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.
38308 Commits
32 Branches
850MB
Tree: ce1e8045e0
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'ce1e8045e0'
${ noResults }
FFmpeg/libavcodec/dnxhdenc.c

1036 lines
36KB
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 Libav.

  10.  *

  11.  * Libav 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.  * Libav 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 Libav; if not, write to the Free Software

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

  24.  */

  25. 

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

  27. 

  28. #include "libavutil/attributes.h"

  29. #include "libavutil/internal.h"

  30. #include "libavutil/opt.h"

  31. #include "avcodec.h"

  32. #include "dsputil.h"

  33. #include "internal.h"

  34. #include "mpegvideo.h"

  35. #include "dnxhdenc.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, {.i64 = 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(int16_t *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(int16_t *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, int16_t *block,

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

  79. {

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

  81.     const int *qmat = ctx->q_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 av_cold 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] == alevel &&

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

  128.                     (!run    || (ctx->cid_table->ac_run_flag  [j] && 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 av_cold 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.     return 0;

  213.  fail:

  214.     return -1;

  215. }

  216. 

  217. static av_cold int dnxhd_init_rc(DNXHDEncContext *ctx)

  218. {

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

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

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

  222. 

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

  224.     ctx->qscale = 1;

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

  226.     return 0;

  227.  fail:

  228.     return -1;

  229. }

  230. 

  231. static av_cold int dnxhd_encode_init(AVCodecContext *avctx)

  232. {

  233.     DNXHDEncContext *ctx = avctx->priv_data;

  234.     int i, index, bit_depth;

  235. 

  236.     switch (avctx->pix_fmt) {

  237.     case AV_PIX_FMT_YUV422P:

  238.         bit_depth = 8;

  239.         break;

  240.     case AV_PIX_FMT_YUV422P10:

  241.         bit_depth = 10;

  242.         break;

  243.     default:

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

  245.         return -1;

  246.     }

  247. 

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

  249.     if (!ctx->cid) {

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

  251.         return -1;

  252.     }

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

  254. 

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

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

  257. 

  258.     ctx->m.avctx = avctx;

  259.     ctx->m.mb_intra = 1;

  260.     ctx->m.h263_aic = 1;

  261. 

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

  263. 

  264.     ff_dsputil_init(&ctx->m.dsp, avctx);

  265.     ff_dct_common_init(&ctx->m);

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

  267.         ctx->m.dct_quantize = ff_dct_quantize_c;

  268. 

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

  270.        ctx->m.dct_quantize = dnxhd_10bit_dct_quantize;

  271.        ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym;

  272.        ctx->block_width_l2 = 4;

  273.     } else {

  274.        ctx->get_pixels_8x4_sym = dnxhd_8bit_get_pixels_8x4_sym;

  275.        ctx->block_width_l2 = 3;

  276.     }

  277. 

  278.     if (ARCH_X86)

  279.         ff_dnxhdenc_init_x86(ctx);

  280. 

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

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

  283. 

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

  285.         ctx->interlaced = 1;

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

  287.     }

  288. 

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

  290. 

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

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

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

  294.         return -1;

  295. 

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

  297.     if (ctx->nitris_compat)

  298.         ctx->min_padding = 1600;

  299. 

  300.     if (dnxhd_init_vlc(ctx) < 0)

  301.         return -1;

  302.     if (dnxhd_init_rc(ctx) < 0)

  303.         return -1;

  304. 

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

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

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

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

  309. 

  310.     ctx->frame.key_frame = 1;

  311.     ctx->frame.pict_type = AV_PICTURE_TYPE_I;

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

  313. 

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

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

  316.         return -1;

  317.     }

  318. 

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

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

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

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

  323.     }

  324. 

  325.     return 0;

  326.  fail: //for FF_ALLOCZ_OR_GOTO

  327.     return -1;

  328. }

  329. 

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

  331. {

  332.     DNXHDEncContext *ctx = avctx->priv_data;

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

  334. 

  335.     memset(buf, 0, 640);

  336. 

  337.     memcpy(buf, header_prefix, 5);

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

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

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

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

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

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

  344. 

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

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

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

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

  349. 

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

  351. 

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

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

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

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

  356. 

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

  358.     return 0;

  359. }

  360. 

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

  362. {

  363.     int nbits;

  364.     if (diff < 0) {

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

  366.         diff--;

  367.     } else {

  368.         nbits = av_log2_16bit(2*diff);

  369.     }

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

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

  372. }

  373. 

  374. static av_always_inline void dnxhd_encode_block(DNXHDEncContext *ctx, int16_t *block, int last_index, int n)

  375. {

  376.     int last_non_zero = 0;

  377.     int slevel, i, j;

  378. 

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

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

  381. 

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

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

  384.         slevel = block[j];

  385.         if (slevel) {

  386.             int run_level = i - last_non_zero - 1;

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

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

  389.             if (run_level)

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

  391.             last_non_zero = i;

  392.         }

  393.     }

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

  395. }

  396. 

  397. static av_always_inline void dnxhd_unquantize_c(DNXHDEncContext *ctx, int16_t *block, int n, int qscale, int last_index)

  398. {

  399.     const uint8_t *weight_matrix;

  400.     int level;

  401.     int i;

  402. 

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

  404. 

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

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

  407.         level = block[j];

  408.         if (level) {

  409.             if (level < 0) {

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

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

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

  413.                         level += 8;

  414.                     level >>= 4;

  415.                 } else {

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

  417.                         level += 32;

  418.                     level >>= 6;

  419.                 }

  420.                 level = -level;

  421.             } else {

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

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

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

  425.                         level += 8;

  426.                     level >>= 4;

  427.                 } else {

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

  429.                         level += 32;

  430.                     level >>= 6;

  431.                 }

  432.             }

  433.             block[j] = level;

  434.         }

  435.     }

  436. }

  437. 

  438. static av_always_inline int dnxhd_ssd_block(int16_t *qblock, int16_t *block)

  439. {

  440.     int score = 0;

  441.     int i;

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

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

  444.     return score;

  445. }

  446. 

  447. static av_always_inline int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, int16_t *block, int last_index)

  448. {

  449.     int last_non_zero = 0;

  450.     int bits = 0;

  451.     int i, j, level;

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

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

  454.         level = block[j];

  455.         if (level) {

  456.             int run_level = i - last_non_zero - 1;

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

  458.             last_non_zero = i;

  459.         }

  460.     }

  461.     return bits;

  462. }

  463. 

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

  465. {

  466.     const int bs = ctx->block_width_l2;

  467.     const int bw = 1 << bs;

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

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

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

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

  472. 

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

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

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

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

  477. 

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

  479.         if (ctx->interlaced) {

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

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

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

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

  484.         } else {

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

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

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

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

  489.         }

  490.     } else {

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

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

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

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

  495.     }

  496. }

  497. 

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

  499. {

  500.     if (i&2) {

  501.         ctx->m.q_intra_matrix16 = ctx->qmatrix_c16;

  502.         ctx->m.q_intra_matrix   = ctx->qmatrix_c;

  503.         return 1 + (i&1);

  504.     } else {

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

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

  507.         return 0;

  508.     }

  509. }

  510. 

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

  512. {

  513.     DNXHDEncContext *ctx = avctx->priv_data;

  514.     int mb_y = jobnr, mb_x;

  515.     int qscale = ctx->qscale;

  516.     LOCAL_ALIGNED_16(int16_t, block, [64]);

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

  518. 

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

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

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

  522. 

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

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

  525.         int ssd     = 0;

  526.         int ac_bits = 0;

  527.         int dc_bits = 0;

  528.         int i;

  529. 

  530.         dnxhd_get_blocks(ctx, mb_x, mb_y);

  531. 

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

  533.             int16_t *src_block = ctx->blocks[i];

  534.             int overflow, nbits, diff, last_index;

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

  536. 

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

  538.             last_index = ctx->m.dct_quantize(&ctx->m, block, i, qscale, &overflow);

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

  540. 

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

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

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

  544. 

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

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

  547. 

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

  549. 

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

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

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

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

  554.             }

  555.         }

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

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

  558.     }

  559.     return 0;

  560. }

  561. 

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

  563. {

  564.     DNXHDEncContext *ctx = avctx->priv_data;

  565.     int mb_y = jobnr, mb_x;

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

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

  568. 

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

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

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

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

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

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

  575.         int i;

  576. 

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

  578. 

  579.         dnxhd_get_blocks(ctx, mb_x, mb_y);

  580. 

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

  582.             int16_t *block = ctx->blocks[i];

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

  584.             int last_index = ctx->m.dct_quantize(&ctx->m, block, i,

  585.                                                  qscale, &overflow);

  586.             //START_TIMER;

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

  588.             //STOP_TIMER("encode_block");

  589.         }

  590.     }

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

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

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

  594.     return 0;

  595. }

  596. 

  597. static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx)

  598. {

  599.     int mb_y, mb_x;

  600.     int offset = 0;

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

  602.         int thread_size;

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

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

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

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

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

  608.         }

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

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

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

  612.         offset += thread_size;

  613.     }

  614. }

  615. 

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

  617. {

  618.     DNXHDEncContext *ctx = avctx->priv_data;

  619.     int mb_y = jobnr, mb_x, x, y;

  620.     int partial_last_row = (mb_y == ctx->m.mb_height - 1) &&

  621.                            ((avctx->height >> ctx->interlaced) & 0xF);

  622. 

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

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

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

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

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

  628.             int sum;

  629.             int varc;

  630. 

  631.             if (!partial_last_row && mb_x * 16 <= avctx->width - 16) {

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

  633.                 varc = ctx->m.dsp.pix_norm1(pix, ctx->m.linesize);

  634.             } else {

  635.                 int bw = FFMIN(avctx->width - 16 * mb_x, 16);

  636.                 int bh = FFMIN((avctx->height >> ctx->interlaced) - 16 * mb_y, 16);

  637.                 sum = varc = 0;

  638.                 for (y = 0; y < bh; y++) {

  639.                     for (x = 0; x < bw; x++) {

  640.                         uint8_t val = pix[x + y * ctx->m.linesize];

  641.                         sum  += val;

  642.                         varc += val * val;

  643.                     }

  644.                 }

  645.             }

  646.             varc = (varc - (((unsigned)sum * sum) >> 8) + 128) >> 8;

  647. 

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

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

  650.         }

  651.     } else { // 10-bit

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

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

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

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

  656.             int sum = 0;

  657.             int sqsum = 0;

  658.             int mean, sqmean;

  659.             int i, j;

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

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

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

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

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

  665.                     sum += sample;

  666.                     sqsum += sample * sample;

  667.                     // 2^10 * 2^10 * 16 * 16 = 2^28, which is less than INT_MAX

  668.                 }

  669.                 pix += linesize;

  670.             }

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

  672.             sqmean = sqsum >> 8;

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

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

  675.         }

  676.     }

  677.     return 0;

  678. }

  679. 

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

  681. {

  682.     int lambda, up_step, down_step;

  683.     int last_lower = INT_MAX, last_higher = 0;

  684.     int x, y, q;

  685. 

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

  687.         ctx->qscale = q;

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

  689.     }

  690.     up_step = down_step = 2<<LAMBDA_FRAC_BITS;

  691.     lambda = ctx->lambda;

  692. 

  693.     for (;;) {

  694.         int bits = 0;

  695.         int end = 0;

  696.         if (lambda == last_higher) {

  697.             lambda++;

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

  699.         }

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

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

  702.                 unsigned min = UINT_MAX;

  703.                 int qscale = 1;

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

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

  706.                     unsigned score = ctx->mb_rc[q][mb].bits*lambda+

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

  708.                     if (score < min) {

  709.                         min = score;

  710.                         qscale = q;

  711.                     }

  712.                 }

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

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

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

  716.             }

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

  718.             if (bits > ctx->frame_bits)

  719.                 break;

  720.         }

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

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

  723.         if (end) {

  724.             if (bits > ctx->frame_bits)

  725.                 return -1;

  726.             break;

  727.         }

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

  729.             last_lower = FFMIN(lambda, last_lower);

  730.             if (last_higher != 0)

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

  732.             else

  733.                 lambda -= down_step;

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

  735.             up_step = 1<<LAMBDA_FRAC_BITS;

  736.             lambda = FFMAX(1, lambda);

  737.             if (lambda == last_lower)

  738.                 break;

  739.         } else {

  740.             last_higher = FFMAX(lambda, last_higher);

  741.             if (last_lower != INT_MAX)

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

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

  744.                 return -1;

  745.             else

  746.                 lambda += up_step;

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

  748.             down_step = 1<<LAMBDA_FRAC_BITS;

  749.         }

  750.     }

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

  752.     ctx->lambda = lambda;

  753.     return 0;

  754. }

  755. 

  756. static int dnxhd_find_qscale(DNXHDEncContext *ctx)

  757. {

  758.     int bits = 0;

  759.     int up_step = 1;

  760.     int down_step = 1;

  761.     int last_higher = 0;

  762.     int last_lower = INT_MAX;

  763.     int qscale;

  764.     int x, y;

  765. 

  766.     qscale = ctx->qscale;

  767.     for (;;) {

  768.         bits = 0;

  769.         ctx->qscale = qscale;

  770.         // XXX avoid recalculating bits

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

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

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

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

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

  776.             if (bits > ctx->frame_bits)

  777.                 break;

  778.         }

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

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

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

  782.             if (qscale == 1)

  783.                 return 1;

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

  785.                 qscale = last_higher;

  786.                 break;

  787.             }

  788.             last_lower = FFMIN(qscale, last_lower);

  789.             if (last_higher != 0)

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

  791.             else

  792.                 qscale -= down_step++;

  793.             if (qscale < 1)

  794.                 qscale = 1;

  795.             up_step = 1;

  796.         } else {

  797.             if (last_lower == qscale + 1)

  798.                 break;

  799.             last_higher = FFMAX(qscale, last_higher);

  800.             if (last_lower != INT_MAX)

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

  802.             else

  803.                 qscale += up_step++;

  804.             down_step = 1;

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

  806.                 return -1;

  807.         }

  808.     }

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

  810.     ctx->qscale = qscale;

  811.     return 0;

  812. }

  813. 

  814. #define BUCKET_BITS 8

  815. #define RADIX_PASSES 4

  816. #define NBUCKETS (1 << BUCKET_BITS)

  817. 

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

  819. {

  820.     value >>= shift;

  821.     value &= NBUCKETS - 1;

  822.     return NBUCKETS - 1 - value;

  823. }

  824. 

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

  826. {

  827.     int i, j;

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

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

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

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

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

  833.             v >>= BUCKET_BITS;

  834.         }

  835.         assert(!v);

  836.     }

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

  838.         int offset = size;

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

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

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

  842.     }

  843. }

  844. 

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

  846. {

  847.     int shift = pass * BUCKET_BITS;

  848.     int i;

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

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

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

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

  853.     }

  854. }

  855. 

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

  857. {

  858.     int buckets[RADIX_PASSES][NBUCKETS];

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

  860.     radix_count(data, size, buckets);

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

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

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

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

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

  866.     }

  867.     av_free(tmp);

  868. }

  869. 

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

  871. {

  872.     int max_bits = 0;

  873.     int ret, x, y;

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

  875.         return -1;

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

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

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

  879.             int delta_bits;

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

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

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

  883.             if (!RC_VARIANCE) {

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

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

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

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

  888.                     : INT_MIN; //avoid increasing qscale

  889.             }

  890.         }

  891.         max_bits += 31; //worst padding

  892.     }

  893.     if (!ret) {

  894.         if (RC_VARIANCE)

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

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

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

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

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

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

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

  902.         }

  903.     }

  904.     return 0;

  905. }

  906. 

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

  908. {

  909.     int i;

  910. 

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

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

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

  914.     }

  915. 

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

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

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

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

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

  921.     }

  922. 

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

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

  925. }

  926. 

  927. static int dnxhd_encode_picture(AVCodecContext *avctx, AVPacket *pkt,

  928.                                 const AVFrame *frame, int *got_packet)

  929. {

  930.     DNXHDEncContext *ctx = avctx->priv_data;

  931.     int first_field = 1;

  932.     int offset, i, ret;

  933.     uint8_t *buf;

  934. 

  935.     if ((ret = ff_alloc_packet(pkt, ctx->cid_table->frame_size)) < 0) {

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

  937.         return ret;

  938.     }

  939.     buf = pkt->data;

  940. 

  941.     dnxhd_load_picture(ctx, frame);

  942. 

  943.  encode_coding_unit:

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

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

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

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

  948.     }

  949. 

  950.     dnxhd_write_header(avctx, buf);

  951. 

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

  953.         ret = dnxhd_encode_rdo(avctx, ctx);

  954.     else

  955.         ret = dnxhd_encode_fast(avctx, ctx);

  956.     if (ret < 0) {

  957.         av_log(avctx, AV_LOG_ERROR,

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

  959.         return -1;

  960.     }

  961. 

  962.     dnxhd_setup_threads_slices(ctx);

  963. 

  964.     offset = 0;

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

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

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

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

  969.     }

  970. 

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

  972. 

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

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

  975. 

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

  977. 

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

  979.         first_field     = 0;

  980.         ctx->cur_field ^= 1;

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

  982.         goto encode_coding_unit;

  983.     }

  984. 

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

  986. 

  987.     pkt->flags |= AV_PKT_FLAG_KEY;

  988.     *got_packet = 1;

  989.     return 0;

  990. }

  991. 

  992. static av_cold int dnxhd_encode_end(AVCodecContext *avctx)

  993. {

  994.     DNXHDEncContext *ctx = avctx->priv_data;

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

  996.     int i;

  997. 

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

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

  1000.     av_freep(&ctx->run_codes);

  1001.     av_freep(&ctx->run_bits);

  1002. 

  1003.     av_freep(&ctx->mb_bits);

  1004.     av_freep(&ctx->mb_qscale);

  1005.     av_freep(&ctx->mb_rc);

  1006.     av_freep(&ctx->mb_cmp);

  1007.     av_freep(&ctx->slice_size);

  1008.     av_freep(&ctx->slice_offs);

  1009. 

  1010.     av_freep(&ctx->qmatrix_c);

  1011.     av_freep(&ctx->qmatrix_l);

  1012.     av_freep(&ctx->qmatrix_c16);

  1013.     av_freep(&ctx->qmatrix_l16);

  1014. 

  1015.     for (i = 1; i < avctx->thread_count; i++)

  1016.         av_freep(&ctx->thread[i]);

  1017. 

  1018.     return 0;

  1019. }

  1020. 

  1021. AVCodec ff_dnxhd_encoder = {

  1022.     .name           = "dnxhd",

  1023.     .long_name      = NULL_IF_CONFIG_SMALL("VC3/DNxHD"),

  1024.     .type           = AVMEDIA_TYPE_VIDEO,

  1025.     .id             = AV_CODEC_ID_DNXHD,

  1026.     .priv_data_size = sizeof(DNXHDEncContext),

  1027.     .init           = dnxhd_encode_init,

  1028.     .encode2        = dnxhd_encode_picture,

  1029.     .close          = dnxhd_encode_end,

  1030.     .capabilities   = CODEC_CAP_SLICE_THREADS,

  1031.     .pix_fmts       = (const enum AVPixelFormat[]){ AV_PIX_FMT_YUV422P,

  1032.                                                   AV_PIX_FMT_YUV422P10,

  1033.                                                   AV_PIX_FMT_NONE },

  1034.     .priv_class     = &class,

  1035. };