }
 225
 226     ctx->thread[0] = ctx;
 227     for (i = 1; i < avctx->thread_count; i++) {
 228         ctx->thread[i] =  av_malloc(sizeof(DNXHDEncContext));
 229         memcpy(ctx->thread[i], ctx, sizeof(DNXHDEncContext));
 230     }
 231
 232     for (i = 0; i < avctx->thread_count; i++) {
 233         ctx->thread[i]->m.start_mb_y = (ctx->m.mb_height*(i  ) + avctx->thread_count/2) / avctx->thread_count;
 234         ctx->thread[i]->m.end_mb_y   = (ctx->m.mb_height*(i+1) + avctx->thread_count/2) / avctx->thread_count;
 235     }
 236
 237     return 0;
 238  fail: //for CHECKED_ALLOCZ
 239     return -1;
 240 }
 241
 242 static int dnxhd_write_header(AVCodecContext *avctx, uint8_t *buf)
 243 {
 244     DNXHDEncContext *ctx = avctx->priv_data;
 245     const uint8_t header_prefix[5] = { 0x00,0x00,0x02,0x80,0x01 };
 246
 247     memcpy(buf, header_prefix, 5);
 248     buf[5] = ctx->interlaced ? ctx->cur_field+2 : 0x01;
 249     buf[6] = 0x80; // crc flag off
 250     buf[7] = 0xa0; // reserved
 251     AV_WB16(buf + 0x18, avctx->height); // ALPF
 252     AV_WB16(buf + 0x1a, avctx->width);  // SPL
 253     AV_WB16(buf + 0x1d, avctx->height); // NAL
 254
 255     buf[0x21] = 0x38; // FIXME 8 bit per comp
 256     buf[0x22] = 0x88 + (ctx->frame.interlaced_frame<<2);
 257     AV_WB32(buf + 0x28, ctx->cid); // CID
 258     buf[0x2c] = ctx->interlaced ? 0 : 0x80;
 259
 260     buf[0x5f] = 0x01; // UDL
 261
 262     buf[0x167] = 0x02; // reserved
 263     AV_WB16(buf + 0x16a, ctx->m.mb_height * 4 + 4); // MSIPS
 264     buf[0x16d] = ctx->m.mb_height; // Ns
 265     buf[0x16f] = 0x10; // reserved
 266
 267     ctx->msip = buf + 0x170;
 268     return 0;
 269 }
 270
 271 static av_always_inline void dnxhd_encode_dc(DNXHDEncContext *ctx, int diff)
 272 {
 273     int nbits;
 274     if (diff < 0) {
 275         nbits = av_log2_16bit(-2*diff);
 276         diff--;
 277     } else {
 278         nbits = av_log2_16bit(2*diff);
 279     }
 280     put_bits(&ctx->m.pb, ctx->cid_table->dc_bits[nbits] + nbits,
 281              (ctx->cid_table->dc_codes[nbits]<<nbits) + (diff & ((1 << nbits) - 1)));
 282 }
 283
 284 static av_always_inline void dnxhd_encode_block(DNXHDEncContext *ctx, DCTELEM *block, int last_index, int n)
 285 {
 286     int last_non_zero = 0;
 287     int offset = 0;
 288     int slevel, i, j;
 289
 290     dnxhd_encode_dc(ctx, block[0] - ctx->m.last_dc[n]);
 291     ctx->m.last_dc[n] = block[0];
 292
 293     for (i = 1; i <= last_index; i++) {
 294         j = ctx->m.intra_scantable.permutated[i];
 295         slevel = block[j];
 296         if (slevel) {
 297             int run_level = i - last_non_zero - 1;
 298             int sign;
 299             MASK_ABS(sign, slevel);
 300             if (slevel > 64) {
 301                 offset = (slevel-1) >> 6;
 302                 slevel = 256 | (slevel & 63); // level 64 is treated as 0
 303             }
 304             if (run_level)
 305                 slevel |= 128;
 306             put_bits(&ctx->m.pb, ctx->table_vlc_bits[slevel]+1, (ctx->table_vlc_codes[slevel]<<1)|(sign&1));
 307             if (offset) {
 308                 put_bits(&ctx->m.pb, 4, offset);
 309                 offset = 0;
 310             }
 311             if (run_level)
 312                 put_bits(&ctx->m.pb, ctx->table_run_bits[run_level], ctx->table_run_codes[run_level]);
 313             last_non_zero = i;
 314         }
 315     }
 316     put_bits(&ctx->m.pb, ctx->table_vlc_bits[0], ctx->table_vlc_codes[0]); // EOB
 317 }
 318
 319 static av_always_inline void dnxhd_unquantize_c(DNXHDEncContext *ctx, DCTELEM *block, int n, int qscale, int last_index)
 320 {
 321     const uint8_t *weight_matrix;
 322     int level;
 323     int i;
 324
 325     weight_matrix = (n&2) ? ctx->cid_table->chroma_weight : ctx->cid_table->luma_weight;
 326
 327     for (i = 1; i <= last_index; i++) {
 328         int j = ctx->m.intra_scantable.permutated[i];
 329         level = block[j];
 330         if (level) {
 331             if (level < 0) {
 332                 level = (1-2*level) * qscale * weight_matrix[i];
 333                 if (weight_matrix[i] != 32)
 334                     level += 32;
 335                 level >>= 6;
 336                 level = -level;
 337             } else {
 338                 level = (2*level+1) * qscale * weight_matrix[i];
 339                 if (weight_matrix[i] != 32)
 340                     level += 32;
 341                 level >>= 6;
 342             }
 343             block[j] = level;
 344         }
 345     }
 346 }
 347
 348 static av_always_inline int dnxhd_ssd_block(DCTELEM *qblock, DCTELEM *block)
 349 {
 350     int score = 0;
 351     int i;
 352     for (i = 0; i < 64; i++)
 353         score += (block[i]-qblock[i])*(block[i]-qblock[i]);
 354     return score;
 355 }
 356
 357 static av_always_inline int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, DCTELEM *block, int last_index)
 358 {
 359     int last_non_zero = 0;
 360     int bits = 0;
 361     int i, j, level;
 362     for (i = 1; i <= last_index; i++) {
 363         j = ctx->m.intra_scantable.permutated[i];
 364         level = block[j];
 365         if (level) {
 366             int run_level = i - last_non_zero - 1;
 367             level = FFABS(level);
 368             if (level > 64) {
 369                 level = 256 | (level & 63); // level 64 is treated as 0
 370                 bits += 4;
 371             }
 372             level |= (!!run_level)<<7;
 373             bits += ctx->table_vlc_bits[level]+1 + ctx->table_run_bits[run_level];
 374             last_non_zero = i;
 375         }
 376     }
 377     return bits;
 378 }
 379
 380 static av_always_inline void dnxhd_get_pixels_4x8(DCTELEM *restrict block, const uint8_t *pixels, int line_size)
 381 {
 382     int i;
 383     for (i = 0; i < 4; i++) {
 384         block[0] = pixels[0];
 385         block[1] = pixels[1];
 386         block[2] = pixels[2];
 387         block[3] = pixels[3];
 388         block[4] = pixels[4];
 389         block[5] = pixels[5];
 390         block[6] = pixels[6];
 391         block[7] = pixels[7];
 392         pixels += line_size;
 393         block += 8;
 394     }
 395     memcpy(block   , block- 8, sizeof(*block)*8);
 396     memcpy(block+ 8, block-16, sizeof(*block)*8);
 397     memcpy(block+16, block-24, sizeof(*block)*8);
 398     memcpy(block+24, block-32, sizeof(*block)*8);
 399 }
 400
 401 static av_always_inline void dnxhd_get_blocks(DNXHDEncContext *ctx, int mb_x, int mb_y)
 402 {
 403     const uint8_t *ptr_y = ctx->thread[0]->src[0] + ((mb_y << 4) * ctx->m.linesize)   + (mb_x << 4);
 404     const uint8_t *ptr_u = ctx->thread[0]->src[1] + ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << 3);
 405     const uint8_t *ptr_v = ctx->thread[0]->src[2] + ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << 3);
 406     DSPContext *dsp = &ctx->m.dsp;
 407
 408     dsp->get_pixels(ctx->blocks[0], ptr_y    , ctx->m.linesize);
 409     dsp->get_pixels(ctx->blocks[1], ptr_y + 8, ctx->m.linesize);
 410     dsp->get_pixels(ctx->blocks[2], ptr_u    , ctx->m.uvlinesize);
 411     dsp->get_pixels(ctx->blocks[3], ptr_v    , ctx->m.uvlinesize);
 412
 413     if (mb_y+1 == ctx->m.mb_height && ctx->m.avctx->height == 1080) {
 414         if (ctx->interlaced) {
 415             dnxhd_get_pixels_4x8(ctx->blocks[4], ptr_y + ctx->dct_y_offset    , ctx->m.linesize);
 416             dnxhd_get_pixels_4x8(ctx->blocks[5], ptr_y + ctx->dct_y_offset + 8, ctx->m.linesize);
 417             dnxhd_get_pixels_4x8(ctx->blocks[6], ptr_u + ctx->dct_uv_offset   , ctx->m.uvlinesize);
 418             dnxhd_get_pixels_4x8(ctx->blocks[7], ptr_v + ctx->dct_uv_offset   , ctx->m.uvlinesize);
 419         } else
 420             memset(ctx->blocks[4], 0, 4*64*sizeof(DCTELEM));
 421     } else {
 422         dsp->get_pixels(ctx->blocks[4], ptr_y + ctx->dct_y_offset    , ctx->m.linesize);
 423         dsp->get_pixels(ctx->blocks[5], ptr_y + ctx->dct_y_offset + 8, ctx->m.linesize);
 424         dsp->get_pixels(ctx->blocks[6], ptr_u + ctx->dct_uv_offset   , ctx->m.uvlinesize);
 425         dsp->get_pixels(ctx->blocks[7], ptr_v + ctx->dct_uv_offset   , ctx->m.uvlinesize);
 426     }
 427 }
 428
 429 static av_always_inline int dnxhd_switch_matrix(DNXHDEncContext *ctx, int i)
 430 {
 431     if (i&2) {
 432         ctx->m.q_intra_matrix16 = ctx->qmatrix_c16;
 433         ctx->m.q_intra_matrix   = ctx->qmatrix_c;
 434         return 1 + (i&1);
 435     } else {
 436         ctx->m.q_intra_matrix16 = ctx->qmatrix_l16;
 437         ctx->m.q_intra_matrix   = ctx->qmatrix_l;
 438         return 0;
 439     }
 440 }
 441
 442 static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg)
 443 {
 444     DNXHDEncContext *ctx = arg;
 445     int mb_y, mb_x;
 446     int qscale = ctx->thread[0]->qscale;
 447
 448     for (mb_y = ctx->m.start_mb_y; mb_y < ctx->m.end_mb_y; mb_y++) {
 449         ctx->m.last_dc[0] =
 450         ctx->m.last_dc[1] =
 451         ctx->m.last_dc[2] = 1024;
 452
 453         for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
 454             unsigned mb = mb_y * ctx->m.mb_width + mb_x;
 455             int ssd     = 0;
 456             int ac_bits = 0;
 457             int dc_bits = 0;
 458             int i;
 459
 460             dnxhd_get_blocks(ctx, mb_x, mb_y);
 461
 462             for (i = 0; i < 8; i++) {
 463                 DECLARE_ALIGNED_16(DCTELEM, block[64]);
 464                 DCTELEM *src_block = ctx->blocks[i];
 465                 int overflow, nbits, diff, last_index;
 466                 int n = dnxhd_switch_matrix(ctx, i);
 467
 468                 memcpy(block, src_block, sizeof(block));
 469                 last_index = ctx->m.dct_quantize((MpegEncContext*)ctx, block, i, qscale, &overflow);
 470                 ac_bits += dnxhd_calc_ac_bits(ctx, block, last_index);
 471
 472                 diff = block[0] - ctx->m.last_dc[n];
 473                 if (diff < 0) nbits = av_log2_16bit(-2*diff);
 474                 else          nbits = av_log2_16bit( 2*diff);
 475                 dc_bits += ctx->cid_table->dc_bits[nbits] + nbits;
 476
 477                 ctx->m.last_dc[n] = block[0];
 478
 479                 if (avctx->mb_decision == FF_MB_DECISION_RD || !RC_VARIANCE) {
 480                     dnxhd_unquantize_c(ctx, block, i, qscale, last_index);
 481                     ctx->m.dsp.idct(block);
 482                     ssd += dnxhd_ssd_block(block, src_block);
 483                 }
 484             }
 485             ctx->mb_rc[qscale][mb].ssd = ssd;
 486             ctx->mb_rc[qscale][mb].bits = ac_bits+dc_bits+12+8*ctx->table_vlc_bits[0];
 487         }
 488     }
 489     return 0;
 490 }
 491
 492 static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg)
 493 {
 494     DNXHDEncContext *ctx = arg;
 495     int mb_y, mb_x;
 496
 497     for (mb_y = ctx->m.start_mb_y; mb_y < ctx->m.end_mb_y; mb_y++) {
 498         ctx->m.last_dc[0] =
 499         ctx->m.last_dc[1] =
 500         ctx->m.last_dc[2] = 1024;
 501         for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
 502             unsigned mb = mb_y * ctx->m.mb_width + mb_x;
 503             int qscale = ctx->mb_qscale[mb];
 504             int i;
 505
 506             put_bits(&ctx->m.pb, 12, qscale<<1);
 507
 508             dnxhd_get_blocks(ctx, mb_x, mb_y);
 509
 510             for (i = 0; i < 8; i++) {
 511                 DCTELEM *block = ctx->blocks[i];
 512                 int last_index, overflow;
 513                 int n = dnxhd_switch_matrix(ctx, i);
 514                 last_index = ctx->m.dct_quantize((MpegEncContext*)ctx, block, i, qscale, &overflow);
 515                 dnxhd_encode_block(ctx, block, last_index, n);
 516             }
 517         }
 518         if (put_bits_count(&ctx->m.pb)&31)
 519             put_bits(&ctx->m.pb, 32-(put_bits_count(&ctx->m.pb)&31), 0);
 520     }
 521     flush_put_bits(&ctx->m.pb);
 522     return 0;
 523 }
 524
 525 static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx, uint8_t *buf)
 526 {
 527     int mb_y, mb_x;
 528     int i, offset = 0;
 529     for (i = 0; i < ctx->m.avctx->thread_count; i++) {
 530         int thread_size = 0;
 531         for (mb_y = ctx->thread[i]->m.start_mb_y; mb_y < ctx->thread[i]->m.end_mb_y; mb_y++) {
 532             ctx->slice_size[mb_y] = 0;
 533             for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
 534                 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
 535                 ctx->slice_size[mb_y] += ctx->mb_bits[mb];
 536             }
 537             ctx->slice_size[mb_y] = (ctx->slice_size[mb_y]+31)&~31;
 538             ctx->slice_size[mb_y] >>= 3;
 539             thread_size += ctx->slice_size[mb_y];
 540         }
 541         init_put_bits(&ctx->thread[i]->m.pb, buf + 640 + offset, thread_size);
 542         offset += thread_size;
 543     }
 544 }
 545
 546 static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg)
 547 {
 548     DNXHDEncContext *ctx = arg;
 549     int mb_y, mb_x;
 550     for (mb_y = ctx->m.start_mb_y; mb_y < ctx->m.end_mb_y; mb_y++) {
 551         for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
 552             unsigned mb  = mb_y * ctx->m.mb_width + mb_x;
 553             uint8_t *pix = ctx->thread[0]->src[0] + ((mb_y<<4) * ctx->m.linesize) + (mb_x<<4);
 554             int sum      = ctx->m.dsp.pix_sum(pix, ctx->m.linesize);
 555             int varc     = (ctx->m.dsp.pix_norm1(pix, ctx->m.linesize) - (((unsigned)(sum*sum))>>8)+128)>>8;
 556             ctx->mb_cmp[mb].value = varc;
 557             ctx->mb_cmp[mb].mb = mb;
 558         }
 559     }
 560     return 0;
 561 }
 562
 563 static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx)
 564 {
 565     int lambda, up_step, down_step;
 566     int last_lower = INT_MAX, last_higher = 0;
 567     int x, y, q;
 568
 569     for (q = 1; q < avctx->qmax; q++) {
 570         ctx->qscale = q;
 571         avctx->execute(avctx, dnxhd_calc_bits_thread, (void**)&ctx->thread[0], NULL, avctx->thread_count);
 572     }
 573     up_step = down_step = 2<<LAMBDA_FRAC_BITS;
 574     lambda = ctx->lambda;
 575
 576     for (;;) {
 577         int bits = 0;
 578         int end = 0;
 579         if (lambda == last_higher) {
 580             lambda++;
 581             end = 1; // need to set final qscales/bits
 582         }
 583         for (y = 0; y < ctx->m.mb_height; y++) {
 584             for (x = 0; x < ctx->m.mb_width; x++) {
 585                 unsigned min = UINT_MAX;
 586                 int qscale = 1;
 587                 int mb = y*ctx->m.mb_width+x;
 588                 for (q = 1; q < avctx->qmax; q++) {
 589                     unsigned score = ctx->mb_rc[q][mb].bits*lambda+(ctx->mb_rc[q][mb].ssd<<LAMBDA_FRAC_BITS);
 590                     if (score < min) {
 591                         min = score;
 592                         qscale = q;
 593                     }
 594                 }
 595                 bits += ctx->mb_rc[qscale][mb].bits;
 596                 ctx->mb_qscale[mb] = qscale;
 597                 ctx->mb_bits[mb] = ctx->mb_rc[qscale][mb].bits;
 598             }
 599             bits = (bits+31)&~31; // padding
 600             if (bits > ctx->frame_bits)
 601                 break;
 602         }
 603         //dprintf(ctx->m.avctx, "lambda %d, up %u, down %u, bits %d, frame %d\n",
 604         //        lambda, last_higher, last_lower, bits, ctx->frame_bits);
 605         if (end) {
 606             if (bits > ctx->frame_bits)
 607                 return -1;
 608             break;
 609         }
 610         if (bits < ctx->frame_bits) {
 611             last_lower = FFMIN(lambda, last_lower);
 612             if (last_higher != 0)
 613                 lambda = (lambda+last_higher)>>1;
 614             else
 615                 lambda -= down_step;
 616             down_step *= 5; // XXX tune ?
 617             up_step = 1<<LAMBDA_FRAC_BITS;
 618             lambda = FFMAX(1, lambda);
 619             if (lambda == last_lower)
 620                 break;
 621         } else {
 622             last_higher = FFMAX(lambda, last_higher);
 623             if (last_lower != INT_MAX)
 624                 lambda = (lambda+last_lower)>>1;
 625             else
 626                 lambda += up_step;
 627             up_step *= 5;
 628             down_step = 1<<LAMBDA_FRAC_BITS;
 629         }
 630     }
 631     //dprintf(ctx->m.avctx, "out lambda %d\n", lambda);
 632     ctx->lambda = lambda;
 633     return 0;
 634 }
 635
 636 static int dnxhd_find_qscale(DNXHDEncContext *ctx)
 637 {
 638     int bits = 0;
 639     int up_step = 1;
 640     int down_step = 1;
 641     int last_higher = 0;
 642     int last_lower = INT_MAX;
 643     int qscale;
 644     int x, y;
 645
 646     qscale = ctx->qscale;
 647     for (;;) {
 648         bits = 0;
 649         ctx->qscale = qscale;
 650         // XXX avoid recalculating bits
 651         ctx->m.avctx->execute(ctx->m.avctx, dnxhd_calc_bits_thread, (void**)&ctx->thread[0], NULL, ctx->m.avctx->thread_count);
 652         for (y = 0; y < ctx->m.mb_height; y++) {
 653             for (x = 0; x < ctx->m.mb_width; x++)
 654                 bits += ctx->mb_rc[qscale][y*ctx->m.mb_width+x].bits;
 655             bits = (bits+31)&~31; // padding
 656             if (bits > ctx->frame_bits)
 657                 break;
 658         }
 659         //dprintf(ctx->m.avctx, "%d, qscale %d, bits %d, frame %d, higher %d, lower %d\n",
 660         //        ctx->m.avctx->frame_number, qscale, bits, ctx->frame_bits, last_higher, last_lower);
 661         if (bits < ctx->frame_bits) {
 662             if (qscale == 1)
 663                 return 1;
 664             if (last_higher == qscale - 1) {
 665                 qscale = last_higher;
 666                 break;
 667             }
 668             last_lower = FFMIN(qscale, last_lower);
 669             if (last_higher != 0)
 670                 qscale = (qscale+last_higher)>>1;
 671             else
 672                 qscale -= down_step++;
 673             if (qscale < 1)
 674                 qscale = 1;
 675             up_step = 1;
 676         } else {
 677             if (last_lower == qscale + 1)
 678                 break;
 679             last_higher = FFMAX(qscale, last_higher);
 680             if (last_lower != INT_MAX)
 681                 qscale = (qscale+last_lower)>>1;
 682             else
 683                 qscale += up_step++;
 684             down_step = 1;
 685             if (qscale >= ctx->m.avctx->qmax)
 686                 return -1;
 687         }
 688     }
 689     //dprintf(ctx->m.avctx, "out qscale %d\n", qscale);
 690     ctx->qscale = qscale;
 691     return 0;
 692 }
 693
 694 static int dnxhd_rc_cmp(const void *a, const void *b)
 695 {
 696     return ((const RCCMPEntry *)b)->value - ((const RCCMPEntry *)a)->value;
 697 }
 698
 699 static int dnxhd_encode_fast(AVCodecContext *avctx, DNXHDEncContext *ctx)
 700 {
 701     int max_bits = 0;
 702     int ret, x, y;
 703     if ((ret = dnxhd_find_qscale(ctx)) < 0)
 704         return -1;
 705     for (y = 0; y < ctx->m.mb_height; y++) {
 706         for (x = 0; x < ctx->m.mb_width; x++) {
 707             int mb = y*ctx->m.mb_width+x;
 708             int delta_bits;
 709             ctx->mb_qscale[mb] = ctx->qscale;
 710             ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale][mb].bits;
 711             max_bits += ctx->mb_rc[ctx->qscale][mb].bits;
 712             if (!RC_VARIANCE) {
 713                 delta_bits = ctx->mb_rc[ctx->qscale][mb].bits-ctx->mb_rc[ctx->qscale+1][mb].bits;
 714                 ctx->mb_cmp[mb].mb = mb;
 715                 ctx->mb_cmp[mb].value = delta_bits ?
 716                     ((ctx->mb_rc[ctx->qscale][mb].ssd-ctx->mb_rc[ctx->qscale+1][mb].ssd)*100)/delta_bits
 717                     : INT_MIN; //avoid increasing qscale
 718             }
 719         }
 720         max_bits += 31; //worst padding
 721     }
 722     if (!ret) {
 723         if (RC_VARIANCE)
 724             avctx->execute(avctx, dnxhd_mb_var_thread, (void**)&ctx->thread[0], NULL, avctx->thread_count);
 725         qsort(ctx->mb_cmp, ctx->m.mb_num, sizeof(RCEntry), dnxhd_rc_cmp);
 726         for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) {
 727             int mb = ctx->mb_cmp[x].mb;
 728             max_bits -= ctx->mb_rc[ctx->qscale][mb].bits - ctx->mb_rc[ctx->qscale+1][mb].bits;
 729             ctx->mb_qscale[mb] = ctx->qscale+1;
 730             ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale+1][mb].bits;
 731         }
 732     }
 733     return 0;
 734 }
 735
 736 static void dnxhd_load_picture(DNXHDEncContext *ctx, const AVFrame *frame)
 737 {
 738     int i;
 739
 740     for (i = 0; i < 3; i++) {
 741         ctx->frame.data[i]     = frame->data[i];
 742         ctx->frame.linesize[i] = frame->linesize[i];
 743     }
 744
 745     for (i = 0; i < ctx->m.avctx->thread_count; i++) {
 746         ctx->thread[i]->m.linesize    = ctx->frame.linesize[0]<<ctx->interlaced;
 747         ctx->thread[i]->m.uvlinesize  = ctx->frame.linesize[1]<<ctx->interlaced;
 748         ctx->thread[i]->dct_y_offset  = ctx->m.linesize  *8;
 749         ctx->thread[i]->dct_uv_offset = ctx->m.uvlinesize*8;
 750     }
 751
 752     ctx->frame.interlaced_frame = frame->interlaced_frame;
 753     ctx->cur_field = frame->interlaced_frame && !frame->top_field_first;
 754 }
 755
 756 static int dnxhd_encode_picture(AVCodecContext *avctx, unsigned char *buf, int buf_size, const void *data)
 757 {
 758     DNXHDEncContext *ctx = avctx->priv_data;
 759     int first_field = 1;
 760     int offset, i, ret;
 761
 762     if (buf_size < ctx->cid_table->frame_size) {
 763         av_log(avctx, AV_LOG_ERROR, "output buffer is too small to compress picture\n");
 764         return -1;
 765     }
 766
 767     dnxhd_load_picture(ctx, data);
 768
 769  encode_coding_unit:
 770     for (i = 0; i < 3; i++) {
 771         ctx->src[i] = ctx->frame.data[i];
 772         if (ctx->interlaced && ctx->cur_field)
 773             ctx->src[i] += ctx->frame.linesize[i];
 774     }
 775
 776     dnxhd_write_header(avctx, buf);
 777
 778     if (avctx->mb_decision == FF_MB_DECISION_RD)
 779         ret = dnxhd_encode_rdo(avctx, ctx);
 780     else
 781         ret = dnxhd_encode_fast(avctx, ctx);
 782     if (ret < 0) {
 783         av_log(avctx, AV_LOG_ERROR, "picture could not fit ratecontrol constraints\n");
 784         return -1;
 785     }
 786
 787     dnxhd_setup_threads_slices(ctx, buf);
 788
 789     offset = 0;
 790     for (i = 0; i < ctx->m.mb_height; i++) {
 791         AV_WB32(ctx->msip + i * 4, offset);
 792         offset += ctx->slice_size[i];
 793         assert(!(ctx->slice_size[i] & 3));
 794     }
 795
 796     avctx->execute(avctx, dnxhd_encode_thread, (void**)&ctx->thread[0], NULL, avctx->thread_count);
 797
 798     AV_WB32(buf + ctx->cid_table->coding_unit_size - 4, 0x600DC0DE); // EOF
 799
 800     if (ctx->interlaced && first_field) {
 801         first_field     = 0;
 802         ctx->cur_field ^= 1;
 803         buf      += ctx->cid_table->coding_unit_size;
 804         buf_size -= ctx->cid_table->coding_unit_size;
 805         goto encode_coding_unit;
 806     }
 807
 808     ctx->frame.quality = ctx->qscale*FF_QP2LAMBDA;
 809
 810     return ctx->cid_table->frame_size;
 811 }
 812
 813 static int dnxhd_encode_end(AVCodecContext *avctx)
 814 {
 815     DNXHDEncContext *ctx = avctx->priv_data;
 816     int i;
 817
 818     av_freep(&ctx->table_vlc_codes);
 819     av_freep(&ctx->table_vlc_bits);
 820     av_freep(&ctx->table_run_codes);
 821     av_freep(&ctx->table_run_bits);
 822
 823     av_freep(&ctx->mb_bits);
 824     av_freep(&ctx->mb_qscale);
 825     av_freep(&ctx->mb_rc);
 826     av_freep(&ctx->mb_cmp);
 827     av_freep(&ctx->slice_size);
 828
 829     av_freep(&ctx->qmatrix_c);
 830     av_freep(&ctx->qmatrix_l);
 831     av_freep(&ctx->qmatrix_c16);
 832     av_freep(&ctx->qmatrix_l16);
 833
 834     for (i = 1; i < avctx->thread_count; i++)
 835         av_freep(&ctx->thread[i]);
 836
 837     return 0;
 838 }
 839
 840 AVCodec dnxhd_encoder = {
 841     "dnxhd",
 842     CODEC_TYPE_VIDEO,
 843     CODEC_ID_DNXHD,
 844     sizeof(DNXHDEncContext),
 845     dnxhd_encode_init,
 846     dnxhd_encode_picture,
 847     dnxhd_encode_end,
 848     .pix_fmts = (enum PixelFormat[]){PIX_FMT_YUV422P, PIX_FMT_NONE},
 849     .long_name = "VC3/DNxHD",
 850 };