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.
37645 Commits
32 Branches
369MB
Tree: d97fa416d6
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'd97fa416d6'
${ noResults }
FFmpeg/libavcodec/dnxhdenc.c

1013 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. #include "internal.h"

  36. 

  37. #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM

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

  39. 

  40. static const AVOption options[]={

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

  42. {NULL}

  43. };

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

  45. 

  46. #define LAMBDA_FRAC_BITS 10

  47. 

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

  49. {

  50.     int i;

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

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

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

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

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

  56.         pixels += line_size;

  57.         block += 8;

  58.     }

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

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

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

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

  63. }

  64. 

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

  66. {

  67.     int i;

  68. 

  69.     block += 32;

  70. 

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

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

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

  74.     }

  75. }

  76. 

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

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

  79. {

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

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

  82.     int last_non_zero = 0;

  83.     int i;

  84. 

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

  86. 

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

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

  89. 

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

  91.         int j = scantable[i];

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

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

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

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

  96.         if (level)

  97.             last_non_zero = i;

  98.     }

  99. 

  100.     return last_non_zero;

  101. }

  102. 

  103. static int dnxhd_init_vlc(DNXHDEncContext *ctx)

  104. {

  105.     int i, j, level, run;

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

  107. 

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

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

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

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

  112. 

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

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

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

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

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

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

  119. 

  120.             MASK_ABS(sign, alevel);

  121.             if (alevel > 64) {

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

  123.                 alevel -= offset<<6;

  124.             }

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

  126.                 if (ctx->cid_table->ac_level[j] >> 1 == alevel &&

  127.                     (!offset || (ctx->cid_table->ac_flags[j] & 1) && offset) &&

  128.                     (!run    || (ctx->cid_table->ac_flags[j] & 2) && run)) {

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

  130.                     if (alevel) {

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

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

  133.                     } else {

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

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

  136.                     }

  137.                     break;

  138.                 }

  139.             }

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

  141.             if (offset) {

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

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

  144.             }

  145.         }

  146.     }

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

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

  149.         assert(run < 63);

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

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

  152.     }

  153.     return 0;

  154.  fail:

  155.     return -1;

  156. }

  157. 

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

  159. {

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

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

  162.     int qscale, i;

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

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

  165. 

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

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

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

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

  170. 

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

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

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

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

  175.         }

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

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

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

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

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

  181.         }

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

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

  184. 

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

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

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

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

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

  190.             }

  191.         }

  192.     } else {

  193.         // 10-bit

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

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

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

  197. 

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

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

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

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

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

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

  204.                 // ((1 << DNX10BIT_QMAT_SHIFT) * (p / s)) / (qscale * weight_table[i])

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

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

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

  208.             }

  209.         }

  210.     }

  211. 

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

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

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

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

  216. 

  217.     return 0;

  218.  fail:

  219.     return -1;

  220. }

  221. 

  222. static int dnxhd_init_rc(DNXHDEncContext *ctx)

  223. {

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

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

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

  227. 

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

  229.     ctx->qscale = 1;

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

  231.     return 0;

  232.  fail:

  233.     return -1;

  234. }

  235. 

  236. static int dnxhd_encode_init(AVCodecContext *avctx)

  237. {

  238.     DNXHDEncContext *ctx = avctx->priv_data;

  239.     int i, index, bit_depth;

  240. 

  241.     switch (avctx->pix_fmt) {

  242.     case PIX_FMT_YUV422P:

  243.         bit_depth = 8;

  244.         break;

  245.     case PIX_FMT_YUV422P10:

  246.         bit_depth = 10;

  247.         break;

  248.     default:

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

  250.         return -1;

  251.     }

  252. 

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

  254.     if (!ctx->cid) {

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

  256.         return -1;

  257.     }

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

  259. 

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

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

  262. 

  263.     ctx->m.avctx = avctx;

  264.     ctx->m.mb_intra = 1;

  265.     ctx->m.h263_aic = 1;

  266. 

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

  268. 

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

  270.     ff_dct_common_init(&ctx->m);

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

  272.         ctx->m.dct_quantize = dct_quantize_c;

  273. 

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

  275.        ctx->m.dct_quantize = dnxhd_10bit_dct_quantize;

  276.        ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym;

  277.        ctx->block_width_l2 = 4;

  278.     } else {

  279.        ctx->get_pixels_8x4_sym = dnxhd_8bit_get_pixels_8x4_sym;

  280.        ctx->block_width_l2 = 3;

  281.     }

  282. 

  283. #if HAVE_MMX

  284.     ff_dnxhd_init_mmx(ctx);

  285. #endif

  286. 

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

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

  289. 

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

  291.         ctx->interlaced = 1;

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

  293.     }

  294. 

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

  296. 

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

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

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

  300.         return -1;

  301. 

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

  303.     if (ctx->nitris_compat)

  304.         ctx->min_padding = 1600;

  305. 

  306.     if (dnxhd_init_vlc(ctx) < 0)

  307.         return -1;

  308.     if (dnxhd_init_rc(ctx) < 0)

  309.         return -1;

  310. 

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

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

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

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

  315. 

  316.     ctx->frame.key_frame = 1;

  317.     ctx->frame.pict_type = AV_PICTURE_TYPE_I;

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

  319. 

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

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

  322.         return -1;

  323.     }

  324. 

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

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

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

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

  329.     }

  330. 

  331.     return 0;

  332.  fail: //for FF_ALLOCZ_OR_GOTO

  333.     return -1;

  334. }

  335. 

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

  337. {

  338.     DNXHDEncContext *ctx = avctx->priv_data;

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

  340. 

  341.     memset(buf, 0, 640);

  342. 

  343.     memcpy(buf, header_prefix, 5);

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

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

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

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

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

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

  350. 

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

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

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

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

  355. 

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

  357. 

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

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

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

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

  362. 

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

  364.     return 0;

  365. }

  366. 

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

  368. {

  369.     int nbits;

  370.     if (diff < 0) {

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

  372.         diff--;

  373.     } else {

  374.         nbits = av_log2_16bit(2*diff);

  375.     }

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

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

  378. }

  379. 

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

  381. {

  382.     int last_non_zero = 0;

  383.     int slevel, i, j;

  384. 

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

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

  387. 

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

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

  390.         slevel = block[j];

  391.         if (slevel) {

  392.             int run_level = i - last_non_zero - 1;

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

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

  395.             if (run_level)

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

  397.             last_non_zero = i;

  398.         }

  399.     }

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

  401. }

  402. 

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

  404. {

  405.     const uint8_t *weight_matrix;

  406.     int level;

  407.     int i;

  408. 

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

  410. 

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

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

  413.         level = block[j];

  414.         if (level) {

  415.             if (level < 0) {

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

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

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

  419.                         level += 8;

  420.                     level >>= 4;

  421.                 } else {

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

  423.                         level += 32;

  424.                     level >>= 6;

  425.                 }

  426.                 level = -level;

  427.             } else {

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

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

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

  431.                         level += 8;

  432.                     level >>= 4;

  433.                 } else {

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

  435.                         level += 32;

  436.                     level >>= 6;

  437.                 }

  438.             }

  439.             block[j] = level;

  440.         }

  441.     }

  442. }

  443. 

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

  445. {

  446.     int score = 0;

  447.     int i;

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

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

  450.     return score;

  451. }

  452. 

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

  454. {

  455.     int last_non_zero = 0;

  456.     int bits = 0;

  457.     int i, j, level;

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

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

  460.         level = block[j];

  461.         if (level) {

  462.             int run_level = i - last_non_zero - 1;

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

  464.             last_non_zero = i;

  465.         }

  466.     }

  467.     return bits;

  468. }

  469. 

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

  471. {

  472.     const int bs = ctx->block_width_l2;

  473.     const int bw = 1 << bs;

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

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

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

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

  478. 

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

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

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

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

  483. 

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

  485.         if (ctx->interlaced) {

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

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

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

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

  490.         } else {

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

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

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

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

  495.         }

  496.     } else {

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

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

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

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

  501.     }

  502. }

  503. 

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

  505. {

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

  507.     return component[i];

  508. }

  509. 

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

  511. {

  512.     DNXHDEncContext *ctx = avctx->priv_data;

  513.     int mb_y = jobnr, mb_x;

  514.     int qscale = ctx->qscale;

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

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

  517. 

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

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

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

  521. 

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

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

  524.         int ssd     = 0;

  525.         int ac_bits = 0;

  526.         int dc_bits = 0;

  527.         int i;

  528. 

  529.         dnxhd_get_blocks(ctx, mb_x, mb_y);

  530. 

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

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

  533.             int overflow, nbits, diff, last_index;

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

  535. 

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

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

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

  539. 

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

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

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

  543. 

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

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

  546. 

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

  548. 

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

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

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

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

  553.             }

  554.         }

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

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

  557.     }

  558.     return 0;

  559. }

  560. 

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

  562. {

  563.     DNXHDEncContext *ctx = avctx->priv_data;

  564.     int mb_y = jobnr, mb_x;

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

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

  567. 

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

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

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

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

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

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

  574.         int i;

  575. 

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

  577. 

  578.         dnxhd_get_blocks(ctx, mb_x, mb_y);

  579. 

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

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

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

  583.             int 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+

  684.                         ((unsigned)ctx->mb_rc[q][mb].ssd<<LAMBDA_FRAC_BITS);

  685.                     if (score < min) {

  686.                         min = score;

  687.                         qscale = q;

  688.                     }

  689.                 }

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

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

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

  693.             }

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

  695.             if (bits > ctx->frame_bits)

  696.                 break;

  697.         }

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

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

  700.         if (end) {

  701.             if (bits > ctx->frame_bits)

  702.                 return -1;

  703.             break;

  704.         }

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

  706.             last_lower = FFMIN(lambda, last_lower);

  707.             if (last_higher != 0)

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

  709.             else

  710.                 lambda -= down_step;

  711.             down_step = FFMIN((int64_t)down_step*5, INT_MAX);

  712.             up_step = 1<<LAMBDA_FRAC_BITS;

  713.             lambda = FFMAX(1, lambda);

  714.             if (lambda == last_lower)

  715.                 break;

  716.         } else {

  717.             last_higher = FFMAX(lambda, last_higher);

  718.             if (last_lower != INT_MAX)

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

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

  721.                 return -1;

  722.             else

  723.                 lambda += up_step;

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

  725.             down_step = 1<<LAMBDA_FRAC_BITS;

  726.         }

  727.     }

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

  729.     ctx->lambda = lambda;

  730.     return 0;

  731. }

  732. 

  733. static int dnxhd_find_qscale(DNXHDEncContext *ctx)

  734. {

  735.     int bits = 0;

  736.     int up_step = 1;

  737.     int down_step = 1;

  738.     int last_higher = 0;

  739.     int last_lower = INT_MAX;

  740.     int qscale;

  741.     int x, y;

  742. 

  743.     qscale = ctx->qscale;

  744.     for (;;) {

  745.         bits = 0;

  746.         ctx->qscale = qscale;

  747.         // XXX avoid recalculating bits

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

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

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

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

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

  753.             if (bits > ctx->frame_bits)

  754.                 break;

  755.         }

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

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

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

  759.             if (qscale == 1)

  760.                 return 1;

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

  762.                 qscale = last_higher;

  763.                 break;

  764.             }

  765.             last_lower = FFMIN(qscale, last_lower);

  766.             if (last_higher != 0)

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

  768.             else

  769.                 qscale -= down_step++;

  770.             if (qscale < 1)

  771.                 qscale = 1;

  772.             up_step = 1;

  773.         } else {

  774.             if (last_lower == qscale + 1)

  775.                 break;

  776.             last_higher = FFMAX(qscale, last_higher);

  777.             if (last_lower != INT_MAX)

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

  779.             else

  780.                 qscale += up_step++;

  781.             down_step = 1;

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

  783.                 return -1;

  784.         }

  785.     }

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

  787.     ctx->qscale = qscale;

  788.     return 0;

  789. }

  790. 

  791. #define BUCKET_BITS 8

  792. #define RADIX_PASSES 4

  793. #define NBUCKETS (1 << BUCKET_BITS)

  794. 

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

  796. {

  797.     value >>= shift;

  798.     value &= NBUCKETS - 1;

  799.     return NBUCKETS - 1 - value;

  800. }

  801. 

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

  803. {

  804.     int i, j;

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

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

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

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

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

  810.             v >>= BUCKET_BITS;

  811.         }

  812.         assert(!v);

  813.     }

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

  815.         int offset = size;

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

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

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

  819.     }

  820. }

  821. 

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

  823. {

  824.     int shift = pass * BUCKET_BITS;

  825.     int i;

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

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

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

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

  830.     }

  831. }

  832. 

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

  834. {

  835.     int buckets[RADIX_PASSES][NBUCKETS];

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

  837.     radix_count(data, size, buckets);

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

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

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

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

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

  843.     }

  844.     av_free(tmp);

  845. }

  846. 

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

  848. {

  849.     int max_bits = 0;

  850.     int ret, x, y;

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

  852.         return -1;

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

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

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

  856.             int delta_bits;

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

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

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

  860.             if (!RC_VARIANCE) {

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

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

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

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

  865.                     : INT_MIN; //avoid increasing qscale

  866.             }

  867.         }

  868.         max_bits += 31; //worst padding

  869.     }

  870.     if (!ret) {

  871.         if (RC_VARIANCE)

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

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

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

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

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

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

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

  879.         }

  880.     }

  881.     return 0;

  882. }

  883. 

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

  885. {

  886.     int i;

  887. 

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

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

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

  891.     }

  892. 

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

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

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

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

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

  898.     }

  899. 

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

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

  902. }

  903. 

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

  905. {

  906.     DNXHDEncContext *ctx = avctx->priv_data;

  907.     int first_field = 1;

  908.     int offset, i, ret;

  909. 

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

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

  912.         return -1;

  913.     }

  914. 

  915.     dnxhd_load_picture(ctx, data);

  916. 

  917.  encode_coding_unit:

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

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

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

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

  922.     }

  923. 

  924.     dnxhd_write_header(avctx, buf);

  925. 

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

  927.         ret = dnxhd_encode_rdo(avctx, ctx);

  928.     else

  929.         ret = dnxhd_encode_fast(avctx, ctx);

  930.     if (ret < 0) {

  931.         av_log(avctx, AV_LOG_ERROR,

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

  933.         return -1;

  934.     }

  935. 

  936.     dnxhd_setup_threads_slices(ctx);

  937. 

  938.     offset = 0;

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

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

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

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

  943.     }

  944. 

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

  946. 

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

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

  949. 

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

  951. 

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

  953.         first_field     = 0;

  954.         ctx->cur_field ^= 1;

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

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

  957.         goto encode_coding_unit;

  958.     }

  959. 

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

  961. 

  962.     return ctx->cid_table->frame_size;

  963. }

  964. 

  965. static int dnxhd_encode_end(AVCodecContext *avctx)

  966. {

  967.     DNXHDEncContext *ctx = avctx->priv_data;

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

  969.     int i;

  970. 

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

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

  973.     av_freep(&ctx->run_codes);

  974.     av_freep(&ctx->run_bits);

  975. 

  976.     av_freep(&ctx->mb_bits);

  977.     av_freep(&ctx->mb_qscale);

  978.     av_freep(&ctx->mb_rc);

  979.     av_freep(&ctx->mb_cmp);

  980.     av_freep(&ctx->slice_size);

  981.     av_freep(&ctx->slice_offs);

  982. 

  983.     av_freep(&ctx->qmatrix_c);

  984.     av_freep(&ctx->qmatrix_l);

  985.     av_freep(&ctx->qmatrix_c16);

  986.     av_freep(&ctx->qmatrix_l16);

  987. 

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

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

  990. 

  991.     return 0;

  992. }

  993. 

  994. static const AVCodecDefault dnxhd_defaults[] = {

  995.     { "qmax", "1024" }, /* Maximum quantization scale factor allowed for VC-3 */

  996.     { NULL },

  997. };

  998. 

  999. AVCodec ff_dnxhd_encoder = {

  1000.     .name           = "dnxhd",

  1001.     .type           = AVMEDIA_TYPE_VIDEO,

  1002.     .id             = CODEC_ID_DNXHD,

  1003.     .priv_data_size = sizeof(DNXHDEncContext),

  1004.     .init           = dnxhd_encode_init,

  1005.     .encode         = dnxhd_encode_picture,

  1006.     .close          = dnxhd_encode_end,

  1007.     .capabilities = CODEC_CAP_SLICE_THREADS,

  1008.     .pix_fmts = (const enum PixelFormat[]){PIX_FMT_YUV422P, PIX_FMT_YUV422P10, PIX_FMT_NONE},

  1009.     .long_name = NULL_IF_CONFIG_SMALL("VC3/DNxHD"),

  1010.     .priv_class = &class,

  1011.     .defaults       = dnxhd_defaults,

  1012. };