falkTX
/
FFmpeg
mirror of https://github.com/falkTX/FFmpeg.git
1
0
Fork 0
Code Issues 0 Releases 338 Wiki Activity
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
33469 Commits
32 Branches
766MB
Tree: 4b4a6ebd68
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '4b4a6ebd68'
${ noResults }
FFmpeg/libavcodec/dnxhdenc.c

1006 lines
35KB
Raw Blame History 

  1. /*

  2.  * VC3/DNxHD encoder

  3.  * Copyright (c) 2007 Baptiste Coudurier <baptiste dot coudurier at smartjog dot com>

  4.  * Copyright (c) 2011 MirriAd Ltd

  5.  *

  6.  * VC-3 encoder funded by the British Broadcasting Corporation

  7.  * 10 bit support added by MirriAd Ltd, Joseph Artsimovich <joseph@mirriad.com>

  8.  *

  9.  * This file is part of FFmpeg.

  10.  *

  11.  * FFmpeg is free software; you can redistribute it and/or

  12.  * modify it under the terms of the GNU Lesser General Public

  13.  * License as published by the Free Software Foundation; either

  14.  * version 2.1 of the License, or (at your option) any later version.

  15.  *

  16.  * FFmpeg is distributed in the hope that it will be useful,

  17.  * but WITHOUT ANY WARRANTY; without even the implied warranty of

  18.  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU

  19.  * Lesser General Public License for more details.

  20.  *

  21.  * You should have received a copy of the GNU Lesser General Public

  22.  * License along with FFmpeg; if not, write to the Free Software

  23.  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA

  24.  */

  25. 

  26. //#define DEBUG

  27. #define RC_VARIANCE 1 // use variance or ssd for fast rc

  28. 

  29. #include "libavutil/opt.h"

  30. #include "avcodec.h"

  31. #include "dsputil.h"

  32. #include "mpegvideo.h"

  33. #include "mpegvideo_common.h"

  34. #include "dnxhdenc.h"

  35. 

  36. #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM

  37. #define DNX10BIT_QMAT_SHIFT 18 // The largest value that will not lead to overflow for 10bit samples.

  38. 

  39. static const AVOption options[]={

  40.     {"nitris_compat", "encode with Avid Nitris compatibility", offsetof(DNXHDEncContext, nitris_compat), FF_OPT_TYPE_INT, {.dbl = 0}, 0, 1, VE},

  41. {NULL}

  42. };

  43. static const AVClass class = { "dnxhd", av_default_item_name, options, LIBAVUTIL_VERSION_INT };

  44. 

  45. #define LAMBDA_FRAC_BITS 10

  46. 

  47. static void dnxhd_8bit_get_pixels_8x4_sym(DCTELEM *restrict block, const uint8_t *pixels, int line_size)

  48. {

  49.     int i;

  50.     for (i = 0; i < 4; i++) {

  51.         block[0] = pixels[0]; block[1] = pixels[1];

  52.         block[2] = pixels[2]; block[3] = pixels[3];

  53.         block[4] = pixels[4]; block[5] = pixels[5];

  54.         block[6] = pixels[6]; block[7] = pixels[7];

  55.         pixels += line_size;

  56.         block += 8;

  57.     }

  58.     memcpy(block,      block -  8, sizeof(*block) * 8);

  59.     memcpy(block +  8, block - 16, sizeof(*block) * 8);

  60.     memcpy(block + 16, block - 24, sizeof(*block) * 8);

  61.     memcpy(block + 24, block - 32, sizeof(*block) * 8);

  62. }

  63. 

  64. static av_always_inline void dnxhd_10bit_get_pixels_8x4_sym(DCTELEM *restrict block, const uint8_t *pixels, int line_size)

  65. {

  66.     int i;

  67. 

  68.     block += 32;

  69. 

  70.     for (i = 0; i < 4; i++) {

  71.         memcpy(block + i     * 8, pixels + i * line_size, 8 * sizeof(*block));

  72.         memcpy(block - (i+1) * 8, pixels + i * line_size, 8 * sizeof(*block));

  73.     }

  74. }

  75. 

  76. static int dnxhd_10bit_dct_quantize(MpegEncContext *ctx, DCTELEM *block,

  77.                                     int n, int qscale, int *overflow)

  78. {

  79.     const uint8_t *scantable= ctx->intra_scantable.scantable;

  80.     const int *qmat = n<4 ? ctx->q_intra_matrix[qscale] : ctx->q_chroma_intra_matrix[qscale];

  81.     int last_non_zero = 0;

  82.     int i;

  83. 

  84.     ctx->dsp.fdct(block);

  85. 

  86.     // Divide by 4 with rounding, to compensate scaling of DCT coefficients

  87.     block[0] = (block[0] + 2) >> 2;

  88. 

  89.     for (i = 1; i < 64; ++i) {

  90.         int j = scantable[i];

  91.         int sign = block[j] >> 31;

  92.         int level = (block[j] ^ sign) - sign;

  93.         level = level * qmat[j] >> DNX10BIT_QMAT_SHIFT;

  94.         block[j] = (level ^ sign) - sign;

  95.         if (level)

  96.             last_non_zero = i;

  97.     }

  98. 

  99.     return last_non_zero;

  100. }

  101. 

  102. static int dnxhd_init_vlc(DNXHDEncContext *ctx)

  103. {

  104.     int i, j, level, run;

  105.     int max_level = 1<<(ctx->cid_table->bit_depth+2);

  106. 

  107.     FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->vlc_codes, max_level*4*sizeof(*ctx->vlc_codes), fail);

  108.     FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->vlc_bits,  max_level*4*sizeof(*ctx->vlc_bits) , fail);

  109.     FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_codes, 63*2,                                fail);

  110.     FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_bits,  63,                                  fail);

  111. 

  112.     ctx->vlc_codes += max_level*2;

  113.     ctx->vlc_bits  += max_level*2;

  114.     for (level = -max_level; level < max_level; level++) {

  115.         for (run = 0; run < 2; run++) {

  116.             int index = (level<<1)|run;

  117.             int sign, offset = 0, alevel = level;

  118. 

  119.             MASK_ABS(sign, alevel);

  120.             if (alevel > 64) {

  121.                 offset = (alevel-1)>>6;

  122.                 alevel -= offset<<6;

  123.             }

  124.             for (j = 0; j < 257; j++) {

  125.                 if (ctx->cid_table->ac_level[j] == alevel &&

  126.                     (!offset || (ctx->cid_table->ac_index_flag[j] && offset)) &&

  127.                     (!run    || (ctx->cid_table->ac_run_flag  [j] && run))) {

  128.                     assert(!ctx->vlc_codes[index]);

  129.                     if (alevel) {

  130.                         ctx->vlc_codes[index] = (ctx->cid_table->ac_codes[j]<<1)|(sign&1);

  131.                         ctx->vlc_bits [index] = ctx->cid_table->ac_bits[j]+1;

  132.                     } else {

  133.                         ctx->vlc_codes[index] = ctx->cid_table->ac_codes[j];

  134.                         ctx->vlc_bits [index] = ctx->cid_table->ac_bits [j];

  135.                     }

  136.                     break;

  137.                 }

  138.             }

  139.             assert(!alevel || j < 257);

  140.             if (offset) {

  141.                 ctx->vlc_codes[index] = (ctx->vlc_codes[index]<<ctx->cid_table->index_bits)|offset;

  142.                 ctx->vlc_bits [index]+= ctx->cid_table->index_bits;

  143.             }

  144.         }

  145.     }

  146.     for (i = 0; i < 62; i++) {

  147.         int run = ctx->cid_table->run[i];

  148.         assert(run < 63);

  149.         ctx->run_codes[run] = ctx->cid_table->run_codes[i];

  150.         ctx->run_bits [run] = ctx->cid_table->run_bits[i];

  151.     }

  152.     return 0;

  153.  fail:

  154.     return -1;

  155. }

  156. 

  157. static int dnxhd_init_qmat(DNXHDEncContext *ctx, int lbias, int cbias)

  158. {

  159.     // init first elem to 1 to avoid div by 0 in convert_matrix

  160.     uint16_t weight_matrix[64] = {1,}; // convert_matrix needs uint16_t*

  161.     int qscale, i;

  162.     const uint8_t *luma_weight_table   = ctx->cid_table->luma_weight;

  163.     const uint8_t *chroma_weight_table = ctx->cid_table->chroma_weight;

  164. 

  165.     FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l,   (ctx->m.avctx->qmax+1) * 64 *     sizeof(int),      fail);

  166.     FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c,   (ctx->m.avctx->qmax+1) * 64 *     sizeof(int),      fail);

  167.     FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l16, (ctx->m.avctx->qmax+1) * 64 * 2 * sizeof(uint16_t), fail);

  168.     FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c16, (ctx->m.avctx->qmax+1) * 64 * 2 * sizeof(uint16_t), fail);

  169. 

  170.     if (ctx->cid_table->bit_depth == 8) {

  171.         for (i = 1; i < 64; i++) {

  172.             int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]];

  173.             weight_matrix[j] = ctx->cid_table->luma_weight[i];

  174.         }

  175.         ff_convert_matrix(&ctx->m.dsp, ctx->qmatrix_l, ctx->qmatrix_l16, weight_matrix,

  176.                           ctx->m.intra_quant_bias, 1, ctx->m.avctx->qmax, 1);

  177.         for (i = 1; i < 64; i++) {

  178.             int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]];

  179.             weight_matrix[j] = ctx->cid_table->chroma_weight[i];

  180.         }

  181.         ff_convert_matrix(&ctx->m.dsp, ctx->qmatrix_c, ctx->qmatrix_c16, weight_matrix,

  182.                           ctx->m.intra_quant_bias, 1, ctx->m.avctx->qmax, 1);

  183. 

  184.         for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {

  185.             for (i = 0; i < 64; i++) {

  186.                 ctx->qmatrix_l  [qscale]   [i] <<= 2; ctx->qmatrix_c  [qscale]   [i] <<= 2;

  187.                 ctx->qmatrix_l16[qscale][0][i] <<= 2; ctx->qmatrix_l16[qscale][1][i] <<= 2;

  188.                 ctx->qmatrix_c16[qscale][0][i] <<= 2; ctx->qmatrix_c16[qscale][1][i] <<= 2;

  189.             }

  190.         }

  191.     } else {

  192.         // 10-bit

  193.         for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {

  194.             for (i = 1; i < 64; i++) {

  195.                 int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]];

  196. 

  197.                 // The quantization formula from the VC-3 standard is:

  198.                 // quantized = sign(block[i]) * floor(abs(block[i]/s) * p / (qscale * weight_table[i]))

  199.                 // Where p is 32 for 8-bit samples and 8 for 10-bit ones.

  200.                 // The s factor compensates scaling of DCT coefficients done by the DCT routines,

  201.                 // and therefore is not present in standard.  It's 8 for 8-bit samples and 4 for 10-bit ones.

  202.                 // We want values of ctx->qtmatrix_l and ctx->qtmatrix_r to be:

  203.                 // ((1 << DNX10BIT_QMAT_SHIFT) * (p / s)) / (qscale * weight_table[i])

  204.                 // For 10-bit samples, p / s == 2

  205.                 ctx->qmatrix_l[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) / (qscale * luma_weight_table[i]);

  206.                 ctx->qmatrix_c[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) / (qscale * chroma_weight_table[i]);

  207.             }

  208.         }

  209.     }

  210. 

  211.     ctx->m.q_chroma_intra_matrix16 = ctx->qmatrix_c16;

  212.     ctx->m.q_chroma_intra_matrix   = ctx->qmatrix_c;

  213.     ctx->m.q_intra_matrix16        = ctx->qmatrix_l16;

  214.     ctx->m.q_intra_matrix          = ctx->qmatrix_l;

  215. 

  216.     return 0;

  217.  fail:

  218.     return -1;

  219. }

  220. 

  221. static int dnxhd_init_rc(DNXHDEncContext *ctx)

  222. {

  223.     FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_rc, 8160*ctx->m.avctx->qmax*sizeof(RCEntry), fail);

  224.     if (ctx->m.avctx->mb_decision != FF_MB_DECISION_RD)

  225.         FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_cmp, ctx->m.mb_num*sizeof(RCCMPEntry), fail);

  226. 

  227.     ctx->frame_bits = (ctx->cid_table->coding_unit_size - 640 - 4 - ctx->min_padding) * 8;

  228.     ctx->qscale = 1;

  229.     ctx->lambda = 2<<LAMBDA_FRAC_BITS; // qscale 2

  230.     return 0;

  231.  fail:

  232.     return -1;

  233. }

  234. 

  235. static int dnxhd_encode_init(AVCodecContext *avctx)

  236. {

  237.     DNXHDEncContext *ctx = avctx->priv_data;

  238.     int i, index, bit_depth;

  239. 

  240.     switch (avctx->pix_fmt) {

  241.     case PIX_FMT_YUV422P:

  242.         bit_depth = 8;

  243.         break;

  244.     case PIX_FMT_YUV422P10:

  245.         bit_depth = 10;

  246.         break;

  247.     default:

  248.         av_log(avctx, AV_LOG_ERROR, "pixel format is incompatible with DNxHD\n");

  249.         return -1;

  250.     }

  251. 

  252.     ctx->cid = ff_dnxhd_find_cid(avctx, bit_depth);

  253.     if (!ctx->cid) {

  254.         av_log(avctx, AV_LOG_ERROR, "video parameters incompatible with DNxHD\n");

  255.         return -1;

  256.     }

  257.     av_log(avctx, AV_LOG_DEBUG, "cid %d\n", ctx->cid);

  258. 

  259.     index = ff_dnxhd_get_cid_table(ctx->cid);

  260.     ctx->cid_table = &ff_dnxhd_cid_table[index];

  261. 

  262.     ctx->m.avctx = avctx;

  263.     ctx->m.mb_intra = 1;

  264.     ctx->m.h263_aic = 1;

  265. 

  266.     avctx->bits_per_raw_sample = ctx->cid_table->bit_depth;

  267. 

  268.     dsputil_init(&ctx->m.dsp, avctx);

  269.     ff_dct_common_init(&ctx->m);

  270.     if (!ctx->m.dct_quantize)

  271.         ctx->m.dct_quantize = dct_quantize_c;

  272. 

  273.     if (ctx->cid_table->bit_depth == 10) {

  274.        ctx->m.dct_quantize = dnxhd_10bit_dct_quantize;

  275.        ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym;

  276.        ctx->block_width_l2 = 4;

  277.     } else {

  278.        ctx->get_pixels_8x4_sym = dnxhd_8bit_get_pixels_8x4_sym;

  279.        ctx->block_width_l2 = 3;

  280.     }

  281. 

  282. #if HAVE_MMX

  283.     ff_dnxhd_init_mmx(ctx);

  284. #endif

  285. 

  286.     ctx->m.mb_height = (avctx->height + 15) / 16;

  287.     ctx->m.mb_width  = (avctx->width  + 15) / 16;

  288. 

  289.     if (avctx->flags & CODEC_FLAG_INTERLACED_DCT) {

  290.         ctx->interlaced = 1;

  291.         ctx->m.mb_height /= 2;

  292.     }

  293. 

  294.     ctx->m.mb_num = ctx->m.mb_height * ctx->m.mb_width;

  295. 

  296.     if (avctx->intra_quant_bias != FF_DEFAULT_QUANT_BIAS)

  297.         ctx->m.intra_quant_bias = avctx->intra_quant_bias;

  298.     if (dnxhd_init_qmat(ctx, ctx->m.intra_quant_bias, 0) < 0) // XXX tune lbias/cbias

  299.         return -1;

  300. 

  301.     // Avid Nitris hardware decoder requires a minimum amount of padding in the coding unit payload

  302.     if (ctx->nitris_compat)

  303.         ctx->min_padding = 1600;

  304. 

  305.     if (dnxhd_init_vlc(ctx) < 0)

  306.         return -1;

  307.     if (dnxhd_init_rc(ctx) < 0)

  308.         return -1;

  309. 

  310.     FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_size, ctx->m.mb_height*sizeof(uint32_t), fail);

  311.     FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_offs, ctx->m.mb_height*sizeof(uint32_t), fail);

  312.     FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_bits,    ctx->m.mb_num   *sizeof(uint16_t), fail);

  313.     FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_qscale,  ctx->m.mb_num   *sizeof(uint8_t),  fail);

  314. 

  315.     ctx->frame.key_frame = 1;

  316.     ctx->frame.pict_type = AV_PICTURE_TYPE_I;

  317.     ctx->m.avctx->coded_frame = &ctx->frame;

  318. 

  319.     if (avctx->thread_count > MAX_THREADS) {

  320.         av_log(avctx, AV_LOG_ERROR, "too many threads\n");

  321.         return -1;

  322.     }

  323. 

  324.     ctx->thread[0] = ctx;

  325.     for (i = 1; i < avctx->thread_count; i++) {

  326.         ctx->thread[i] =  av_malloc(sizeof(DNXHDEncContext));

  327.         memcpy(ctx->thread[i], ctx, sizeof(DNXHDEncContext));

  328.     }

  329. 

  330.     return 0;

  331.  fail: //for FF_ALLOCZ_OR_GOTO

  332.     return -1;

  333. }

  334. 

  335. static int dnxhd_write_header(AVCodecContext *avctx, uint8_t *buf)

  336. {

  337.     DNXHDEncContext *ctx = avctx->priv_data;

  338.     const uint8_t header_prefix[5] = { 0x00,0x00,0x02,0x80,0x01 };

  339. 

  340.     memset(buf, 0, 640);

  341. 

  342.     memcpy(buf, header_prefix, 5);

  343.     buf[5] = ctx->interlaced ? ctx->cur_field+2 : 0x01;

  344.     buf[6] = 0x80; // crc flag off

  345.     buf[7] = 0xa0; // reserved

  346.     AV_WB16(buf + 0x18, avctx->height>>ctx->interlaced); // ALPF

  347.     AV_WB16(buf + 0x1a, avctx->width);  // SPL

  348.     AV_WB16(buf + 0x1d, avctx->height>>ctx->interlaced); // NAL

  349. 

  350.     buf[0x21] = ctx->cid_table->bit_depth == 10 ? 0x58 : 0x38;

  351.     buf[0x22] = 0x88 + (ctx->interlaced<<2);

  352.     AV_WB32(buf + 0x28, ctx->cid); // CID

  353.     buf[0x2c] = ctx->interlaced ? 0 : 0x80;

  354. 

  355.     buf[0x5f] = 0x01; // UDL

  356. 

  357.     buf[0x167] = 0x02; // reserved

  358.     AV_WB16(buf + 0x16a, ctx->m.mb_height * 4 + 4); // MSIPS

  359.     buf[0x16d] = ctx->m.mb_height; // Ns

  360.     buf[0x16f] = 0x10; // reserved

  361. 

  362.     ctx->msip = buf + 0x170;

  363.     return 0;

  364. }

  365. 

  366. static av_always_inline void dnxhd_encode_dc(DNXHDEncContext *ctx, int diff)

  367. {

  368.     int nbits;

  369.     if (diff < 0) {

  370.         nbits = av_log2_16bit(-2*diff);

  371.         diff--;

  372.     } else {

  373.         nbits = av_log2_16bit(2*diff);

  374.     }

  375.     put_bits(&ctx->m.pb, ctx->cid_table->dc_bits[nbits] + nbits,

  376.              (ctx->cid_table->dc_codes[nbits]<<nbits) + (diff & ((1 << nbits) - 1)));

  377. }

  378. 

  379. static av_always_inline void dnxhd_encode_block(DNXHDEncContext *ctx, DCTELEM *block, int last_index, int n)

  380. {

  381.     int last_non_zero = 0;

  382.     int slevel, i, j;

  383. 

  384.     dnxhd_encode_dc(ctx, block[0] - ctx->m.last_dc[n]);

  385.     ctx->m.last_dc[n] = block[0];

  386. 

  387.     for (i = 1; i <= last_index; i++) {

  388.         j = ctx->m.intra_scantable.permutated[i];

  389.         slevel = block[j];

  390.         if (slevel) {

  391.             int run_level = i - last_non_zero - 1;

  392.             int rlevel = (slevel<<1)|!!run_level;

  393.             put_bits(&ctx->m.pb, ctx->vlc_bits[rlevel], ctx->vlc_codes[rlevel]);

  394.             if (run_level)

  395.                 put_bits(&ctx->m.pb, ctx->run_bits[run_level], ctx->run_codes[run_level]);

  396.             last_non_zero = i;

  397.         }

  398.     }

  399.     put_bits(&ctx->m.pb, ctx->vlc_bits[0], ctx->vlc_codes[0]); // EOB

  400. }

  401. 

  402. static av_always_inline void dnxhd_unquantize_c(DNXHDEncContext *ctx, DCTELEM *block, int n, int qscale, int last_index)

  403. {

  404.     const uint8_t *weight_matrix;

  405.     int level;

  406.     int i;

  407. 

  408.     weight_matrix = (n&2) ? ctx->cid_table->chroma_weight : ctx->cid_table->luma_weight;

  409. 

  410.     for (i = 1; i <= last_index; i++) {

  411.         int j = ctx->m.intra_scantable.permutated[i];

  412.         level = block[j];

  413.         if (level) {

  414.             if (level < 0) {

  415.                 level = (1-2*level) * qscale * weight_matrix[i];

  416.                 if (ctx->cid_table->bit_depth == 10) {

  417.                     if (weight_matrix[i] != 8)

  418.                         level += 8;

  419.                     level >>= 4;

  420.                 } else {

  421.                     if (weight_matrix[i] != 32)

  422.                         level += 32;

  423.                     level >>= 6;

  424.                 }

  425.                 level = -level;

  426.             } else {

  427.                 level = (2*level+1) * qscale * weight_matrix[i];

  428.                 if (ctx->cid_table->bit_depth == 10) {

  429.                     if (weight_matrix[i] != 8)

  430.                         level += 8;

  431.                     level >>= 4;

  432.                 } else {

  433.                     if (weight_matrix[i] != 32)

  434.                         level += 32;

  435.                     level >>= 6;

  436.                 }

  437.             }

  438.             block[j] = level;

  439.         }

  440.     }

  441. }

  442. 

  443. static av_always_inline int dnxhd_ssd_block(DCTELEM *qblock, DCTELEM *block)

  444. {

  445.     int score = 0;

  446.     int i;

  447.     for (i = 0; i < 64; i++)

  448.         score += (block[i] - qblock[i]) * (block[i] - qblock[i]);

  449.     return score;

  450. }

  451. 

  452. static av_always_inline int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, DCTELEM *block, int last_index)

  453. {

  454.     int last_non_zero = 0;

  455.     int bits = 0;

  456.     int i, j, level;

  457.     for (i = 1; i <= last_index; i++) {

  458.         j = ctx->m.intra_scantable.permutated[i];

  459.         level = block[j];

  460.         if (level) {

  461.             int run_level = i - last_non_zero - 1;

  462.             bits += ctx->vlc_bits[(level<<1)|!!run_level]+ctx->run_bits[run_level];

  463.             last_non_zero = i;

  464.         }

  465.     }

  466.     return bits;

  467. }

  468. 

  469. static av_always_inline void dnxhd_get_blocks(DNXHDEncContext *ctx, int mb_x, int mb_y)

  470. {

  471.     const int bs = ctx->block_width_l2;

  472.     const int bw = 1 << bs;

  473.     const uint8_t *ptr_y = ctx->thread[0]->src[0] + ((mb_y << 4) * ctx->m.linesize)   + (mb_x << bs+1);

  474.     const uint8_t *ptr_u = ctx->thread[0]->src[1] + ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs);

  475.     const uint8_t *ptr_v = ctx->thread[0]->src[2] + ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs);

  476.     DSPContext *dsp = &ctx->m.dsp;

  477. 

  478.     dsp->get_pixels(ctx->blocks[0], ptr_y,      ctx->m.linesize);

  479.     dsp->get_pixels(ctx->blocks[1], ptr_y + bw, ctx->m.linesize);

  480.     dsp->get_pixels(ctx->blocks[2], ptr_u,      ctx->m.uvlinesize);

  481.     dsp->get_pixels(ctx->blocks[3], ptr_v,      ctx->m.uvlinesize);

  482. 

  483.     if (mb_y+1 == ctx->m.mb_height && ctx->m.avctx->height == 1080) {

  484.         if (ctx->interlaced) {

  485.             ctx->get_pixels_8x4_sym(ctx->blocks[4], ptr_y + ctx->dct_y_offset,      ctx->m.linesize);

  486.             ctx->get_pixels_8x4_sym(ctx->blocks[5], ptr_y + ctx->dct_y_offset + bw, ctx->m.linesize);

  487.             ctx->get_pixels_8x4_sym(ctx->blocks[6], ptr_u + ctx->dct_uv_offset,     ctx->m.uvlinesize);

  488.             ctx->get_pixels_8x4_sym(ctx->blocks[7], ptr_v + ctx->dct_uv_offset,     ctx->m.uvlinesize);

  489.         } else {

  490.             dsp->clear_block(ctx->blocks[4]);

  491.             dsp->clear_block(ctx->blocks[5]);

  492.             dsp->clear_block(ctx->blocks[6]);

  493.             dsp->clear_block(ctx->blocks[7]);

  494.         }

  495.     } else {

  496.         dsp->get_pixels(ctx->blocks[4], ptr_y + ctx->dct_y_offset,      ctx->m.linesize);

  497.         dsp->get_pixels(ctx->blocks[5], ptr_y + ctx->dct_y_offset + bw, ctx->m.linesize);

  498.         dsp->get_pixels(ctx->blocks[6], ptr_u + ctx->dct_uv_offset,     ctx->m.uvlinesize);

  499.         dsp->get_pixels(ctx->blocks[7], ptr_v + ctx->dct_uv_offset,     ctx->m.uvlinesize);

  500.     }

  501. }

  502. 

  503. static av_always_inline int dnxhd_switch_matrix(DNXHDEncContext *ctx, int i)

  504. {

  505.     const static uint8_t component[8]={0,0,1,2,0,0,1,2};

  506.     return component[i];

  507. }

  508. 

  509. static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg, int jobnr, int threadnr)

  510. {

  511.     DNXHDEncContext *ctx = avctx->priv_data;

  512.     int mb_y = jobnr, mb_x;

  513.     int qscale = ctx->qscale;

  514.     LOCAL_ALIGNED_16(DCTELEM, block, [64]);

  515.     ctx = ctx->thread[threadnr];

  516. 

  517.     ctx->m.last_dc[0] =

  518.     ctx->m.last_dc[1] =

  519.     ctx->m.last_dc[2] = 1 << (ctx->cid_table->bit_depth + 2);

  520. 

  521.     for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {

  522.         unsigned mb = mb_y * ctx->m.mb_width + mb_x;

  523.         int ssd     = 0;

  524.         int ac_bits = 0;

  525.         int dc_bits = 0;

  526.         int i;

  527. 

  528.         dnxhd_get_blocks(ctx, mb_x, mb_y);

  529. 

  530.         for (i = 0; i < 8; i++) {

  531.             DCTELEM *src_block = ctx->blocks[i];

  532.             int overflow, nbits, diff, last_index;

  533.             int n = dnxhd_switch_matrix(ctx, i);

  534. 

  535.             memcpy(block, src_block, 64*sizeof(*block));

  536.             last_index = ctx->m.dct_quantize(&ctx->m, block, 4&(2*i), qscale, &overflow);

  537.             ac_bits += dnxhd_calc_ac_bits(ctx, block, last_index);

  538. 

  539.             diff = block[0] - ctx->m.last_dc[n];

  540.             if (diff < 0) nbits = av_log2_16bit(-2*diff);

  541.             else          nbits = av_log2_16bit( 2*diff);

  542. 

  543.             assert(nbits < ctx->cid_table->bit_depth + 4);

  544.             dc_bits += ctx->cid_table->dc_bits[nbits] + nbits;

  545. 

  546.             ctx->m.last_dc[n] = block[0];

  547. 

  548.             if (avctx->mb_decision == FF_MB_DECISION_RD || !RC_VARIANCE) {

  549.                 dnxhd_unquantize_c(ctx, block, i, qscale, last_index);

  550.                 ctx->m.dsp.idct(block);

  551.                 ssd += dnxhd_ssd_block(block, src_block);

  552.             }

  553.         }

  554.         ctx->mb_rc[qscale][mb].ssd = ssd;

  555.         ctx->mb_rc[qscale][mb].bits = ac_bits+dc_bits+12+8*ctx->vlc_bits[0];

  556.     }

  557.     return 0;

  558. }

  559. 

  560. static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg, int jobnr, int threadnr)

  561. {

  562.     DNXHDEncContext *ctx = avctx->priv_data;

  563.     int mb_y = jobnr, mb_x;

  564.     ctx = ctx->thread[threadnr];

  565.     init_put_bits(&ctx->m.pb, (uint8_t *)arg + 640 + ctx->slice_offs[jobnr], ctx->slice_size[jobnr]);

  566. 

  567.     ctx->m.last_dc[0] =

  568.     ctx->m.last_dc[1] =

  569.     ctx->m.last_dc[2] = 1 << (ctx->cid_table->bit_depth + 2);

  570.     for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {

  571.         unsigned mb = mb_y * ctx->m.mb_width + mb_x;

  572.         int qscale = ctx->mb_qscale[mb];

  573.         int i;

  574. 

  575.         put_bits(&ctx->m.pb, 12, qscale<<1);

  576. 

  577.         dnxhd_get_blocks(ctx, mb_x, mb_y);

  578. 

  579.         for (i = 0; i < 8; i++) {

  580.             DCTELEM *block = ctx->blocks[i];

  581.             int last_index, overflow;

  582.             int n = dnxhd_switch_matrix(ctx, i);

  583.             last_index = ctx->m.dct_quantize(&ctx->m, block, 4&(2*i), qscale, &overflow);

  584.             //START_TIMER;

  585.             dnxhd_encode_block(ctx, block, last_index, n);

  586.             //STOP_TIMER("encode_block");

  587.         }

  588.     }

  589.     if (put_bits_count(&ctx->m.pb)&31)

  590.         put_bits(&ctx->m.pb, 32-(put_bits_count(&ctx->m.pb)&31), 0);

  591.     flush_put_bits(&ctx->m.pb);

  592.     return 0;

  593. }

  594. 

  595. static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx)

  596. {

  597.     int mb_y, mb_x;

  598.     int offset = 0;

  599.     for (mb_y = 0; mb_y < ctx->m.mb_height; mb_y++) {

  600.         int thread_size;

  601.         ctx->slice_offs[mb_y] = offset;

  602.         ctx->slice_size[mb_y] = 0;

  603.         for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {

  604.             unsigned mb = mb_y * ctx->m.mb_width + mb_x;

  605.             ctx->slice_size[mb_y] += ctx->mb_bits[mb];

  606.         }

  607.         ctx->slice_size[mb_y] = (ctx->slice_size[mb_y]+31)&~31;

  608.         ctx->slice_size[mb_y] >>= 3;

  609.         thread_size = ctx->slice_size[mb_y];

  610.         offset += thread_size;

  611.     }

  612. }

  613. 

  614. static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg, int jobnr, int threadnr)

  615. {

  616.     DNXHDEncContext *ctx = avctx->priv_data;

  617.     int mb_y = jobnr, mb_x;

  618.     ctx = ctx->thread[threadnr];

  619.     if (ctx->cid_table->bit_depth == 8) {

  620.         uint8_t *pix = ctx->thread[0]->src[0] + ((mb_y<<4) * ctx->m.linesize);

  621.         for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x, pix += 16) {

  622.             unsigned mb  = mb_y * ctx->m.mb_width + mb_x;

  623.             int sum = ctx->m.dsp.pix_sum(pix, ctx->m.linesize);

  624.             int varc = (ctx->m.dsp.pix_norm1(pix, ctx->m.linesize) - (((unsigned)(sum*sum))>>8)+128)>>8;

  625.             ctx->mb_cmp[mb].value = varc;

  626.             ctx->mb_cmp[mb].mb = mb;

  627.         }

  628.     } else { // 10-bit

  629.         int const linesize = ctx->m.linesize >> 1;

  630.         for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x) {

  631.             uint16_t *pix = (uint16_t*)ctx->thread[0]->src[0] + ((mb_y << 4) * linesize) + (mb_x << 4);

  632.             unsigned mb  = mb_y * ctx->m.mb_width + mb_x;

  633.             int sum = 0;

  634.             int sqsum = 0;

  635.             int mean, sqmean;

  636.             int i, j;

  637.             // Macroblocks are 16x16 pixels, unlike DCT blocks which are 8x8.

  638.             for (i = 0; i < 16; ++i) {

  639.                 for (j = 0; j < 16; ++j) {

  640.                     // Turn 16-bit pixels into 10-bit ones.

  641.                     int const sample = (unsigned)pix[j] >> 6;

  642.                     sum += sample;

  643.                     sqsum += sample * sample;

  644.                     // 2^10 * 2^10 * 16 * 16 = 2^28, which is less than INT_MAX

  645.                 }

  646.                 pix += linesize;

  647.             }

  648.             mean = sum >> 8; // 16*16 == 2^8

  649.             sqmean = sqsum >> 8;

  650.             ctx->mb_cmp[mb].value = sqmean - mean * mean;

  651.             ctx->mb_cmp[mb].mb = mb;

  652.         }

  653.     }

  654.     return 0;

  655. }

  656. 

  657. static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx)

  658. {

  659.     int lambda, up_step, down_step;

  660.     int last_lower = INT_MAX, last_higher = 0;

  661.     int x, y, q;

  662. 

  663.     for (q = 1; q < avctx->qmax; q++) {

  664.         ctx->qscale = q;

  665.         avctx->execute2(avctx, dnxhd_calc_bits_thread, NULL, NULL, ctx->m.mb_height);

  666.     }

  667.     up_step = down_step = 2<<LAMBDA_FRAC_BITS;

  668.     lambda = ctx->lambda;

  669. 

  670.     for (;;) {

  671.         int bits = 0;

  672.         int end = 0;

  673.         if (lambda == last_higher) {

  674.             lambda++;

  675.             end = 1; // need to set final qscales/bits

  676.         }

  677.         for (y = 0; y < ctx->m.mb_height; y++) {

  678.             for (x = 0; x < ctx->m.mb_width; x++) {

  679.                 unsigned min = UINT_MAX;

  680.                 int qscale = 1;

  681.                 int mb = y*ctx->m.mb_width+x;

  682.                 for (q = 1; q < avctx->qmax; q++) {

  683.                     unsigned score = ctx->mb_rc[q][mb].bits*lambda+(ctx->mb_rc[q][mb].ssd<<LAMBDA_FRAC_BITS);

  684.                     if (score < min) {

  685.                         min = score;

  686.                         qscale = q;

  687.                     }

  688.                 }

  689.                 bits += ctx->mb_rc[qscale][mb].bits;

  690.                 ctx->mb_qscale[mb] = qscale;

  691.                 ctx->mb_bits[mb] = ctx->mb_rc[qscale][mb].bits;

  692.             }

  693.             bits = (bits+31)&~31; // padding

  694.             if (bits > ctx->frame_bits)

  695.                 break;

  696.         }

  697.         //av_dlog(ctx->m.avctx, "lambda %d, up %u, down %u, bits %d, frame %d\n",

  698.         //        lambda, last_higher, last_lower, bits, ctx->frame_bits);

  699.         if (end) {

  700.             if (bits > ctx->frame_bits)

  701.                 return -1;

  702.             break;

  703.         }

  704.         if (bits < ctx->frame_bits) {

  705.             last_lower = FFMIN(lambda, last_lower);

  706.             if (last_higher != 0)

  707.                 lambda = (lambda+last_higher)>>1;

  708.             else

  709.                 lambda -= down_step;

  710.             down_step *= 5; // XXX tune ?

  711.             up_step = 1<<LAMBDA_FRAC_BITS;

  712.             lambda = FFMAX(1, lambda);

  713.             if (lambda == last_lower)

  714.                 break;

  715.         } else {

  716.             last_higher = FFMAX(lambda, last_higher);

  717.             if (last_lower != INT_MAX)

  718.                 lambda = (lambda+last_lower)>>1;

  719.             else if ((int64_t)lambda + up_step > INT_MAX)

  720.                 return -1;

  721.             else

  722.                 lambda += up_step;

  723.             up_step = FFMIN((int64_t)up_step*5, INT_MAX);

  724.             down_step = 1<<LAMBDA_FRAC_BITS;

  725.         }

  726.     }

  727.     //av_dlog(ctx->m.avctx, "out lambda %d\n", lambda);

  728.     ctx->lambda = lambda;

  729.     return 0;

  730. }

  731. 

  732. static int dnxhd_find_qscale(DNXHDEncContext *ctx)

  733. {

  734.     int bits = 0;

  735.     int up_step = 1;

  736.     int down_step = 1;

  737.     int last_higher = 0;

  738.     int last_lower = INT_MAX;

  739.     int qscale;

  740.     int x, y;

  741. 

  742.     qscale = ctx->qscale;

  743.     for (;;) {

  744.         bits = 0;

  745.         ctx->qscale = qscale;

  746.         // XXX avoid recalculating bits

  747.         ctx->m.avctx->execute2(ctx->m.avctx, dnxhd_calc_bits_thread, NULL, NULL, ctx->m.mb_height);

  748.         for (y = 0; y < ctx->m.mb_height; y++) {

  749.             for (x = 0; x < ctx->m.mb_width; x++)

  750.                 bits += ctx->mb_rc[qscale][y*ctx->m.mb_width+x].bits;

  751.             bits = (bits+31)&~31; // padding

  752.             if (bits > ctx->frame_bits)

  753.                 break;

  754.         }

  755.         //av_dlog(ctx->m.avctx, "%d, qscale %d, bits %d, frame %d, higher %d, lower %d\n",

  756.         //        ctx->m.avctx->frame_number, qscale, bits, ctx->frame_bits, last_higher, last_lower);

  757.         if (bits < ctx->frame_bits) {

  758.             if (qscale == 1)

  759.                 return 1;

  760.             if (last_higher == qscale - 1) {

  761.                 qscale = last_higher;

  762.                 break;

  763.             }

  764.             last_lower = FFMIN(qscale, last_lower);

  765.             if (last_higher != 0)

  766.                 qscale = (qscale+last_higher)>>1;

  767.             else

  768.                 qscale -= down_step++;

  769.             if (qscale < 1)

  770.                 qscale = 1;

  771.             up_step = 1;

  772.         } else {

  773.             if (last_lower == qscale + 1)

  774.                 break;

  775.             last_higher = FFMAX(qscale, last_higher);

  776.             if (last_lower != INT_MAX)

  777.                 qscale = (qscale+last_lower)>>1;

  778.             else

  779.                 qscale += up_step++;

  780.             down_step = 1;

  781.             if (qscale >= ctx->m.avctx->qmax)

  782.                 return -1;

  783.         }

  784.     }

  785.     //av_dlog(ctx->m.avctx, "out qscale %d\n", qscale);

  786.     ctx->qscale = qscale;

  787.     return 0;

  788. }

  789. 

  790. #define BUCKET_BITS 8

  791. #define RADIX_PASSES 4

  792. #define NBUCKETS (1 << BUCKET_BITS)

  793. 

  794. static inline int get_bucket(int value, int shift)

  795. {

  796.     value >>= shift;

  797.     value &= NBUCKETS - 1;

  798.     return NBUCKETS - 1 - value;

  799. }

  800. 

  801. static void radix_count(const RCCMPEntry *data, int size, int buckets[RADIX_PASSES][NBUCKETS])

  802. {

  803.     int i, j;

  804.     memset(buckets, 0, sizeof(buckets[0][0]) * RADIX_PASSES * NBUCKETS);

  805.     for (i = 0; i < size; i++) {

  806.         int v = data[i].value;

  807.         for (j = 0; j < RADIX_PASSES; j++) {

  808.             buckets[j][get_bucket(v, 0)]++;

  809.             v >>= BUCKET_BITS;

  810.         }

  811.         assert(!v);

  812.     }

  813.     for (j = 0; j < RADIX_PASSES; j++) {

  814.         int offset = size;

  815.         for (i = NBUCKETS - 1; i >= 0; i--)

  816.             buckets[j][i] = offset -= buckets[j][i];

  817.         assert(!buckets[j][0]);

  818.     }

  819. }

  820. 

  821. static void radix_sort_pass(RCCMPEntry *dst, const RCCMPEntry *data, int size, int buckets[NBUCKETS], int pass)

  822. {

  823.     int shift = pass * BUCKET_BITS;

  824.     int i;

  825.     for (i = 0; i < size; i++) {

  826.         int v = get_bucket(data[i].value, shift);

  827.         int pos = buckets[v]++;

  828.         dst[pos] = data[i];

  829.     }

  830. }

  831. 

  832. static void radix_sort(RCCMPEntry *data, int size)

  833. {

  834.     int buckets[RADIX_PASSES][NBUCKETS];

  835.     RCCMPEntry *tmp = av_malloc(sizeof(*tmp) * size);

  836.     radix_count(data, size, buckets);

  837.     radix_sort_pass(tmp, data, size, buckets[0], 0);

  838.     radix_sort_pass(data, tmp, size, buckets[1], 1);

  839.     if (buckets[2][NBUCKETS - 1] || buckets[3][NBUCKETS - 1]) {

  840.         radix_sort_pass(tmp, data, size, buckets[2], 2);

  841.         radix_sort_pass(data, tmp, size, buckets[3], 3);

  842.     }

  843.     av_free(tmp);

  844. }

  845. 

  846. static int dnxhd_encode_fast(AVCodecContext *avctx, DNXHDEncContext *ctx)

  847. {

  848.     int max_bits = 0;

  849.     int ret, x, y;

  850.     if ((ret = dnxhd_find_qscale(ctx)) < 0)

  851.         return -1;

  852.     for (y = 0; y < ctx->m.mb_height; y++) {

  853.         for (x = 0; x < ctx->m.mb_width; x++) {

  854.             int mb = y*ctx->m.mb_width+x;

  855.             int delta_bits;

  856.             ctx->mb_qscale[mb] = ctx->qscale;

  857.             ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale][mb].bits;

  858.             max_bits += ctx->mb_rc[ctx->qscale][mb].bits;

  859.             if (!RC_VARIANCE) {

  860.                 delta_bits = ctx->mb_rc[ctx->qscale][mb].bits-ctx->mb_rc[ctx->qscale+1][mb].bits;

  861.                 ctx->mb_cmp[mb].mb = mb;

  862.                 ctx->mb_cmp[mb].value = delta_bits ?

  863.                     ((ctx->mb_rc[ctx->qscale][mb].ssd-ctx->mb_rc[ctx->qscale+1][mb].ssd)*100)/delta_bits

  864.                     : INT_MIN; //avoid increasing qscale

  865.             }

  866.         }

  867.         max_bits += 31; //worst padding

  868.     }

  869.     if (!ret) {

  870.         if (RC_VARIANCE)

  871.             avctx->execute2(avctx, dnxhd_mb_var_thread, NULL, NULL, ctx->m.mb_height);

  872.         radix_sort(ctx->mb_cmp, ctx->m.mb_num);

  873.         for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) {

  874.             int mb = ctx->mb_cmp[x].mb;

  875.             max_bits -= ctx->mb_rc[ctx->qscale][mb].bits - ctx->mb_rc[ctx->qscale+1][mb].bits;

  876.             ctx->mb_qscale[mb] = ctx->qscale+1;

  877.             ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale+1][mb].bits;

  878.         }

  879.     }

  880.     return 0;

  881. }

  882. 

  883. static void dnxhd_load_picture(DNXHDEncContext *ctx, const AVFrame *frame)

  884. {

  885.     int i;

  886. 

  887.     for (i = 0; i < 3; i++) {

  888.         ctx->frame.data[i]     = frame->data[i];

  889.         ctx->frame.linesize[i] = frame->linesize[i];

  890.     }

  891. 

  892.     for (i = 0; i < ctx->m.avctx->thread_count; i++) {

  893.         ctx->thread[i]->m.linesize    = ctx->frame.linesize[0]<<ctx->interlaced;

  894.         ctx->thread[i]->m.uvlinesize  = ctx->frame.linesize[1]<<ctx->interlaced;

  895.         ctx->thread[i]->dct_y_offset  = ctx->m.linesize  *8;

  896.         ctx->thread[i]->dct_uv_offset = ctx->m.uvlinesize*8;

  897.     }

  898. 

  899.     ctx->frame.interlaced_frame = frame->interlaced_frame;

  900.     ctx->cur_field = frame->interlaced_frame && !frame->top_field_first;

  901. }

  902. 

  903. static int dnxhd_encode_picture(AVCodecContext *avctx, unsigned char *buf, int buf_size, void *data)

  904. {

  905.     DNXHDEncContext *ctx = avctx->priv_data;

  906.     int first_field = 1;

  907.     int offset, i, ret;

  908. 

  909.     if (buf_size < ctx->cid_table->frame_size) {

  910.         av_log(avctx, AV_LOG_ERROR, "output buffer is too small to compress picture\n");

  911.         return -1;

  912.     }

  913. 

  914.     dnxhd_load_picture(ctx, data);

  915. 

  916.  encode_coding_unit:

  917.     for (i = 0; i < 3; i++) {

  918.         ctx->src[i] = ctx->frame.data[i];

  919.         if (ctx->interlaced && ctx->cur_field)

  920.             ctx->src[i] += ctx->frame.linesize[i];

  921.     }

  922. 

  923.     dnxhd_write_header(avctx, buf);

  924. 

  925.     if (avctx->mb_decision == FF_MB_DECISION_RD)

  926.         ret = dnxhd_encode_rdo(avctx, ctx);

  927.     else

  928.         ret = dnxhd_encode_fast(avctx, ctx);

  929.     if (ret < 0) {

  930.         av_log(avctx, AV_LOG_ERROR,

  931.                "picture could not fit ratecontrol constraints, increase qmax\n");

  932.         return -1;

  933.     }

  934. 

  935.     dnxhd_setup_threads_slices(ctx);

  936. 

  937.     offset = 0;

  938.     for (i = 0; i < ctx->m.mb_height; i++) {

  939.         AV_WB32(ctx->msip + i * 4, offset);

  940.         offset += ctx->slice_size[i];

  941.         assert(!(ctx->slice_size[i] & 3));

  942.     }

  943. 

  944.     avctx->execute2(avctx, dnxhd_encode_thread, buf, NULL, ctx->m.mb_height);

  945. 

  946.     assert(640 + offset + 4 <= ctx->cid_table->coding_unit_size);

  947.     memset(buf + 640 + offset, 0, ctx->cid_table->coding_unit_size - 4 - offset - 640);

  948. 

  949.     AV_WB32(buf + ctx->cid_table->coding_unit_size - 4, 0x600DC0DE); // EOF

  950. 

  951.     if (ctx->interlaced && first_field) {

  952.         first_field     = 0;

  953.         ctx->cur_field ^= 1;

  954.         buf      += ctx->cid_table->coding_unit_size;

  955.         buf_size -= ctx->cid_table->coding_unit_size;

  956.         goto encode_coding_unit;

  957.     }

  958. 

  959.     ctx->frame.quality = ctx->qscale*FF_QP2LAMBDA;

  960. 

  961.     return ctx->cid_table->frame_size;

  962. }

  963. 

  964. static int dnxhd_encode_end(AVCodecContext *avctx)

  965. {

  966.     DNXHDEncContext *ctx = avctx->priv_data;

  967.     int max_level = 1<<(ctx->cid_table->bit_depth+2);

  968.     int i;

  969. 

  970.     av_free(ctx->vlc_codes-max_level*2);

  971.     av_free(ctx->vlc_bits -max_level*2);

  972.     av_freep(&ctx->run_codes);

  973.     av_freep(&ctx->run_bits);

  974. 

  975.     av_freep(&ctx->mb_bits);

  976.     av_freep(&ctx->mb_qscale);

  977.     av_freep(&ctx->mb_rc);

  978.     av_freep(&ctx->mb_cmp);

  979.     av_freep(&ctx->slice_size);

  980.     av_freep(&ctx->slice_offs);

  981. 

  982.     av_freep(&ctx->qmatrix_c);

  983.     av_freep(&ctx->qmatrix_l);

  984.     av_freep(&ctx->qmatrix_c16);

  985.     av_freep(&ctx->qmatrix_l16);

  986. 

  987.     for (i = 1; i < avctx->thread_count; i++)

  988.         av_freep(&ctx->thread[i]);

  989. 

  990.     return 0;

  991. }

  992. 

  993. AVCodec ff_dnxhd_encoder = {

  994.     .name           = "dnxhd",

  995.     .type           = AVMEDIA_TYPE_VIDEO,

  996.     .id             = CODEC_ID_DNXHD,

  997.     .priv_data_size = sizeof(DNXHDEncContext),

  998.     .init           = dnxhd_encode_init,

  999.     .encode         = dnxhd_encode_picture,

  1000.     .close          = dnxhd_encode_end,

  1001.     .capabilities = CODEC_CAP_SLICE_THREADS,

  1002.     .pix_fmts = (const enum PixelFormat[]){PIX_FMT_YUV422P, PIX_FMT_YUV422P10, PIX_FMT_NONE},

  1003.     .long_name = NULL_IF_CONFIG_SMALL("VC3/DNxHD"),

  1004.     .priv_class = &class,

  1005. };