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.
34063 Commits
32 Branches
940MB
Tree: 25b51b2c44
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '25b51b2c44'
${ noResults }
FFmpeg/libavcodec/dnxhdenc.c

1016 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 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 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 "internal.h"

  33. #include "mpegvideo.h"

  34. #include "mpegvideo_common.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, {.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 = 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 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 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 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 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 PIX_FMT_YUV422P:

  238.         bit_depth = 8;

  239.         break;

  240.     case 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 HAVE_MMX

  279.     ff_dnxhd_init_mmx(ctx);

  280. #endif

  281. 

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

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

  284. 

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

  286.         ctx->interlaced = 1;

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

  288.     }

  289. 

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

  291. 

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

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

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

  295.         return -1;

  296. 

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

  298.     if (ctx->nitris_compat)

  299.         ctx->min_padding = 1600;

  300. 

  301.     if (dnxhd_init_vlc(ctx) < 0)

  302.         return -1;

  303.     if (dnxhd_init_rc(ctx) < 0)

  304.         return -1;

  305. 

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

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

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

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

  310. 

  311.     ctx->frame.key_frame = 1;

  312.     ctx->frame.pict_type = AV_PICTURE_TYPE_I;

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

  314. 

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

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

  317.         return -1;

  318.     }

  319. 

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

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

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

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

  324.     }

  325. 

  326.     return 0;

  327.  fail: //for FF_ALLOCZ_OR_GOTO

  328.     return -1;

  329. }

  330. 

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

  332. {

  333.     DNXHDEncContext *ctx = avctx->priv_data;

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

  335. 

  336.     memset(buf, 0, 640);

  337. 

  338.     memcpy(buf, header_prefix, 5);

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

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

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

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

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

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

  345. 

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

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

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

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

  350. 

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

  352. 

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

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

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

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

  357. 

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

  359.     return 0;

  360. }

  361. 

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

  363. {

  364.     int nbits;

  365.     if (diff < 0) {

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

  367.         diff--;

  368.     } else {

  369.         nbits = av_log2_16bit(2*diff);

  370.     }

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

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

  373. }

  374. 

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

  376. {

  377.     int last_non_zero = 0;

  378.     int slevel, i, j;

  379. 

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

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

  382. 

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

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

  385.         slevel = block[j];

  386.         if (slevel) {

  387.             int run_level = i - last_non_zero - 1;

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

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

  390.             if (run_level)

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

  392.             last_non_zero = i;

  393.         }

  394.     }

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

  396. }

  397. 

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

  399. {

  400.     const uint8_t *weight_matrix;

  401.     int level;

  402.     int i;

  403. 

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

  405. 

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

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

  408.         level = block[j];

  409.         if (level) {

  410.             if (level < 0) {

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

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

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

  414.                         level += 8;

  415.                     level >>= 4;

  416.                 } else {

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

  418.                         level += 32;

  419.                     level >>= 6;

  420.                 }

  421.                 level = -level;

  422.             } else {

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

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

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

  426.                         level += 8;

  427.                     level >>= 4;

  428.                 } else {

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

  430.                         level += 32;

  431.                     level >>= 6;

  432.                 }

  433.             }

  434.             block[j] = level;

  435.         }

  436.     }

  437. }

  438. 

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

  440. {

  441.     int score = 0;

  442.     int i;

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

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

  445.     return score;

  446. }

  447. 

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

  449. {

  450.     int last_non_zero = 0;

  451.     int bits = 0;

  452.     int i, j, level;

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

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

  455.         level = block[j];

  456.         if (level) {

  457.             int run_level = i - last_non_zero - 1;

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

  459.             last_non_zero = i;

  460.         }

  461.     }

  462.     return bits;

  463. }

  464. 

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

  466. {

  467.     const int bs = ctx->block_width_l2;

  468.     const int bw = 1 << bs;

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

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

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

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

  473. 

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

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

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

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

  478. 

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

  480.         if (ctx->interlaced) {

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

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

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

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

  485.         } else {

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

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

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

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

  490.         }

  491.     } else {

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

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

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

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

  496.     }

  497. }

  498. 

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

  500. {

  501.     if (i&2) {

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

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

  504.         return 1 + (i&1);

  505.     } else {

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

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

  508.         return 0;

  509.     }

  510. }

  511. 

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

  513. {

  514.     DNXHDEncContext *ctx = avctx->priv_data;

  515.     int mb_y = jobnr, mb_x;

  516.     int qscale = ctx->qscale;

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

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

  519. 

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

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

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

  523. 

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

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

  526.         int ssd     = 0;

  527.         int ac_bits = 0;

  528.         int dc_bits = 0;

  529.         int i;

  530. 

  531.         dnxhd_get_blocks(ctx, mb_x, mb_y);

  532. 

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

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

  535.             int overflow, nbits, diff, last_index;

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

  537. 

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

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

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

  541. 

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

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

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

  545. 

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

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

  548. 

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

  550. 

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

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

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

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

  555.             }

  556.         }

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

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

  559.     }

  560.     return 0;

  561. }

  562. 

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

  564. {

  565.     DNXHDEncContext *ctx = avctx->priv_data;

  566.     int mb_y = jobnr, mb_x;

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

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

  569. 

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

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

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

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

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

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

  576.         int i;

  577. 

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

  579. 

  580.         dnxhd_get_blocks(ctx, mb_x, mb_y);

  581. 

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

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

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

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

  586.                                                  qscale, &overflow);

  587.             //START_TIMER;

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

  589.             //STOP_TIMER("encode_block");

  590.         }

  591.     }

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

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

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

  595.     return 0;

  596. }

  597. 

  598. static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx)

  599. {

  600.     int mb_y, mb_x;

  601.     int offset = 0;

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

  603.         int thread_size;

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

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

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

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

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

  609.         }

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

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

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

  613.         offset += thread_size;

  614.     }

  615. }

  616. 

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

  618. {

  619.     DNXHDEncContext *ctx = avctx->priv_data;

  620.     int mb_y = jobnr, mb_x;

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

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

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

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

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

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

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

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

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

  630.         }

  631.     } else { // 10-bit

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

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

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

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

  636.             int sum = 0;

  637.             int sqsum = 0;

  638.             int mean, sqmean;

  639.             int i, j;

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

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

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

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

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

  645.                     sum += sample;

  646.                     sqsum += sample * sample;

  647.                     // 2^10 * 2^10 * 16 * 16 = 2^28, which is less than INT_MAX

  648.                 }

  649.                 pix += linesize;

  650.             }

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

  652.             sqmean = sqsum >> 8;

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

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

  655.         }

  656.     }

  657.     return 0;

  658. }

  659. 

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

  661. {

  662.     int lambda, up_step, down_step;

  663.     int last_lower = INT_MAX, last_higher = 0;

  664.     int x, y, q;

  665. 

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

  667.         ctx->qscale = q;

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

  669.     }

  670.     up_step = down_step = 2<<LAMBDA_FRAC_BITS;

  671.     lambda = ctx->lambda;

  672. 

  673.     for (;;) {

  674.         int bits = 0;

  675.         int end = 0;

  676.         if (lambda == last_higher) {

  677.             lambda++;

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

  679.         }

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

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

  682.                 unsigned min = UINT_MAX;

  683.                 int qscale = 1;

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

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

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

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

  688.                     if (score < min) {

  689.                         min = score;

  690.                         qscale = q;

  691.                     }

  692.                 }

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

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

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

  696.             }

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

  698.             if (bits > ctx->frame_bits)

  699.                 break;

  700.         }

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

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

  703.         if (end) {

  704.             if (bits > ctx->frame_bits)

  705.                 return -1;

  706.             break;

  707.         }

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

  709.             last_lower = FFMIN(lambda, last_lower);

  710.             if (last_higher != 0)

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

  712.             else

  713.                 lambda -= down_step;

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

  715.             up_step = 1<<LAMBDA_FRAC_BITS;

  716.             lambda = FFMAX(1, lambda);

  717.             if (lambda == last_lower)

  718.                 break;

  719.         } else {

  720.             last_higher = FFMAX(lambda, last_higher);

  721.             if (last_lower != INT_MAX)

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

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

  724.                 return -1;

  725.             else

  726.                 lambda += up_step;

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

  728.             down_step = 1<<LAMBDA_FRAC_BITS;

  729.         }

  730.     }

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

  732.     ctx->lambda = lambda;

  733.     return 0;

  734. }

  735. 

  736. static int dnxhd_find_qscale(DNXHDEncContext *ctx)

  737. {

  738.     int bits = 0;

  739.     int up_step = 1;

  740.     int down_step = 1;

  741.     int last_higher = 0;

  742.     int last_lower = INT_MAX;

  743.     int qscale;

  744.     int x, y;

  745. 

  746.     qscale = ctx->qscale;

  747.     for (;;) {

  748.         bits = 0;

  749.         ctx->qscale = qscale;

  750.         // XXX avoid recalculating bits

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

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

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

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

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

  756.             if (bits > ctx->frame_bits)

  757.                 break;

  758.         }

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

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

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

  762.             if (qscale == 1)

  763.                 return 1;

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

  765.                 qscale = last_higher;

  766.                 break;

  767.             }

  768.             last_lower = FFMIN(qscale, last_lower);

  769.             if (last_higher != 0)

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

  771.             else

  772.                 qscale -= down_step++;

  773.             if (qscale < 1)

  774.                 qscale = 1;

  775.             up_step = 1;

  776.         } else {

  777.             if (last_lower == qscale + 1)

  778.                 break;

  779.             last_higher = FFMAX(qscale, last_higher);

  780.             if (last_lower != INT_MAX)

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

  782.             else

  783.                 qscale += up_step++;

  784.             down_step = 1;

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

  786.                 return -1;

  787.         }

  788.     }

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

  790.     ctx->qscale = qscale;

  791.     return 0;

  792. }

  793. 

  794. #define BUCKET_BITS 8

  795. #define RADIX_PASSES 4

  796. #define NBUCKETS (1 << BUCKET_BITS)

  797. 

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

  799. {

  800.     value >>= shift;

  801.     value &= NBUCKETS - 1;

  802.     return NBUCKETS - 1 - value;

  803. }

  804. 

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

  806. {

  807.     int i, j;

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

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

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

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

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

  813.             v >>= BUCKET_BITS;

  814.         }

  815.         assert(!v);

  816.     }

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

  818.         int offset = size;

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

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

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

  822.     }

  823. }

  824. 

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

  826. {

  827.     int shift = pass * BUCKET_BITS;

  828.     int i;

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

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

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

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

  833.     }

  834. }

  835. 

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

  837. {

  838.     int buckets[RADIX_PASSES][NBUCKETS];

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

  840.     radix_count(data, size, buckets);

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

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

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

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

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

  846.     }

  847.     av_free(tmp);

  848. }

  849. 

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

  851. {

  852.     int max_bits = 0;

  853.     int ret, x, y;

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

  855.         return -1;

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

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

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

  859.             int delta_bits;

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

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

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

  863.             if (!RC_VARIANCE) {

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

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

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

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

  868.                     : INT_MIN; //avoid increasing qscale

  869.             }

  870.         }

  871.         max_bits += 31; //worst padding

  872.     }

  873.     if (!ret) {

  874.         if (RC_VARIANCE)

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

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

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

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

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

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

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

  882.         }

  883.     }

  884.     return 0;

  885. }

  886. 

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

  888. {

  889.     int i;

  890. 

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

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

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

  894.     }

  895. 

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

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

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

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

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

  901.     }

  902. 

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

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

  905. }

  906. 

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

  908.                                 const AVFrame *frame, int *got_packet)

  909. {

  910.     DNXHDEncContext *ctx = avctx->priv_data;

  911.     int first_field = 1;

  912.     int offset, i, ret;

  913.     uint8_t *buf;

  914. 

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

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

  917.         return ret;

  918.     }

  919.     buf = pkt->data;

  920. 

  921.     dnxhd_load_picture(ctx, frame);

  922. 

  923.  encode_coding_unit:

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

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

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

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

  928.     }

  929. 

  930.     dnxhd_write_header(avctx, buf);

  931. 

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

  933.         ret = dnxhd_encode_rdo(avctx, ctx);

  934.     else

  935.         ret = dnxhd_encode_fast(avctx, ctx);

  936.     if (ret < 0) {

  937.         av_log(avctx, AV_LOG_ERROR,

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

  939.         return -1;

  940.     }

  941. 

  942.     dnxhd_setup_threads_slices(ctx);

  943. 

  944.     offset = 0;

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

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

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

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

  949.     }

  950. 

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

  952. 

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

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

  955. 

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

  957. 

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

  959.         first_field     = 0;

  960.         ctx->cur_field ^= 1;

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

  962.         goto encode_coding_unit;

  963.     }

  964. 

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

  966. 

  967.     pkt->flags |= AV_PKT_FLAG_KEY;

  968.     *got_packet = 1;

  969.     return 0;

  970. }

  971. 

  972. static int dnxhd_encode_end(AVCodecContext *avctx)

  973. {

  974.     DNXHDEncContext *ctx = avctx->priv_data;

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

  976.     int i;

  977. 

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

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

  980.     av_freep(&ctx->run_codes);

  981.     av_freep(&ctx->run_bits);

  982. 

  983.     av_freep(&ctx->mb_bits);

  984.     av_freep(&ctx->mb_qscale);

  985.     av_freep(&ctx->mb_rc);

  986.     av_freep(&ctx->mb_cmp);

  987.     av_freep(&ctx->slice_size);

  988.     av_freep(&ctx->slice_offs);

  989. 

  990.     av_freep(&ctx->qmatrix_c);

  991.     av_freep(&ctx->qmatrix_l);

  992.     av_freep(&ctx->qmatrix_c16);

  993.     av_freep(&ctx->qmatrix_l16);

  994. 

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

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

  997. 

  998.     return 0;

  999. }

  1000. 

  1001. AVCodec ff_dnxhd_encoder = {

  1002.     .name           = "dnxhd",

  1003.     .type           = AVMEDIA_TYPE_VIDEO,

  1004.     .id             = CODEC_ID_DNXHD,

  1005.     .priv_data_size = sizeof(DNXHDEncContext),

  1006.     .init           = dnxhd_encode_init,

  1007.     .encode2        = dnxhd_encode_picture,

  1008.     .close          = dnxhd_encode_end,

  1009.     .capabilities   = CODEC_CAP_SLICE_THREADS,

  1010.     .pix_fmts       = (const enum PixelFormat[]){ PIX_FMT_YUV422P,

  1011.                                                   PIX_FMT_YUV422P10,

  1012.                                                   PIX_FMT_NONE },

  1013.     .long_name      = NULL_IF_CONFIG_SMALL("VC3/DNxHD"),

  1014.     .priv_class     = &class,

  1015. };