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 "dnxhdenc.h"
  31
  32 int dct_quantize_c(MpegEncContext *s, DCTELEM *block, int n, int qscale, int *overflow);
  33
  34 #define LAMBDA_FRAC_BITS 10
  35
  36 static av_always_inline void dnxhd_get_pixels_8x4(DCTELEM *restrict block, const uint8_t *pixels, int line_size)
  37 {
  38     int i;
  39     for (i = 0; i < 4; i++) {
  40         block[0] = pixels[0]; block[1] = pixels[1];
  41         block[2] = pixels[2]; block[3] = pixels[3];
  42         block[4] = pixels[4]; block[5] = pixels[5];
  43         block[6] = pixels[6]; block[7] = pixels[7];
  44         pixels += line_size;
  45         block += 8;
  46     }
  47     memcpy(block   , block- 8, sizeof(*block)*8);
  48     memcpy(block+ 8, block-16, sizeof(*block)*8);
  49     memcpy(block+16, block-24, sizeof(*block)*8);
  50     memcpy(block+24, block-32, sizeof(*block)*8);
  51 }
  52
  53 static int dnxhd_init_vlc(DNXHDEncContext *ctx)
  54 {
  55     int i, j, level, run;
  56     int max_level = 1<<(ctx->cid_table->bit_depth+2);
  57
  58     CHECKED_ALLOCZ(ctx->vlc_codes, max_level*4*sizeof(*ctx->vlc_codes));
  59     CHECKED_ALLOCZ(ctx->vlc_bits,  max_level*4*sizeof(*ctx->vlc_bits));
  60     CHECKED_ALLOCZ(ctx->run_codes, 63*2);
  61     CHECKED_ALLOCZ(ctx->run_bits,    63);
  62
  63     ctx->vlc_codes += max_level*2;
  64     ctx->vlc_bits  += max_level*2;
  65     for (level = -max_level; level < max_level; level++) {
  66         for (run = 0; run < 2; run++) {
  67             int index = (level<<1)|run;
  68             int sign, offset = 0, alevel = level;
  69
  70             MASK_ABS(sign, alevel);
  71             if (alevel > 64) {
  72                 offset = (alevel-1)>>6;
  73                 alevel -= offset<<6;
  74             }
  75             for (j = 0; j < 257; j++) {
  76                 if (ctx->cid_table->ac_level[j] == alevel &&
  77                     (!offset || (ctx->cid_table->ac_index_flag[j] && offset)) &&
  78                     (!run    || (ctx->cid_table->ac_run_flag  [j] && run))) {
  79                     assert(!ctx->vlc_codes[index]);
  80                     if (alevel) {
  81                         ctx->vlc_codes[index] = (ctx->cid_table->ac_codes[j]<<1)|(sign&1);
  82                         ctx->vlc_bits [index] = ctx->cid_table->ac_bits[j]+1;
  83                     } else {
  84                         ctx->vlc_codes[index] = ctx->cid_table->ac_codes[j];
  85                         ctx->vlc_bits [index] = ctx->cid_table->ac_bits [j];
  86                     }
  87                     break;
  88                 }
  89             }
  90             assert(!alevel || j < 257);
  91             if (offset) {
  92                 ctx->vlc_codes[index] = (ctx->vlc_codes[index]<<ctx->cid_table->index_bits)|offset;
  93                 ctx->vlc_bits [index]+= ctx->cid_table->index_bits;
  94             }
  95         }
  96     }
  97     for (i = 0; i < 62; i++) {
  98         int run = ctx->cid_table->run[i];
  99         assert(run < 63);
 100         ctx->run_codes[run] = ctx->cid_table->run_codes[i];
 101         ctx->run_bits [run] = ctx->cid_table->run_bits[i];
 102     }
 103     return 0;
 104  fail:
 105     return -1;
 106 }
 107
 108 static int dnxhd_init_qmat(DNXHDEncContext *ctx, int lbias, int cbias)
 109 {
 110     // init first elem to 1 to avoid div by 0 in convert_matrix
 111     uint16_t weight_matrix[64] = {1,}; // convert_matrix needs uint16_t*
 112     int qscale, i;
 113
 114     CHECKED_ALLOCZ(ctx->qmatrix_l,   (ctx->m.avctx->qmax+1) * 64 * sizeof(int));
 115     CHECKED_ALLOCZ(ctx->qmatrix_c,   (ctx->m.avctx->qmax+1) * 64 * sizeof(int));
 116     CHECKED_ALLOCZ(ctx->qmatrix_l16, (ctx->m.avctx->qmax+1) * 64 * 2 * sizeof(uint16_t));
 117     CHECKED_ALLOCZ(ctx->qmatrix_c16, (ctx->m.avctx->qmax+1) * 64 * 2 * sizeof(uint16_t));
 118
 119     for (i = 1; i < 64; i++) {
 120         int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]];
 121         weight_matrix[j] = ctx->cid_table->luma_weight[i];
 122     }
 123     ff_convert_matrix(&ctx->m.dsp, ctx->qmatrix_l, ctx->qmatrix_l16, weight_matrix,
 124                       ctx->m.intra_quant_bias, 1, ctx->m.avctx->qmax, 1);
 125     for (i = 1; i < 64; i++) {
 126         int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]];
 127         weight_matrix[j] = ctx->cid_table->chroma_weight[i];
 128     }
 129     ff_convert_matrix(&ctx->m.dsp, ctx->qmatrix_c, ctx->qmatrix_c16, weight_matrix,
 130                       ctx->m.intra_quant_bias, 1, ctx->m.avctx->qmax, 1);
 131     for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
 132         for (i = 0; i < 64; i++) {
 133             ctx->qmatrix_l  [qscale]   [i] <<= 2; ctx->qmatrix_c  [qscale]   [i] <<= 2;
 134             ctx->qmatrix_l16[qscale][0][i] <<= 2; ctx->qmatrix_l16[qscale][1][i] <<= 2;
 135             ctx->qmatrix_c16[qscale][0][i] <<= 2; ctx->qmatrix_c16[qscale][1][i] <<= 2;
 136         }
 137     }
 138     return 0;
 139  fail:
 140     return -1;
 141 }
 142
 143 static int dnxhd_init_rc(DNXHDEncContext *ctx)
 144 {
 145     CHECKED_ALLOCZ(ctx->mb_rc, 8160*ctx->m.avctx->qmax*sizeof(RCEntry));
 146     if (ctx->m.avctx->mb_decision != FF_MB_DECISION_RD)
 147         CHECKED_ALLOCZ(ctx->mb_cmp, ctx->m.mb_num*sizeof(RCCMPEntry));
 148
 149     ctx->frame_bits = (ctx->cid_table->coding_unit_size - 640 - 4) * 8;
 150     ctx->qscale = 1;
 151     ctx->lambda = 2<<LAMBDA_FRAC_BITS; // qscale 2
 152     return 0;
 153  fail:
 154     return -1;
 155 }
 156
 157 static int dnxhd_encode_init(AVCodecContext *avctx)
 158 {
 159     DNXHDEncContext *ctx = avctx->priv_data;
 160     int i, index;
 161
 162     ctx->cid = ff_dnxhd_find_cid(avctx);
 163     if (!ctx->cid || avctx->pix_fmt != PIX_FMT_YUV422P) {
 164         av_log(avctx, AV_LOG_ERROR, "video parameters incompatible with DNxHD\n");
 165         return -1;
 166     }
 167     av_log(avctx, AV_LOG_DEBUG, "cid %d\n", ctx->cid);
 168
 169     index = ff_dnxhd_get_cid_table(ctx->cid);
 170     ctx->cid_table = &ff_dnxhd_cid_table[index];
 171
 172     ctx->m.avctx = avctx;
 173     ctx->m.mb_intra = 1;
 174     ctx->m.h263_aic = 1;
 175
 176     ctx->get_pixels_8x4_sym = dnxhd_get_pixels_8x4;
 177
 178     dsputil_init(&ctx->m.dsp, avctx);
 179     ff_dct_common_init(&ctx->m);
 180 #if HAVE_MMX
 181     ff_dnxhd_init_mmx(ctx);
 182 #endif
 183     if (!ctx->m.dct_quantize)
 184         ctx->m.dct_quantize = dct_quantize_c;
 185
 186     ctx->m.mb_height = (avctx->height + 15) / 16;
 187     ctx->m.mb_width  = (avctx->width  + 15) / 16;
 188
 189     if (avctx->flags & CODEC_FLAG_INTERLACED_DCT) {
 190         ctx->interlaced = 1;
 191         ctx->m.mb_height /= 2;
 192     }
 193
 194     ctx->m.mb_num = ctx->m.mb_height * ctx->m.mb_width;
 195
 196     if (avctx->intra_quant_bias != FF_DEFAULT_QUANT_BIAS)
 197         ctx->m.intra_quant_bias = avctx->intra_quant_bias;
 198     if (dnxhd_init_qmat(ctx, ctx->m.intra_quant_bias, 0) < 0) // XXX tune lbias/cbias
 199         return -1;
 200
 201     if (dnxhd_init_vlc(ctx) < 0)
 202         return -1;
 203     if (dnxhd_init_rc(ctx) < 0)
 204         return -1;
 205
 206     CHECKED_ALLOCZ(ctx->slice_size, ctx->m.mb_height*sizeof(uint32_t));
 207     CHECKED_ALLOCZ(ctx->mb_bits,    ctx->m.mb_num   *sizeof(uint16_t));
 208     CHECKED_ALLOCZ(ctx->mb_qscale,  ctx->m.mb_num   *sizeof(uint8_t));
 209
 210     ctx->frame.key_frame = 1;
 211     ctx->frame.pict_type = FF_I_TYPE;
 212     ctx->m.avctx->coded_frame = &ctx->frame;
 213
 214     if (avctx->thread_count > MAX_THREADS || (avctx->thread_count > ctx->m.mb_height)) {
 215         av_log(avctx, AV_LOG_ERROR, "too many threads\n");
 216         return -1;
 217     }
 218
 219     ctx->thread[0] = ctx;
 220     for (i = 1; i < avctx->thread_count; i++) {
 221         ctx->thread[i] =  av_malloc(sizeof(DNXHDEncContext));
 222         memcpy(ctx->thread[i], ctx, sizeof(DNXHDEncContext));
 223     }
 224
 225     for (i = 0; i < avctx->thread_count; i++) {
 226         ctx->thread[i]->m.start_mb_y = (ctx->m.mb_height*(i  ) + avctx->thread_count/2) / avctx->thread_count;
 227         ctx->thread[i]->m.end_mb_y   = (ctx->m.mb_height*(i+1) + avctx->thread_count/2) / avctx->thread_count;
 228     }
 229
 230     return 0;
 231  fail: //for CHECKED_ALLOCZ
 232     return -1;
 233 }
 234
 235 static int dnxhd_write_header(AVCodecContext *avctx, uint8_t *buf)
 236 {
 237     DNXHDEncContext *ctx = avctx->priv_data;
 238     const uint8_t header_prefix[5] = { 0x00,0x00,0x02,0x80,0x01 };
 239
 240     memcpy(buf, header_prefix, 5);
 241     buf[5] = ctx->interlaced ? ctx->cur_field+2 : 0x01;
 242     buf[6] = 0x80; // crc flag off
 243     buf[7] = 0xa0; // reserved
 244     AV_WB16(buf + 0x18, avctx->height); // ALPF
 245     AV_WB16(buf + 0x1a, avctx->width);  // SPL
 246     AV_WB16(buf + 0x1d, avctx->height); // NAL
 247
 248     buf[0x21] = 0x38; // FIXME 8 bit per comp
 249     buf[0x22] = 0x88 + (ctx->frame.interlaced_frame<<2);
 250     AV_WB32(buf + 0x28, ctx->cid); // CID
 251     buf[0x2c] = ctx->interlaced ? 0 : 0x80;
 252
 253     buf[0x5f] = 0x01; // UDL
 254
 255     buf[0x167] = 0x02; // reserved
 256     AV_WB16(buf + 0x16a, ctx->m.mb_height * 4 + 4); // MSIPS
 257     buf[0x16d] = ctx->m.mb_height; // Ns
 258     buf[0x16f] = 0x10; // reserved
 259
 260     ctx->msip = buf + 0x170;
 261     return 0;
 262 }
 263
 264 static av_always_inline void dnxhd_encode_dc(DNXHDEncContext *ctx, int diff)
 265 {
 266     int nbits;
 267     if (diff < 0) {
 268         nbits = av_log2_16bit(-2*diff);
 269         diff--;
 270     } else {
 271         nbits = av_log2_16bit(2*diff);
 272     }
 273     put_bits(&ctx->m.pb, ctx->cid_table->dc_bits[nbits] + nbits,
 274              (ctx->cid_table->dc_codes[nbits]<<nbits) + (diff & ((1 << nbits) - 1)));
 275 }
 276
 277 static av_always_inline void dnxhd_encode_block(DNXHDEncContext *ctx, DCTELEM *block, int last_index, int n)
 278 {
 279     int last_non_zero = 0;
 280     int slevel, i, j;
 281
 282     dnxhd_encode_dc(ctx, block[0] - ctx->m.last_dc[n]);
 283     ctx->m.last_dc[n] = block[0];
 284
 285     for (i = 1; i <= last_index; i++) {
 286         j = ctx->m.intra_scantable.permutated[i];
 287         slevel = block[j];
 288         if (slevel) {
 289             int run_level = i - last_non_zero - 1;
 290             int rlevel = (slevel<<1)|!!run_level;
 291             put_bits(&ctx->m.pb, ctx->vlc_bits[rlevel], ctx->vlc_codes[rlevel]);
 292             if (run_level)
 293                 put_bits(&ctx->m.pb, ctx->run_bits[run_level], ctx->run_codes[run_level]);
 294             last_non_zero = i;
 295         }
 296     }
 297     put_bits(&ctx->m.pb, ctx->vlc_bits[0], ctx->vlc_codes[0]); // EOB
 298 }
 299
 300 static av_always_inline void dnxhd_unquantize_c(DNXHDEncContext *ctx, DCTELEM *block, int n, int qscale, int last_index)
 301 {
 302     const uint8_t *weight_matrix;
 303     int level;
 304     int i;
 305
 306     weight_matrix = (n&2) ? ctx->cid_table->chroma_weight : ctx->cid_table->luma_weight;
 307
 308     for (i = 1; i <= last_index; i++) {
 309         int j = ctx->m.intra_scantable.permutated[i];
 310         level = block[j];
 311         if (level) {
 312             if (level < 0) {
 313                 level = (1-2*level) * qscale * weight_matrix[i];
 314                 if (weight_matrix[i] != 32)
 315                     level += 32;
 316                 level >>= 6;
 317                 level = -level;
 318             } else {
 319                 level = (2*level+1) * qscale * weight_matrix[i];
 320                 if (weight_matrix[i] != 32)
 321                     level += 32;
 322                 level >>= 6;
 323             }
 324             block[j] = level;
 325         }
 326     }
 327 }
 328
 329 static av_always_inline int dnxhd_ssd_block(DCTELEM *qblock, DCTELEM *block)
 330 {
 331     int score = 0;
 332     int i;
 333     for (i = 0; i < 64; i++)
 334         score += (block[i]-qblock[i])*(block[i]-qblock[i]);
 335     return score;
 336 }
 337
 338 static av_always_inline int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, DCTELEM *block, int last_index)
 339 {
 340     int last_non_zero = 0;
 341     int bits = 0;
 342     int i, j, level;
 343     for (i = 1; i <= last_index; i++) {
 344         j = ctx->m.intra_scantable.permutated[i];
 345         level = block[j];
 346         if (level) {
 347             int run_level = i - last_non_zero - 1;
 348             bits += ctx->vlc_bits[(level<<1)|!!run_level]+ctx->run_bits[run_level];
 349             last_non_zero = i;
 350         }
 351     }
 352     return bits;
 353 }
 354
 355 static av_always_inline void dnxhd_get_blocks(DNXHDEncContext *ctx, int mb_x, int mb_y)
 356 {
 357     const uint8_t *ptr_y = ctx->thread[0]->src[0] + ((mb_y << 4) * ctx->m.linesize)   + (mb_x << 4);
 358     const uint8_t *ptr_u = ctx->thread[0]->src[1] + ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << 3);
 359     const uint8_t *ptr_v = ctx->thread[0]->src[2] + ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << 3);
 360     DSPContext *dsp = &ctx->m.dsp;
 361
 362     dsp->get_pixels(ctx->blocks[0], ptr_y    , ctx->m.linesize);
 363     dsp->get_pixels(ctx->blocks[1], ptr_y + 8, ctx->m.linesize);
 364     dsp->get_pixels(ctx->blocks[2], ptr_u    , ctx->m.uvlinesize);
 365     dsp->get_pixels(ctx->blocks[3], ptr_v    , ctx->m.uvlinesize);
 366
 367     if (mb_y+1 == ctx->m.mb_height && ctx->m.avctx->height == 1080) {
 368         if (ctx->interlaced) {
 369             ctx->get_pixels_8x4_sym(ctx->blocks[4], ptr_y + ctx->dct_y_offset    , ctx->m.linesize);
 370             ctx->get_pixels_8x4_sym(ctx->blocks[5], ptr_y + ctx->dct_y_offset + 8, ctx->m.linesize);
 371             ctx->get_pixels_8x4_sym(ctx->blocks[6], ptr_u + ctx->dct_uv_offset   , ctx->m.uvlinesize);
 372             ctx->get_pixels_8x4_sym(ctx->blocks[7], ptr_v + ctx->dct_uv_offset   , ctx->m.uvlinesize);
 373         } else {
 374             dsp->clear_block(ctx->blocks[4]); dsp->clear_block(ctx->blocks[5]);
 375             dsp->clear_block(ctx->blocks[6]); dsp->clear_block(ctx->blocks[7]);
 376         }
 377     } else {
 378         dsp->get_pixels(ctx->blocks[4], ptr_y + ctx->dct_y_offset    , ctx->m.linesize);
 379         dsp->get_pixels(ctx->blocks[5], ptr_y + ctx->dct_y_offset + 8, ctx->m.linesize);
 380         dsp->get_pixels(ctx->blocks[6], ptr_u + ctx->dct_uv_offset   , ctx->m.uvlinesize);
 381         dsp->get_pixels(ctx->blocks[7], ptr_v + ctx->dct_uv_offset   , ctx->m.uvlinesize);
 382     }
 383 }
 384
 385 static av_always_inline int dnxhd_switch_matrix(DNXHDEncContext *ctx, int i)
 386 {
 387     if (i&2) {
 388         ctx->m.q_intra_matrix16 = ctx->qmatrix_c16;
 389         ctx->m.q_intra_matrix   = ctx->qmatrix_c;
 390         return 1 + (i&1);
 391     } else {
 392         ctx->m.q_intra_matrix16 = ctx->qmatrix_l16;
 393         ctx->m.q_intra_matrix   = ctx->qmatrix_l;
 394         return 0;
 395     }
 396 }
 397
 398 static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg)
 399 {
 400     DNXHDEncContext *ctx = *(void**)arg;
 401     int mb_y, mb_x;
 402     int qscale = ctx->thread[0]->qscale;
 403
 404     for (mb_y = ctx->m.start_mb_y; mb_y < ctx->m.end_mb_y; mb_y++) {
 405         ctx->m.last_dc[0] =
 406         ctx->m.last_dc[1] =
 407         ctx->m.last_dc[2] = 1024;
 408
 409         for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
 410             unsigned mb = mb_y * ctx->m.mb_width + mb_x;
 411             int ssd     = 0;
 412             int ac_bits = 0;
 413             int dc_bits = 0;
 414             int i;
 415
 416             dnxhd_get_blocks(ctx, mb_x, mb_y);
 417
 418             for (i = 0; i < 8; i++) {
 419                 DECLARE_ALIGNED_16(DCTELEM, block[64]);
 420                 DCTELEM *src_block = ctx->blocks[i];
 421                 int overflow, nbits, diff, last_index;
 422                 int n = dnxhd_switch_matrix(ctx, i);
 423
 424                 memcpy(block, src_block, sizeof(block));
 425                 last_index = ctx->m.dct_quantize((MpegEncContext*)ctx, block, i, qscale, &overflow);
 426                 ac_bits += dnxhd_calc_ac_bits(ctx, block, last_index);
 427
 428                 diff = block[0] - ctx->m.last_dc[n];
 429                 if (diff < 0) nbits = av_log2_16bit(-2*diff);
 430                 else          nbits = av_log2_16bit( 2*diff);
 431                 dc_bits += ctx->cid_table->dc_bits[nbits] + nbits;
 432
 433                 ctx->m.last_dc[n] = block[0];
 434
 435                 if (avctx->mb_decision == FF_MB_DECISION_RD || !RC_VARIANCE) {
 436                     dnxhd_unquantize_c(ctx, block, i, qscale, last_index);
 437                     ctx->m.dsp.idct(block);
 438                     ssd += dnxhd_ssd_block(block, src_block);
 439                 }
 440             }
 441             ctx->mb_rc[qscale][mb].ssd = ssd;
 442             ctx->mb_rc[qscale][mb].bits = ac_bits+dc_bits+12+8*ctx->vlc_bits[0];
 443         }
 444     }
 445     return 0;
 446 }
 447
 448 static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg)
 449 {
 450     DNXHDEncContext *ctx = *(void**)arg;
 451     int mb_y, mb_x;
 452
 453     for (mb_y = ctx->m.start_mb_y; mb_y < ctx->m.end_mb_y; mb_y++) {
 454         ctx->m.last_dc[0] =
 455         ctx->m.last_dc[1] =
 456         ctx->m.last_dc[2] = 1024;
 457         for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
 458             unsigned mb = mb_y * ctx->m.mb_width + mb_x;
 459             int qscale = ctx->mb_qscale[mb];
 460             int i;
 461
 462             put_bits(&ctx->m.pb, 12, qscale<<1);
 463
 464             dnxhd_get_blocks(ctx, mb_x, mb_y);
 465
 466             for (i = 0; i < 8; i++) {
 467                 DCTELEM *block = ctx->blocks[i];
 468                 int last_index, overflow;
 469                 int n = dnxhd_switch_matrix(ctx, i);
 470                 last_index = ctx->m.dct_quantize((MpegEncContext*)ctx, block, i, qscale, &overflow);
 471                 //START_TIMER;
 472                 dnxhd_encode_block(ctx, block, last_index, n);
 473                 //STOP_TIMER("encode_block");
 474             }
 475         }
 476         if (put_bits_count(&ctx->m.pb)&31)
 477             put_bits(&ctx->m.pb, 32-(put_bits_count(&ctx->m.pb)&31), 0);
 478     }
 479     flush_put_bits(&ctx->m.pb);
 480     return 0;
 481 }
 482
 483 static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx, uint8_t *buf)
 484 {
 485     int mb_y, mb_x;
 486     int i, offset = 0;
 487     for (i = 0; i < ctx->m.avctx->thread_count; i++) {
 488         int thread_size = 0;
 489         for (mb_y = ctx->thread[i]->m.start_mb_y; mb_y < ctx->thread[i]->m.end_mb_y; mb_y++) {
 490             ctx->slice_size[mb_y] = 0;
 491             for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
 492                 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
 493                 ctx->slice_size[mb_y] += ctx->mb_bits[mb];
 494             }
 495             ctx->slice_size[mb_y] = (ctx->slice_size[mb_y]+31)&~31;
 496             ctx->slice_size[mb_y] >>= 3;
 497             thread_size += ctx->slice_size[mb_y];
 498         }
 499         init_put_bits(&ctx->thread[i]->m.pb, buf + 640 + offset, thread_size);
 500         offset += thread_size;
 501     }
 502 }
 503
 504 static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg)
 505 {
 506     DNXHDEncContext *ctx = *(void**)arg;
 507     int mb_y, mb_x;
 508     for (mb_y = ctx->m.start_mb_y; mb_y < ctx->m.end_mb_y; mb_y++) {
 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             uint8_t *pix = ctx->thread[0]->src[0] + ((mb_y<<4) * ctx->m.linesize) + (mb_x<<4);
 512             int sum      = ctx->m.dsp.pix_sum(pix, ctx->m.linesize);
 513             int varc     = (ctx->m.dsp.pix_norm1(pix, ctx->m.linesize) - (((unsigned)(sum*sum))>>8)+128)>>8;
 514             ctx->mb_cmp[mb].value = varc;
 515             ctx->mb_cmp[mb].mb = mb;
 516         }
 517     }
 518     return 0;
 519 }
 520
 521 static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx)
 522 {
 523     int lambda, up_step, down_step;
 524     int last_lower = INT_MAX, last_higher = 0;
 525     int x, y, q;
 526
 527     for (q = 1; q < avctx->qmax; q++) {
 528         ctx->qscale = q;
 529         avctx->execute(avctx, dnxhd_calc_bits_thread, (void**)&ctx->thread[0], NULL, avctx->thread_count, sizeof(void*));
 530     }
 531     up_step = down_step = 2<<LAMBDA_FRAC_BITS;
 532     lambda = ctx->lambda;
 533
 534     for (;;) {
 535         int bits = 0;
 536         int end = 0;
 537         if (lambda == last_higher) {
 538             lambda++;
 539             end = 1; // need to set final qscales/bits
 540         }
 541         for (y = 0; y < ctx->m.mb_height; y++) {
 542             for (x = 0; x < ctx->m.mb_width; x++) {
 543                 unsigned min = UINT_MAX;
 544                 int qscale = 1;
 545                 int mb = y*ctx->m.mb_width+x;
 546                 for (q = 1; q < avctx->qmax; q++) {
 547                     unsigned score = ctx->mb_rc[q][mb].bits*lambda+(ctx->mb_rc[q][mb].ssd<<LAMBDA_FRAC_BITS);
 548                     if (score < min) {
 549                         min = score;
 550                         qscale = q;
 551                     }
 552                 }
 553                 bits += ctx->mb_rc[qscale][mb].bits;
 554                 ctx->mb_qscale[mb] = qscale;
 555                 ctx->mb_bits[mb] = ctx->mb_rc[qscale][mb].bits;
 556             }
 557             bits = (bits+31)&~31; // padding
 558             if (bits > ctx->frame_bits)
 559                 break;
 560         }
 561         //dprintf(ctx->m.avctx, "lambda %d, up %u, down %u, bits %d, frame %d\n",
 562         //        lambda, last_higher, last_lower, bits, ctx->frame_bits);
 563         if (end) {
 564             if (bits > ctx->frame_bits)
 565                 return -1;
 566             break;
 567         }
 568         if (bits < ctx->frame_bits) {
 569             last_lower = FFMIN(lambda, last_lower);
 570             if (last_higher != 0)
 571                 lambda = (lambda+last_higher)>>1;
 572             else
 573                 lambda -= down_step;
 574             down_step *= 5; // XXX tune ?
 575             up_step = 1<<LAMBDA_FRAC_BITS;
 576             lambda = FFMAX(1, lambda);
 577             if (lambda == last_lower)
 578                 break;
 579         } else {
 580             last_higher = FFMAX(lambda, last_higher);
 581             if (last_lower != INT_MAX)
 582                 lambda = (lambda+last_lower)>>1;
 583             else
 584                 lambda += up_step;
 585             up_step *= 5;
 586             down_step = 1<<LAMBDA_FRAC_BITS;
 587         }
 588     }
 589     //dprintf(ctx->m.avctx, "out lambda %d\n", lambda);
 590     ctx->lambda = lambda;
 591     return 0;
 592 }
 593
 594 static int dnxhd_find_qscale(DNXHDEncContext *ctx)
 595 {
 596     int bits = 0;
 597     int up_step = 1;
 598     int down_step = 1;
 599     int last_higher = 0;
 600     int last_lower = INT_MAX;
 601     int qscale;
 602     int x, y;
 603
 604     qscale = ctx->qscale;
 605     for (;;) {
 606         bits = 0;
 607         ctx->qscale = qscale;
 608         // XXX avoid recalculating bits
 609         ctx->m.avctx->execute(ctx->m.avctx, dnxhd_calc_bits_thread, (void**)&ctx->thread[0], NULL, ctx->m.avctx->thread_count, sizeof(void*));
 610         for (y = 0; y < ctx->m.mb_height; y++) {
 611             for (x = 0; x < ctx->m.mb_width; x++)
 612                 bits += ctx->mb_rc[qscale][y*ctx->m.mb_width+x].bits;
 613             bits = (bits+31)&~31; // padding
 614             if (bits > ctx->frame_bits)
 615                 break;
 616         }
 617         //dprintf(ctx->m.avctx, "%d, qscale %d, bits %d, frame %d, higher %d, lower %d\n",
 618         //        ctx->m.avctx->frame_number, qscale, bits, ctx->frame_bits, last_higher, last_lower);
 619         if (bits < ctx->frame_bits) {
 620             if (qscale == 1)
 621                 return 1;
 622             if (last_higher == qscale - 1) {
 623                 qscale = last_higher;
 624                 break;
 625             }
 626             last_lower = FFMIN(qscale, last_lower);
 627             if (last_higher != 0)
 628                 qscale = (qscale+last_higher)>>1;
 629             else
 630                 qscale -= down_step++;
 631             if (qscale < 1)
 632                 qscale = 1;
 633             up_step = 1;
 634         } else {
 635             if (last_lower == qscale + 1)
 636                 break;
 637             last_higher = FFMAX(qscale, last_higher);
 638             if (last_lower != INT_MAX)
 639                 qscale = (qscale+last_lower)>>1;
 640             else
 641                 qscale += up_step++;
 642             down_step = 1;
 643             if (qscale >= ctx->m.avctx->qmax)
 644                 return -1;
 645         }
 646     }
 647     //dprintf(ctx->m.avctx, "out qscale %d\n", qscale);
 648     ctx->qscale = qscale;
 649     return 0;
 650 }
 651
 652 static int dnxhd_rc_cmp(const void *a, const void *b)
 653 {
 654     return ((const RCCMPEntry *)b)->value - ((const RCCMPEntry *)a)->value;
 655 }
 656
 657 static int dnxhd_encode_fast(AVCodecContext *avctx, DNXHDEncContext *ctx)
 658 {
 659     int max_bits = 0;
 660     int ret, x, y;
 661     if ((ret = dnxhd_find_qscale(ctx)) < 0)
 662         return -1;
 663     for (y = 0; y < ctx->m.mb_height; y++) {
 664         for (x = 0; x < ctx->m.mb_width; x++) {
 665             int mb = y*ctx->m.mb_width+x;
 666             int delta_bits;
 667             ctx->mb_qscale[mb] = ctx->qscale;
 668             ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale][mb].bits;
 669             max_bits += ctx->mb_rc[ctx->qscale][mb].bits;
 670             if (!RC_VARIANCE) {
 671                 delta_bits = ctx->mb_rc[ctx->qscale][mb].bits-ctx->mb_rc[ctx->qscale+1][mb].bits;
 672                 ctx->mb_cmp[mb].mb = mb;
 673                 ctx->mb_cmp[mb].value = delta_bits ?
 674                     ((ctx->mb_rc[ctx->qscale][mb].ssd-ctx->mb_rc[ctx->qscale+1][mb].ssd)*100)/delta_bits
 675                     : INT_MIN; //avoid increasing qscale
 676             }
 677         }
 678         max_bits += 31; //worst padding
 679     }
 680     if (!ret) {
 681         if (RC_VARIANCE)
 682             avctx->execute(avctx, dnxhd_mb_var_thread, (void**)&ctx->thread[0], NULL, avctx->thread_count, sizeof(void*));
 683         qsort(ctx->mb_cmp, ctx->m.mb_num, sizeof(RCEntry), dnxhd_rc_cmp);
 684         for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) {
 685             int mb = ctx->mb_cmp[x].mb;
 686             max_bits -= ctx->mb_rc[ctx->qscale][mb].bits - ctx->mb_rc[ctx->qscale+1][mb].bits;
 687             ctx->mb_qscale[mb] = ctx->qscale+1;
 688             ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale+1][mb].bits;
 689         }
 690     }
 691     return 0;
 692 }
 693
 694 static void dnxhd_load_picture(DNXHDEncContext *ctx, const AVFrame *frame)
 695 {
 696     int i;
 697
 698     for (i = 0; i < 3; i++) {
 699         ctx->frame.data[i]     = frame->data[i];
 700         ctx->frame.linesize[i] = frame->linesize[i];
 701     }
 702
 703     for (i = 0; i < ctx->m.avctx->thread_count; i++) {
 704         ctx->thread[i]->m.linesize    = ctx->frame.linesize[0]<<ctx->interlaced;
 705         ctx->thread[i]->m.uvlinesize  = ctx->frame.linesize[1]<<ctx->interlaced;
 706         ctx->thread[i]->dct_y_offset  = ctx->m.linesize  *8;
 707         ctx->thread[i]->dct_uv_offset = ctx->m.uvlinesize*8;
 708     }
 709
 710     ctx->frame.interlaced_frame = frame->interlaced_frame;
 711     ctx->cur_field = frame->interlaced_frame && !frame->top_field_first;
 712 }
 713
 714 static int dnxhd_encode_picture(AVCodecContext *avctx, unsigned char *buf, int buf_size, void *data)
 715 {
 716     DNXHDEncContext *ctx = avctx->priv_data;
 717     int first_field = 1;
 718     int offset, i, ret;
 719
 720     if (buf_size < ctx->cid_table->frame_size) {
 721         av_log(avctx, AV_LOG_ERROR, "output buffer is too small to compress picture\n");
 722         return -1;
 723     }
 724
 725     dnxhd_load_picture(ctx, data);
 726
 727  encode_coding_unit:
 728     for (i = 0; i < 3; i++) {
 729         ctx->src[i] = ctx->frame.data[i];
 730         if (ctx->interlaced && ctx->cur_field)
 731             ctx->src[i] += ctx->frame.linesize[i];
 732     }
 733
 734     dnxhd_write_header(avctx, buf);
 735
 736     if (avctx->mb_decision == FF_MB_DECISION_RD)
 737         ret = dnxhd_encode_rdo(avctx, ctx);
 738     else
 739         ret = dnxhd_encode_fast(avctx, ctx);
 740     if (ret < 0) {
 741         av_log(avctx, AV_LOG_ERROR, "picture could not fit ratecontrol constraints\n");
 742         return -1;
 743     }
 744
 745     dnxhd_setup_threads_slices(ctx, buf);
 746
 747     offset = 0;
 748     for (i = 0; i < ctx->m.mb_height; i++) {
 749         AV_WB32(ctx->msip + i * 4, offset);
 750         offset += ctx->slice_size[i];
 751         assert(!(ctx->slice_size[i] & 3));
 752     }
 753
 754     avctx->execute(avctx, dnxhd_encode_thread, (void**)&ctx->thread[0], NULL, avctx->thread_count, sizeof(void*));
 755
 756     AV_WB32(buf + ctx->cid_table->coding_unit_size - 4, 0x600DC0DE); // EOF
 757
 758     if (ctx->interlaced && first_field) {
 759         first_field     = 0;
 760         ctx->cur_field ^= 1;
 761         buf      += ctx->cid_table->coding_unit_size;
 762         buf_size -= ctx->cid_table->coding_unit_size;
 763         goto encode_coding_unit;
 764     }
 765
 766     ctx->frame.quality = ctx->qscale*FF_QP2LAMBDA;
 767
 768     return ctx->cid_table->frame_size;
 769 }
 770
 771 static int dnxhd_encode_end(AVCodecContext *avctx)
 772 {
 773     DNXHDEncContext *ctx = avctx->priv_data;
 774     int max_level = 1<<(ctx->cid_table->bit_depth+2);
 775     int i;
 776
 777     av_free(ctx->vlc_codes-max_level*2);
 778     av_free(ctx->vlc_bits -max_level*2);
 779     av_freep(&ctx->run_codes);
 780     av_freep(&ctx->run_bits);
 781
 782     av_freep(&ctx->mb_bits);
 783     av_freep(&ctx->mb_qscale);
 784     av_freep(&ctx->mb_rc);
 785     av_freep(&ctx->mb_cmp);
 786     av_freep(&ctx->slice_size);
 787
 788     av_freep(&ctx->qmatrix_c);
 789     av_freep(&ctx->qmatrix_l);
 790     av_freep(&ctx->qmatrix_c16);
 791     av_freep(&ctx->qmatrix_l16);
 792
 793     for (i = 1; i < avctx->thread_count; i++)
 794         av_freep(&ctx->thread[i]);
 795
 796     return 0;
 797 }
 798
 799 AVCodec dnxhd_encoder = {
 800     "dnxhd",
 801     CODEC_TYPE_VIDEO,
 802     CODEC_ID_DNXHD,
 803     sizeof(DNXHDEncContext),
 804     dnxhd_encode_init,
 805     dnxhd_encode_picture,
 806     dnxhd_encode_end,
 807     .pix_fmts = (enum PixelFormat[]){PIX_FMT_YUV422P, PIX_FMT_NONE},
 808     .long_name = NULL_IF_CONFIG_SMALL("VC3/DNxHD"),
 809 };