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 Indexes 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     uint32_t *vlc_codes;
  71     uint8_t  *vlc_bits;
  72     uint16_t *run_codes;
  73     uint8_t  *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, j, level, run;
  94     int max_level = 1<<(ctx->cid_table->bit_depth+2);
  95
  96     CHECKED_ALLOCZ(ctx->vlc_codes, max_level*4*sizeof(*ctx->vlc_codes));
  97     CHECKED_ALLOCZ(ctx->vlc_bits,  max_level*4*sizeof(*ctx->vlc_bits));
  98     CHECKED_ALLOCZ(ctx->run_codes, 63*2);
  99     CHECKED_ALLOCZ(ctx->run_bits,    63);
 100
 101     ctx->vlc_codes += max_level*2;
 102     ctx->vlc_bits  += max_level*2;
 103     for (level = -max_level; level < max_level; level++) {
 104         for (run = 0; run < 2; run++) {
 105             int index = (level<<1)|run;
 106             int sign, offset = 0, alevel = level;
 107
 108             MASK_ABS(sign, alevel);
 109             if (alevel > 64) {
 110                 offset = (alevel-1)>>6;
 111                 alevel -= offset<<6;
 112             }
 113             for (j = 0; j < 257; j++) {
 114                 if (ctx->cid_table->ac_level[j] == alevel &&
 115                     (!offset || (ctx->cid_table->ac_index_flag[j] && offset)) &&
 116                     (!run    || (ctx->cid_table->ac_run_flag  [j] && run))) {
 117                     assert(!ctx->vlc_codes[index]);
 118                     if (alevel) {
 119                         ctx->vlc_codes[index] = (ctx->cid_table->ac_codes[j]<<1)|(sign&1);
 120                         ctx->vlc_bits [index] = ctx->cid_table->ac_bits[j]+1;
 121                     } else {
 122                         ctx->vlc_codes[index] = ctx->cid_table->ac_codes[j];
 123                         ctx->vlc_bits [index] = ctx->cid_table->ac_bits [j];
 124                     }
 125                     break;
 126                 }
 127             }
 128             assert(!alevel || j < 257);
 129             if (offset) {
 130                 ctx->vlc_codes[index] = (ctx->vlc_codes[index]<<ctx->cid_table->index_bits)|offset;
 131                 ctx->vlc_bits [index]+= ctx->cid_table->index_bits;
 132             }
 133         }
 134     }
 135     for (i = 0; i < 62; i++) {
 136         int run = ctx->cid_table->run[i];
 137         assert(run < 63);
 138         ctx->run_codes[run] = ctx->cid_table->run_codes[i];
 139         ctx->run_bits [run] = ctx->cid_table->run_bits[i];
 140     }
 141     return 0;
 142  fail:
 143     return -1;
 144 }
 145
 146 static int dnxhd_init_qmat(DNXHDEncContext *ctx, int lbias, int cbias)
 147 {
 148     // init first elem to 1 to avoid div by 0 in convert_matrix
 149     uint16_t weight_matrix[64] = {1,}; // convert_matrix needs uint16_t*
 150     int qscale, i;
 151
 152     CHECKED_ALLOCZ(ctx->qmatrix_l,   (ctx->m.avctx->qmax+1) * 64 * sizeof(int));
 153     CHECKED_ALLOCZ(ctx->qmatrix_c,   (ctx->m.avctx->qmax+1) * 64 * sizeof(int));
 154     CHECKED_ALLOCZ(ctx->qmatrix_l16, (ctx->m.avctx->qmax+1) * 64 * 2 * sizeof(uint16_t));
 155     CHECKED_ALLOCZ(ctx->qmatrix_c16, (ctx->m.avctx->qmax+1) * 64 * 2 * sizeof(uint16_t));
 156
 157     for (i = 1; i < 64; i++) {
 158         int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]];
 159         weight_matrix[j] = ctx->cid_table->luma_weight[i];
 160     }
 161     ff_convert_matrix(&ctx->m.dsp, ctx->qmatrix_l, ctx->qmatrix_l16, weight_matrix,
 162                       ctx->m.intra_quant_bias, 1, ctx->m.avctx->qmax, 1);
 163     for (i = 1; i < 64; i++) {
 164         int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]];
 165         weight_matrix[j] = ctx->cid_table->chroma_weight[i];
 166     }
 167     ff_convert_matrix(&ctx->m.dsp, ctx->qmatrix_c, ctx->qmatrix_c16, weight_matrix,
 168                       ctx->m.intra_quant_bias, 1, ctx->m.avctx->qmax, 1);
 169     for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
 170         for (i = 0; i < 64; i++) {
 171             ctx->qmatrix_l  [qscale]   [i] <<= 2; ctx->qmatrix_c  [qscale]   [i] <<= 2;
 172             ctx->qmatrix_l16[qscale][0][i] <<= 2; ctx->qmatrix_l16[qscale][1][i] <<= 2;
 173             ctx->qmatrix_c16[qscale][0][i] <<= 2; ctx->qmatrix_c16[qscale][1][i] <<= 2;
 174         }
 175     }
 176     return 0;
 177  fail:
 178     return -1;
 179 }
 180
 181 static int dnxhd_init_rc(DNXHDEncContext *ctx)
 182 {
 183     CHECKED_ALLOCZ(ctx->mb_rc, 8160*ctx->m.avctx->qmax*sizeof(RCEntry));
 184     if (ctx->m.avctx->mb_decision != FF_MB_DECISION_RD)
 185         CHECKED_ALLOCZ(ctx->mb_cmp, ctx->m.mb_num*sizeof(RCCMPEntry));
 186
 187     ctx->frame_bits = (ctx->cid_table->coding_unit_size - 640 - 4) * 8;
 188     ctx->qscale = 1;
 189     ctx->lambda = 2<<LAMBDA_FRAC_BITS; // qscale 2
 190     return 0;
 191  fail:
 192     return -1;
 193 }
 194
 195 static int dnxhd_encode_init(AVCodecContext *avctx)
 196 {
 197     DNXHDEncContext *ctx = avctx->priv_data;
 198     int i, index;
 199
 200     ctx->cid = ff_dnxhd_find_cid(avctx);
 201     if (!ctx->cid || avctx->pix_fmt != PIX_FMT_YUV422P) {
 202         av_log(avctx, AV_LOG_ERROR, "video parameters incompatible with DNxHD\n");
 203         return -1;
 204     }
 205     av_log(avctx, AV_LOG_DEBUG, "cid %d\n", ctx->cid);
 206
 207     index = ff_dnxhd_get_cid_table(ctx->cid);
 208     ctx->cid_table = &ff_dnxhd_cid_table[index];
 209
 210     ctx->m.avctx = avctx;
 211     ctx->m.mb_intra = 1;
 212     ctx->m.h263_aic = 1;
 213
 214     dsputil_init(&ctx->m.dsp, avctx);
 215     ff_dct_common_init(&ctx->m);
 216     if (!ctx->m.dct_quantize)
 217         ctx->m.dct_quantize = dct_quantize_c;
 218
 219     ctx->m.mb_height = (avctx->height + 15) / 16;
 220     ctx->m.mb_width  = (avctx->width  + 15) / 16;
 221
 222     if (avctx->flags & CODEC_FLAG_INTERLACED_DCT) {
 223         ctx->interlaced = 1;
 224         ctx->m.mb_height /= 2;
 225     }
 226
 227     ctx->m.mb_num = ctx->m.mb_height * ctx->m.mb_width;
 228
 229     if (avctx->intra_quant_bias != FF_DEFAULT_QUANT_BIAS)
 230         ctx->m.intra_quant_bias = avctx->intra_quant_bias;
 231     if (dnxhd_init_qmat(ctx, ctx->m.intra_quant_bias, 0) < 0) // XXX tune lbias/cbias
 232         return -1;
 233
 234     if (dnxhd_init_vlc(ctx) < 0)
 235         return -1;
 236     if (dnxhd_init_rc(ctx) < 0)
 237         return -1;
 238
 239     CHECKED_ALLOCZ(ctx->slice_size, ctx->m.mb_height*sizeof(uint32_t));
 240     CHECKED_ALLOCZ(ctx->mb_bits,    ctx->m.mb_num   *sizeof(uint16_t));
 241     CHECKED_ALLOCZ(ctx->mb_qscale,  ctx->m.mb_num   *sizeof(uint8_t));
 242
 243     ctx->frame.key_frame = 1;
 244     ctx->frame.pict_type = FF_I_TYPE;
 245     ctx->m.avctx->coded_frame = &ctx->frame;
 246
 247     if (avctx->thread_count > MAX_THREADS || (avctx->thread_count > ctx->m.mb_height)) {
 248         av_log(avctx, AV_LOG_ERROR, "too many threads\n");
 249         return -1;
 250     }
 251
 252     ctx->thread[0] = ctx;
 253     for (i = 1; i < avctx->thread_count; i++) {
 254         ctx->thread[i] =  av_malloc(sizeof(DNXHDEncContext));
 255         memcpy(ctx->thread[i], ctx, sizeof(DNXHDEncContext));
 256     }
 257
 258     for (i = 0; i < avctx->thread_count; i++) {
 259         ctx->thread[i]->m.start_mb_y = (ctx->m.mb_height*(i  ) + avctx->thread_count/2) / avctx->thread_count;
 260         ctx->thread[i]->m.end_mb_y   = (ctx->m.mb_height*(i+1) + avctx->thread_count/2) / avctx->thread_count;
 261     }
 262
 263     return 0;
 264  fail: //for CHECKED_ALLOCZ
 265     return -1;
 266 }
 267
 268 static int dnxhd_write_header(AVCodecContext *avctx, uint8_t *buf)
 269 {
 270     DNXHDEncContext *ctx = avctx->priv_data;
 271     const uint8_t header_prefix[5] = { 0x00,0x00,0x02,0x80,0x01 };
 272
 273     memcpy(buf, header_prefix, 5);
 274     buf[5] = ctx->interlaced ? ctx->cur_field+2 : 0x01;
 275     buf[6] = 0x80; // crc flag off
 276     buf[7] = 0xa0; // reserved
 277     AV_WB16(buf + 0x18, avctx->height); // ALPF
 278     AV_WB16(buf + 0x1a, avctx->width);  // SPL
 279     AV_WB16(buf + 0x1d, avctx->height); // NAL
 280
 281     buf[0x21] = 0x38; // FIXME 8 bit per comp
 282     buf[0x22] = 0x88 + (ctx->frame.interlaced_frame<<2);
 283     AV_WB32(buf + 0x28, ctx->cid); // CID
 284     buf[0x2c] = ctx->interlaced ? 0 : 0x80;
 285
 286     buf[0x5f] = 0x01; // UDL
 287
 288     buf[0x167] = 0x02; // reserved
 289     AV_WB16(buf + 0x16a, ctx->m.mb_height * 4 + 4); // MSIPS
 290     buf[0x16d] = ctx->m.mb_height; // Ns
 291     buf[0x16f] = 0x10; // reserved
 292
 293     ctx->msip = buf + 0x170;
 294     return 0;
 295 }
 296
 297 static av_always_inline void dnxhd_encode_dc(DNXHDEncContext *ctx, int diff)
 298 {
 299     int nbits;
 300     if (diff < 0) {
 301         nbits = av_log2_16bit(-2*diff);
 302         diff--;
 303     } else {
 304         nbits = av_log2_16bit(2*diff);
 305     }
 306     put_bits(&ctx->m.pb, ctx->cid_table->dc_bits[nbits] + nbits,
 307              (ctx->cid_table->dc_codes[nbits]<<nbits) + (diff & ((1 << nbits) - 1)));
 308 }
 309
 310 static av_always_inline void dnxhd_encode_block(DNXHDEncContext *ctx, DCTELEM *block, int last_index, int n)
 311 {
 312     int last_non_zero = 0;
 313     int slevel, i, j;
 314
 315     dnxhd_encode_dc(ctx, block[0] - ctx->m.last_dc[n]);
 316     ctx->m.last_dc[n] = block[0];
 317
 318     for (i = 1; i <= last_index; i++) {
 319         j = ctx->m.intra_scantable.permutated[i];
 320         slevel = block[j];
 321         if (slevel) {
 322             int run_level = i - last_non_zero - 1;
 323             int rlevel = (slevel<<1)|!!run_level;
 324             put_bits(&ctx->m.pb, ctx->vlc_bits[rlevel], ctx->vlc_codes[rlevel]);
 325             if (run_level)
 326                 put_bits(&ctx->m.pb, ctx->run_bits[run_level], ctx->run_codes[run_level]);
 327             last_non_zero = i;
 328         }
 329     }
 330     put_bits(&ctx->m.pb, ctx->vlc_bits[0], ctx->vlc_codes[0]); // EOB
 331 }
 332
 333 static av_always_inline void dnxhd_unquantize_c(DNXHDEncContext *ctx, DCTELEM *block, int n, int qscale, int last_index)
 334 {
 335     const uint8_t *weight_matrix;
 336     int level;
 337     int i;
 338
 339     weight_matrix = (n&2) ? ctx->cid_table->chroma_weight : ctx->cid_table->luma_weight;
 340
 341     for (i = 1; i <= last_index; i++) {
 342         int j = ctx->m.intra_scantable.permutated[i];
 343         level = block[j];
 344         if (level) {
 345             if (level < 0) {
 346                 level = (1-2*level) * qscale * weight_matrix[i];
 347                 if (weight_matrix[i] != 32)
 348                     level += 32;
 349                 level >>= 6;
 350                 level = -level;
 351             } else {
 352                 level = (2*level+1) * qscale * weight_matrix[i];
 353                 if (weight_matrix[i] != 32)
 354                     level += 32;
 355                 level >>= 6;
 356             }
 357             block[j] = level;
 358         }
 359     }
 360 }
 361
 362 static av_always_inline int dnxhd_ssd_block(DCTELEM *qblock, DCTELEM *block)
 363 {
 364     int score = 0;
 365     int i;
 366     for (i = 0; i < 64; i++)
 367         score += (block[i]-qblock[i])*(block[i]-qblock[i]);
 368     return score;
 369 }
 370
 371 static av_always_inline int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, DCTELEM *block, int last_index)
 372 {
 373     int last_non_zero = 0;
 374     int bits = 0;
 375     int i, j, level;
 376     for (i = 1; i <= last_index; i++) {
 377         j = ctx->m.intra_scantable.permutated[i];
 378         level = block[j];
 379         if (level) {
 380             int run_level = i - last_non_zero - 1;
 381             bits += ctx->vlc_bits[(level<<1)|!!run_level]+ctx->run_bits[run_level];
 382             last_non_zero = i;
 383         }
 384     }
 385     return bits;
 386 }
 387
 388 static av_always_inline void dnxhd_get_pixels_4x8(DCTELEM *restrict block, const uint8_t *pixels, int line_size)
 389 {
 390     int i;
 391     for (i = 0; i < 4; i++) {
 392         block[0] = pixels[0];
 393         block[1] = pixels[1];
 394         block[2] = pixels[2];
 395         block[3] = pixels[3];
 396         block[4] = pixels[4];
 397         block[5] = pixels[5];
 398         block[6] = pixels[6];
 399         block[7] = pixels[7];
 400         pixels += line_size;
 401         block += 8;
 402     }
 403     memcpy(block   , block- 8, sizeof(*block)*8);
 404     memcpy(block+ 8, block-16, sizeof(*block)*8);
 405     memcpy(block+16, block-24, sizeof(*block)*8);
 406     memcpy(block+24, block-32, sizeof(*block)*8);
 407 }
 408
 409 static av_always_inline void dnxhd_get_blocks(DNXHDEncContext *ctx, int mb_x, int mb_y)
 410 {
 411     const uint8_t *ptr_y = ctx->thread[0]->src[0] + ((mb_y << 4) * ctx->m.linesize)   + (mb_x << 4);
 412     const uint8_t *ptr_u = ctx->thread[0]->src[1] + ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << 3);
 413     const uint8_t *ptr_v = ctx->thread[0]->src[2] + ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << 3);
 414     DSPContext *dsp = &ctx->m.dsp;
 415
 416     dsp->get_pixels(ctx->blocks[0], ptr_y    , ctx->m.linesize);
 417     dsp->get_pixels(ctx->blocks[1], ptr_y + 8, ctx->m.linesize);
 418     dsp->get_pixels(ctx->blocks[2], ptr_u    , ctx->m.uvlinesize);
 419     dsp->get_pixels(ctx->blocks[3], ptr_v    , ctx->m.uvlinesize);
 420
 421     if (mb_y+1 == ctx->m.mb_height && ctx->m.avctx->height == 1080) {
 422         if (ctx->interlaced) {
 423             dnxhd_get_pixels_4x8(ctx->blocks[4], ptr_y + ctx->dct_y_offset    , ctx->m.linesize);
 424             dnxhd_get_pixels_4x8(ctx->blocks[5], ptr_y + ctx->dct_y_offset + 8, ctx->m.linesize);
 425             dnxhd_get_pixels_4x8(ctx->blocks[6], ptr_u + ctx->dct_uv_offset   , ctx->m.uvlinesize);
 426             dnxhd_get_pixels_4x8(ctx->blocks[7], ptr_v + ctx->dct_uv_offset   , ctx->m.uvlinesize);
 427         } else
 428             memset(ctx->blocks[4], 0, 4*64*sizeof(DCTELEM));
 429     } else {
 430         dsp->get_pixels(ctx->blocks[4], ptr_y + ctx->dct_y_offset    , ctx->m.linesize);
 431         dsp->get_pixels(ctx->blocks[5], ptr_y + ctx->dct_y_offset + 8, ctx->m.linesize);
 432         dsp->get_pixels(ctx->blocks[6], ptr_u + ctx->dct_uv_offset   , ctx->m.uvlinesize);
 433         dsp->get_pixels(ctx->blocks[7], ptr_v + ctx->dct_uv_offset   , ctx->m.uvlinesize);
 434     }
 435 }
 436
 437 static av_always_inline int dnxhd_switch_matrix(DNXHDEncContext *ctx, int i)
 438 {
 439     if (i&2) {
 440         ctx->m.q_intra_matrix16 = ctx->qmatrix_c16;
 441         ctx->m.q_intra_matrix   = ctx->qmatrix_c;
 442         return 1 + (i&1);
 443     } else {
 444         ctx->m.q_intra_matrix16 = ctx->qmatrix_l16;
 445         ctx->m.q_intra_matrix   = ctx->qmatrix_l;
 446         return 0;
 447     }
 448 }
 449
 450 static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg)
 451 {
 452     DNXHDEncContext *ctx = arg;
 453     int mb_y, mb_x;
 454     int qscale = ctx->thread[0]->qscale;
 455
 456     for (mb_y = ctx->m.start_mb_y; mb_y < ctx->m.end_mb_y; mb_y++) {
 457         ctx->m.last_dc[0] =
 458         ctx->m.last_dc[1] =
 459         ctx->m.last_dc[2] = 1024;
 460
 461         for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
 462             unsigned mb = mb_y * ctx->m.mb_width + mb_x;
 463             int ssd     = 0;
 464             int ac_bits = 0;
 465             int dc_bits = 0;
 466             int i;
 467
 468             dnxhd_get_blocks(ctx, mb_x, mb_y);
 469
 470             for (i = 0; i < 8; i++) {
 471                 DECLARE_ALIGNED_16(DCTELEM, block[64]);
 472                 DCTELEM *src_block = ctx->blocks[i];
 473                 int overflow, nbits, diff, last_index;
 474                 int n = dnxhd_switch_matrix(ctx, i);
 475
 476                 memcpy(block, src_block, sizeof(block));
 477                 last_index = ctx->m.dct_quantize((MpegEncContext*)ctx, block, i, qscale, &overflow);
 478                 ac_bits += dnxhd_calc_ac_bits(ctx, block, last_index);
 479
 480                 diff = block[0] - ctx->m.last_dc[n];
 481                 if (diff < 0) nbits = av_log2_16bit(-2*diff);
 482                 else          nbits = av_log2_16bit( 2*diff);
 483                 dc_bits += ctx->cid_table->dc_bits[nbits] + nbits;
 484
 485                 ctx->m.last_dc[n] = block[0];
 486
 487                 if (avctx->mb_decision == FF_MB_DECISION_RD || !RC_VARIANCE) {
 488                     dnxhd_unquantize_c(ctx, block, i, qscale, last_index);
 489                     ctx->m.dsp.idct(block);
 490                     ssd += dnxhd_ssd_block(block, src_block);
 491                 }
 492             }
 493             ctx->mb_rc[qscale][mb].ssd = ssd;
 494             ctx->mb_rc[qscale][mb].bits = ac_bits+dc_bits+12+8*ctx->vlc_bits[0];
 495         }
 496     }
 497     return 0;
 498 }
 499
 500 static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg)
 501 {
 502     DNXHDEncContext *ctx = arg;
 503     int mb_y, mb_x;
 504
 505     for (mb_y = ctx->m.start_mb_y; mb_y < ctx->m.end_mb_y; mb_y++) {
 506         ctx->m.last_dc[0] =
 507         ctx->m.last_dc[1] =
 508         ctx->m.last_dc[2] = 1024;
 509         for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
 510             unsigned mb = mb_y * ctx->m.mb_width + mb_x;
 511             int qscale = ctx->mb_qscale[mb];
 512             int i;
 513
 514             put_bits(&ctx->m.pb, 12, qscale<<1);
 515
 516             dnxhd_get_blocks(ctx, mb_x, mb_y);
 517
 518             for (i = 0; i < 8; i++) {
 519                 DCTELEM *block = ctx->blocks[i];
 520                 int last_index, overflow;
 521                 int n = dnxhd_switch_matrix(ctx, i);
 522                 last_index = ctx->m.dct_quantize((MpegEncContext*)ctx, block, i, qscale, &overflow);
 523                 //START_TIMER;
 524                 dnxhd_encode_block(ctx, block, last_index, n);
 525                 //STOP_TIMER("encode_block");
 526             }
 527         }
 528         if (put_bits_count(&ctx->m.pb)&31)
 529             put_bits(&ctx->m.pb, 32-(put_bits_count(&ctx->m.pb)&31), 0);
 530     }
 531     flush_put_bits(&ctx->m.pb);
 532     return 0;
 533 }
 534
 535 static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx, uint8_t *buf)
 536 {
 537     int mb_y, mb_x;
 538     int i, offset = 0;
 539     for (i = 0; i < ctx->m.avctx->thread_count; i++) {
 540         int thread_size = 0;
 541         for (mb_y = ctx->thread[i]->m.start_mb_y; mb_y < ctx->thread[i]->m.end_mb_y; mb_y++) {
 542             ctx->slice_size[mb_y] = 0;
 543             for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
 544                 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
 545                 ctx->slice_size[mb_y] += ctx->mb_bits[mb];
 546             }
 547             ctx->slice_size[mb_y] = (ctx->slice_size[mb_y]+31)&~31;
 548             ctx->slice_size[mb_y] >>= 3;
 549             thread_size += ctx->slice_size[mb_y];
 550         }
 551         init_put_bits(&ctx->thread[i]->m.pb, buf + 640 + offset, thread_size);
 552         offset += thread_size;
 553     }
 554 }
 555
 556 static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg)
 557 {
 558     DNXHDEncContext *ctx = arg;
 559     int mb_y, mb_x;
 560     for (mb_y = ctx->m.start_mb_y; mb_y < ctx->m.end_mb_y; mb_y++) {
 561         for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
 562             unsigned mb  = mb_y * ctx->m.mb_width + mb_x;
 563             uint8_t *pix = ctx->thread[0]->src[0] + ((mb_y<<4) * ctx->m.linesize) + (mb_x<<4);
 564             int sum      = ctx->m.dsp.pix_sum(pix, ctx->m.linesize);
 565             int varc     = (ctx->m.dsp.pix_norm1(pix, ctx->m.linesize) - (((unsigned)(sum*sum))>>8)+128)>>8;
 566             ctx->mb_cmp[mb].value = varc;
 567             ctx->mb_cmp[mb].mb = mb;
 568         }
 569     }
 570     return 0;
 571 }
 572
 573 static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx)
 574 {
 575     int lambda, up_step, down_step;
 576     int last_lower = INT_MAX, last_higher = 0;
 577     int x, y, q;
 578
 579     for (q = 1; q < avctx->qmax; q++) {
 580         ctx->qscale = q;
 581         avctx->execute(avctx, dnxhd_calc_bits_thread, (void**)&ctx->thread[0], NULL, avctx->thread_count);
 582     }
 583     up_step = down_step = 2<<LAMBDA_FRAC_BITS;
 584     lambda = ctx->lambda;
 585
 586     for (;;) {
 587         int bits = 0;
 588         int end = 0;
 589         if (lambda == last_higher) {
 590             lambda++;
 591             end = 1; // need to set final qscales/bits
 592         }
 593         for (y = 0; y < ctx->m.mb_height; y++) {
 594             for (x = 0; x < ctx->m.mb_width; x++) {
 595                 unsigned min = UINT_MAX;
 596                 int qscale = 1;
 597                 int mb = y*ctx->m.mb_width+x;
 598                 for (q = 1; q < avctx->qmax; q++) {
 599                     unsigned score = ctx->mb_rc[q][mb].bits*lambda+(ctx->mb_rc[q][mb].ssd<<LAMBDA_FRAC_BITS);
 600                     if (score < min) {
 601                         min = score;
 602                         qscale = q;
 603                     }
 604                 }
 605                 bits += ctx->mb_rc[qscale][mb].bits;
 606                 ctx->mb_qscale[mb] = qscale;
 607                 ctx->mb_bits[mb] = ctx->mb_rc[qscale][mb].bits;
 608             }
 609             bits = (bits+31)&~31; // padding
 610             if (bits > ctx->frame_bits)
 611                 break;
 612         }
 613         //dprintf(ctx->m.avctx, "lambda %d, up %u, down %u, bits %d, frame %d\n",
 614         //        lambda, last_higher, last_lower, bits, ctx->frame_bits);
 615         if (end) {
 616             if (bits > ctx->frame_bits)
 617                 return -1;
 618             break;
 619         }
 620         if (bits < ctx->frame_bits) {
 621             last_lower = FFMIN(lambda, last_lower);
 622             if (last_higher != 0)
 623                 lambda = (lambda+last_higher)>>1;
 624             else
 625                 lambda -= down_step;
 626             down_step *= 5; // XXX tune ?
 627             up_step = 1<<LAMBDA_FRAC_BITS;
 628             lambda = FFMAX(1, lambda);
 629             if (lambda == last_lower)
 630                 break;
 631         } else {
 632             last_higher = FFMAX(lambda, last_higher);
 633             if (last_lower != INT_MAX)
 634                 lambda = (lambda+last_lower)>>1;
 635             else
 636                 lambda += up_step;
 637             up_step *= 5;
 638             down_step = 1<<LAMBDA_FRAC_BITS;
 639         }
 640     }
 641     //dprintf(ctx->m.avctx, "out lambda %d\n", lambda);
 642     ctx->lambda = lambda;
 643     return 0;
 644 }
 645
 646 static int dnxhd_find_qscale(DNXHDEncContext *ctx)
 647 {
 648     int bits = 0;
 649     int up_step = 1;
 650     int down_step = 1;
 651     int last_higher = 0;
 652     int last_lower = INT_MAX;
 653     int qscale;
 654     int x, y;
 655
 656     qscale = ctx->qscale;
 657     for (;;) {
 658         bits = 0;
 659         ctx->qscale = qscale;
 660         // XXX avoid recalculating bits
 661         ctx->m.avctx->execute(ctx->m.avctx, dnxhd_calc_bits_thread, (void**)&ctx->thread[0], NULL, ctx->m.avctx->thread_count);
 662         for (y = 0; y < ctx->m.mb_height; y++) {
 663             for (x = 0; x < ctx->m.mb_width; x++)
 664                 bits += ctx->mb_rc[qscale][y*ctx->m.mb_width+x].bits;
 665             bits = (bits+31)&~31; // padding
 666             if (bits > ctx->frame_bits)
 667                 break;
 668         }
 669         //dprintf(ctx->m.avctx, "%d, qscale %d, bits %d, frame %d, higher %d, lower %d\n",
 670         //        ctx->m.avctx->frame_number, qscale, bits, ctx->frame_bits, last_higher, last_lower);
 671         if (bits < ctx->frame_bits) {
 672             if (qscale == 1)
 673                 return 1;
 674             if (last_higher == qscale - 1) {
 675                 qscale = last_higher;
 676                 break;
 677             }
 678             last_lower = FFMIN(qscale, last_lower);
 679             if (last_higher != 0)
 680                 qscale = (qscale+last_higher)>>1;
 681             else
 682                 qscale -= down_step++;
 683             if (qscale < 1)
 684                 qscale = 1;
 685             up_step = 1;
 686         } else {
 687             if (last_lower == qscale + 1)
 688                 break;
 689             last_higher = FFMAX(qscale, last_higher);
 690             if (last_lower != INT_MAX)
 691                 qscale = (qscale+last_lower)>>1;
 692             else
 693                 qscale += up_step++;
 694             down_step = 1;
 695             if (qscale >= ctx->m.avctx->qmax)
 696                 return -1;
 697         }
 698     }
 699     //dprintf(ctx->m.avctx, "out qscale %d\n", qscale);
 700     ctx->qscale = qscale;
 701     return 0;
 702 }
 703
 704 static int dnxhd_rc_cmp(const void *a, const void *b)
 705 {
 706     return ((const RCCMPEntry *)b)->value - ((const RCCMPEntry *)a)->value;
 707 }
 708
 709 static int dnxhd_encode_fast(AVCodecContext *avctx, DNXHDEncContext *ctx)
 710 {
 711     int max_bits = 0;
 712     int ret, x, y;
 713     if ((ret = dnxhd_find_qscale(ctx)) < 0)
 714         return -1;
 715     for (y = 0; y < ctx->m.mb_height; y++) {
 716         for (x = 0; x < ctx->m.mb_width; x++) {
 717             int mb = y*ctx->m.mb_width+x;
 718             int delta_bits;
 719             ctx->mb_qscale[mb] = ctx->qscale;
 720             ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale][mb].bits;
 721             max_bits += ctx->mb_rc[ctx->qscale][mb].bits;
 722             if (!RC_VARIANCE) {
 723                 delta_bits = ctx->mb_rc[ctx->qscale][mb].bits-ctx->mb_rc[ctx->qscale+1][mb].bits;
 724                 ctx->mb_cmp[mb].mb = mb;
 725                 ctx->mb_cmp[mb].value = delta_bits ?
 726                     ((ctx->mb_rc[ctx->qscale][mb].ssd-ctx->mb_rc[ctx->qscale+1][mb].ssd)*100)/delta_bits
 727                     : INT_MIN; //avoid increasing qscale
 728             }
 729         }
 730         max_bits += 31; //worst padding
 731     }
 732     if (!ret) {
 733         if (RC_VARIANCE)
 734             avctx->execute(avctx, dnxhd_mb_var_thread, (void**)&ctx->thread[0], NULL, avctx->thread_count);
 735         qsort(ctx->mb_cmp, ctx->m.mb_num, sizeof(RCEntry), dnxhd_rc_cmp);
 736         for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) {
 737             int mb = ctx->mb_cmp[x].mb;
 738             max_bits -= ctx->mb_rc[ctx->qscale][mb].bits - ctx->mb_rc[ctx->qscale+1][mb].bits;
 739             ctx->mb_qscale[mb] = ctx->qscale+1;
 740             ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale+1][mb].bits;
 741         }
 742     }
 743     return 0;
 744 }
 745
 746 static void dnxhd_load_picture(DNXHDEncContext *ctx, const AVFrame *frame)
 747 {
 748     int i;
 749
 750     for (i = 0; i < 3; i++) {
 751         ctx->frame.data[i]     = frame->data[i];
 752         ctx->frame.linesize[i] = frame->linesize[i];
 753     }
 754
 755     for (i = 0; i < ctx->m.avctx->thread_count; i++) {
 756         ctx->thread[i]->m.linesize    = ctx->frame.linesize[0]<<ctx->interlaced;
 757         ctx->thread[i]->m.uvlinesize  = ctx->frame.linesize[1]<<ctx->interlaced;
 758         ctx->thread[i]->dct_y_offset  = ctx->m.linesize  *8;
 759         ctx->thread[i]->dct_uv_offset = ctx->m.uvlinesize*8;
 760     }
 761
 762     ctx->frame.interlaced_frame = frame->interlaced_frame;
 763     ctx->cur_field = frame->interlaced_frame && !frame->top_field_first;
 764 }
 765
 766 static int dnxhd_encode_picture(AVCodecContext *avctx, unsigned char *buf, int buf_size, const void *data)
 767 {
 768     DNXHDEncContext *ctx = avctx->priv_data;
 769     int first_field = 1;
 770     int offset, i, ret;
 771
 772     if (buf_size < ctx->cid_table->frame_size) {
 773         av_log(avctx, AV_LOG_ERROR, "output buffer is too small to compress picture\n");
 774         return -1;
 775     }
 776
 777     dnxhd_load_picture(ctx, data);
 778
 779  encode_coding_unit:
 780     for (i = 0; i < 3; i++) {
 781         ctx->src[i] = ctx->frame.data[i];
 782         if (ctx->interlaced && ctx->cur_field)
 783             ctx->src[i] += ctx->frame.linesize[i];
 784     }
 785
 786     dnxhd_write_header(avctx, buf);
 787
 788     if (avctx->mb_decision == FF_MB_DECISION_RD)
 789         ret = dnxhd_encode_rdo(avctx, ctx);
 790     else
 791         ret = dnxhd_encode_fast(avctx, ctx);
 792     if (ret < 0) {
 793         av_log(avctx, AV_LOG_ERROR, "picture could not fit ratecontrol constraints\n");
 794         return -1;
 795     }
 796
 797     dnxhd_setup_threads_slices(ctx, buf);
 798
 799     offset = 0;
 800     for (i = 0; i < ctx->m.mb_height; i++) {
 801         AV_WB32(ctx->msip + i * 4, offset);
 802         offset += ctx->slice_size[i];
 803         assert(!(ctx->slice_size[i] & 3));
 804     }
 805
 806     avctx->execute(avctx, dnxhd_encode_thread, (void**)&ctx->thread[0], NULL, avctx->thread_count);
 807
 808     AV_WB32(buf + ctx->cid_table->coding_unit_size - 4, 0x600DC0DE); // EOF
 809
 810     if (ctx->interlaced && first_field) {
 811         first_field     = 0;
 812         ctx->cur_field ^= 1;
 813         buf      += ctx->cid_table->coding_unit_size;
 814         buf_size -= ctx->cid_table->coding_unit_size;
 815         goto encode_coding_unit;
 816     }
 817
 818     ctx->frame.quality = ctx->qscale*FF_QP2LAMBDA;
 819
 820     return ctx->cid_table->frame_size;
 821 }
 822
 823 static int dnxhd_encode_end(AVCodecContext *avctx)
 824 {
 825     DNXHDEncContext *ctx = avctx->priv_data;
 826     int max_level = 1<<(ctx->cid_table->bit_depth+2);
 827     int i;
 828
 829     av_free(ctx->vlc_codes-max_level*2);
 830     av_free(ctx->vlc_bits -max_level*2);
 831     av_freep(&ctx->run_codes);
 832     av_freep(&ctx->run_bits);
 833
 834     av_freep(&ctx->mb_bits);
 835     av_freep(&ctx->mb_qscale);
 836     av_freep(&ctx->mb_rc);
 837     av_freep(&ctx->mb_cmp);
 838     av_freep(&ctx->slice_size);
 839
 840     av_freep(&ctx->qmatrix_c);
 841     av_freep(&ctx->qmatrix_l);
 842     av_freep(&ctx->qmatrix_c16);
 843     av_freep(&ctx->qmatrix_l16);
 844
 845     for (i = 1; i < avctx->thread_count; i++)
 846         av_freep(&ctx->thread[i]);
 847
 848     return 0;
 849 }
 850
 851 AVCodec dnxhd_encoder = {
 852     "dnxhd",
 853     CODEC_TYPE_VIDEO,
 854     CODEC_ID_DNXHD,
 855     sizeof(DNXHDEncContext),
 856     dnxhd_encode_init,
 857     dnxhd_encode_picture,
 858     dnxhd_encode_end,
 859     .pix_fmts = (enum PixelFormat[]){PIX_FMT_YUV422P, PIX_FMT_NONE},
 860     .long_name = NULL_IF_CONFIG_SMALL("VC3/DNxHD"),
 861 };