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.
43309 Commits
32 Branches
325MB
Tree: 7a72695c05
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '7a72695c05'
${ noResults }
FFmpeg/libavcodec/dnxhdenc.c

1024 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 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 "internal.h"

  33. #include "mpegvideo.h"

  34. #include "mpegvideo_common.h"

  35. #include "dnxhdenc.h"

  36. #include "internal.h"

  37. 

  38. #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM

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

  40. 

  41. static const AVOption options[]={

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

  43. {NULL}

  44. };

  45. 

  46. static const AVClass class = {

  47.     .class_name = "dnxhd",

  48.     .item_name  = av_default_item_name,

  49.     .option     = options,

  50.     .version    = LIBAVUTIL_VERSION_INT,

  51. };

  52. 

  53. #define LAMBDA_FRAC_BITS 10

  54. 

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

  56. {

  57.     int i;

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

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

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

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

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

  63.         pixels += line_size;

  64.         block += 8;

  65.     }

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

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

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

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

  70. }

  71. 

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

  73. {

  74.     int i;

  75. 

  76.     block += 32;

  77. 

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

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

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

  81.     }

  82. }

  83. 

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

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

  86. {

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

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

  89.     int last_non_zero = 0;

  90.     int i;

  91. 

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

  93. 

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

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

  96. 

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

  98.         int j = scantable[i];

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

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

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

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

  103.         if (level)

  104.             last_non_zero = i;

  105.     }

  106. 

  107.     return last_non_zero;

  108. }

  109. 

  110. static int dnxhd_init_vlc(DNXHDEncContext *ctx)

  111. {

  112.     int i, j, level, run;

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

  114. 

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

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

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

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

  119. 

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

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

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

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

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

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

  126. 

  127.             MASK_ABS(sign, alevel);

  128.             if (alevel > 64) {

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

  130.                 alevel -= offset<<6;

  131.             }

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

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

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

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

  136.                     av_assert1(!ctx->vlc_codes[index]);

  137.                     if (alevel) {

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

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

  140.                     } else {

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

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

  143.                     }

  144.                     break;

  145.                 }

  146.             }

  147.             av_assert0(!alevel || j < 257);

  148.             if (offset) {

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

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

  151.             }

  152.         }

  153.     }

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

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

  156.         av_assert0(run < 63);

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

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

  159.     }

  160.     return 0;

  161.  fail:

  162.     return -1;

  163. }

  164. 

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

  166. {

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

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

  169.     int qscale, i;

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

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

  172. 

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

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

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

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

  177. 

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

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

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

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

  182.         }

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

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

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

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

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

  188.         }

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

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

  191. 

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

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

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

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

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

  197.             }

  198.         }

  199.     } else {

  200.         // 10-bit

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

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

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

  204. 

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

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

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

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

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

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

  211.                 // ((1 << DNX10BIT_QMAT_SHIFT) * (p / s)) / (qscale * weight_table[i])

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

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

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

  215.             }

  216.         }

  217.     }

  218. 

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

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

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

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

  223. 

  224.     return 0;

  225.  fail:

  226.     return -1;

  227. }

  228. 

  229. static int dnxhd_init_rc(DNXHDEncContext *ctx)

  230. {

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

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

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

  234. 

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

  236.     ctx->qscale = 1;

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

  238.     return 0;

  239.  fail:

  240.     return -1;

  241. }

  242. 

  243. static int dnxhd_encode_init(AVCodecContext *avctx)

  244. {

  245.     DNXHDEncContext *ctx = avctx->priv_data;

  246.     int i, index, bit_depth;

  247. 

  248.     switch (avctx->pix_fmt) {

  249.     case PIX_FMT_YUV422P:

  250.         bit_depth = 8;

  251.         break;

  252.     case PIX_FMT_YUV422P10:

  253.         bit_depth = 10;

  254.         break;

  255.     default:

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

  257.         return -1;

  258.     }

  259. 

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

  261.     if (!ctx->cid) {

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

  263.         return -1;

  264.     }

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

  266. 

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

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

  269. 

  270.     ctx->m.avctx = avctx;

  271.     ctx->m.mb_intra = 1;

  272.     ctx->m.h263_aic = 1;

  273. 

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

  275. 

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

  277.     ff_dct_common_init(&ctx->m);

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

  279.         ctx->m.dct_quantize = ff_dct_quantize_c;

  280. 

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

  282.        ctx->m.dct_quantize = dnxhd_10bit_dct_quantize;

  283.        ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym;

  284.        ctx->block_width_l2 = 4;

  285.     } else {

  286.        ctx->get_pixels_8x4_sym = dnxhd_8bit_get_pixels_8x4_sym;

  287.        ctx->block_width_l2 = 3;

  288.     }

  289. 

  290. #if HAVE_MMX

  291.     ff_dnxhd_init_mmx(ctx);

  292. #endif

  293. 

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

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

  296. 

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

  298.         ctx->interlaced = 1;

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

  300.     }

  301. 

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

  303. 

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

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

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

  307.         return -1;

  308. 

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

  310.     if (ctx->nitris_compat)

  311.         ctx->min_padding = 1600;

  312. 

  313.     if (dnxhd_init_vlc(ctx) < 0)

  314.         return -1;

  315.     if (dnxhd_init_rc(ctx) < 0)

  316.         return -1;

  317. 

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

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

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

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

  322. 

  323.     ctx->frame.key_frame = 1;

  324.     ctx->frame.pict_type = AV_PICTURE_TYPE_I;

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

  326. 

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

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

  329.         return -1;

  330.     }

  331. 

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

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

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

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

  336.     }

  337. 

  338.     return 0;

  339.  fail: //for FF_ALLOCZ_OR_GOTO

  340.     return -1;

  341. }

  342. 

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

  344. {

  345.     DNXHDEncContext *ctx = avctx->priv_data;

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

  347. 

  348.     memset(buf, 0, 640);

  349. 

  350.     memcpy(buf, header_prefix, 5);

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

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

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

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

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

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

  357. 

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

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

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

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

  362. 

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

  364. 

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

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

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

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

  369. 

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

  371.     return 0;

  372. }

  373. 

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

  375. {

  376.     int nbits;

  377.     if (diff < 0) {

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

  379.         diff--;

  380.     } else {

  381.         nbits = av_log2_16bit(2*diff);

  382.     }

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

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

  385. }

  386. 

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

  388. {

  389.     int last_non_zero = 0;

  390.     int slevel, i, j;

  391. 

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

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

  394. 

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

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

  397.         slevel = block[j];

  398.         if (slevel) {

  399.             int run_level = i - last_non_zero - 1;

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

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

  402.             if (run_level)

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

  404.             last_non_zero = i;

  405.         }

  406.     }

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

  408. }

  409. 

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

  411. {

  412.     const uint8_t *weight_matrix;

  413.     int level;

  414.     int i;

  415. 

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

  417. 

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

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

  420.         level = block[j];

  421.         if (level) {

  422.             if (level < 0) {

  423.                 level = (1-2*level) * 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.                 level = -level;

  434.             } else {

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

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

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

  438.                         level += 8;

  439.                     level >>= 4;

  440.                 } else {

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

  442.                         level += 32;

  443.                     level >>= 6;

  444.                 }

  445.             }

  446.             block[j] = level;

  447.         }

  448.     }

  449. }

  450. 

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

  452. {

  453.     int score = 0;

  454.     int i;

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

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

  457.     return score;

  458. }

  459. 

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

  461. {

  462.     int last_non_zero = 0;

  463.     int bits = 0;

  464.     int i, j, level;

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

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

  467.         level = block[j];

  468.         if (level) {

  469.             int run_level = i - last_non_zero - 1;

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

  471.             last_non_zero = i;

  472.         }

  473.     }

  474.     return bits;

  475. }

  476. 

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

  478. {

  479.     const int bs = ctx->block_width_l2;

  480.     const int bw = 1 << bs;

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

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

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

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

  485. 

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

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

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

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

  490. 

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

  492.         if (ctx->interlaced) {

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

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

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

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

  497.         } else {

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

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

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

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

  502.         }

  503.     } else {

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

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

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

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

  508.     }

  509. }

  510. 

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

  512. {

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

  514.     return component[i];

  515. }

  516. 

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

  518. {

  519.     DNXHDEncContext *ctx = avctx->priv_data;

  520.     int mb_y = jobnr, mb_x;

  521.     int qscale = ctx->qscale;

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

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

  524. 

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

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

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

  528. 

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

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

  531.         int ssd     = 0;

  532.         int ac_bits = 0;

  533.         int dc_bits = 0;

  534.         int i;

  535. 

  536.         dnxhd_get_blocks(ctx, mb_x, mb_y);

  537. 

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

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

  540.             int overflow, nbits, diff, last_index;

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

  542. 

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

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

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

  546. 

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

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

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

  550. 

  551.             av_assert1(nbits < ctx->cid_table->bit_depth + 4);

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

  553. 

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

  555. 

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

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

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

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

  560.             }

  561.         }

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

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

  564.     }

  565.     return 0;

  566. }

  567. 

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

  569. {

  570.     DNXHDEncContext *ctx = avctx->priv_data;

  571.     int mb_y = jobnr, mb_x;

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

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

  574. 

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

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

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

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

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

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

  581.         int i;

  582. 

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

  584. 

  585.         dnxhd_get_blocks(ctx, mb_x, mb_y);

  586. 

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

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

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

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

  591.             //START_TIMER;

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

  593.             //STOP_TIMER("encode_block");

  594.         }

  595.     }

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

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

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

  599.     return 0;

  600. }

  601. 

  602. static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx)

  603. {

  604.     int mb_y, mb_x;

  605.     int offset = 0;

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

  607.         int thread_size;

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

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

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

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

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

  613.         }

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

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

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

  617.         offset += thread_size;

  618.     }

  619. }

  620. 

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

  622. {

  623.     DNXHDEncContext *ctx = avctx->priv_data;

  624.     int mb_y = jobnr, mb_x;

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

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

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

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

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

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

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

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

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

  634.         }

  635.     } else { // 10-bit

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

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

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

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

  640.             int sum = 0;

  641.             int sqsum = 0;

  642.             int mean, sqmean;

  643.             int i, j;

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

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

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

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

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

  649.                     sum += sample;

  650.                     sqsum += sample * sample;

  651.                     // 2^10 * 2^10 * 16 * 16 = 2^28, which is less than INT_MAX

  652.                 }

  653.                 pix += linesize;

  654.             }

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

  656.             sqmean = sqsum >> 8;

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

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

  659.         }

  660.     }

  661.     return 0;

  662. }

  663. 

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

  665. {

  666.     int lambda, up_step, down_step;

  667.     int last_lower = INT_MAX, last_higher = 0;

  668.     int x, y, q;

  669. 

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

  671.         ctx->qscale = q;

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

  673.     }

  674.     up_step = down_step = 2<<LAMBDA_FRAC_BITS;

  675.     lambda = ctx->lambda;

  676. 

  677.     for (;;) {

  678.         int bits = 0;

  679.         int end = 0;

  680.         if (lambda == last_higher) {

  681.             lambda++;

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

  683.         }

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

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

  686.                 unsigned min = UINT_MAX;

  687.                 int qscale = 1;

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

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

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

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

  692.                     if (score < min) {

  693.                         min = score;

  694.                         qscale = q;

  695.                     }

  696.                 }

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

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

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

  700.             }

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

  702.             if (bits > ctx->frame_bits)

  703.                 break;

  704.         }

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

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

  707.         if (end) {

  708.             if (bits > ctx->frame_bits)

  709.                 return -1;

  710.             break;

  711.         }

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

  713.             last_lower = FFMIN(lambda, last_lower);

  714.             if (last_higher != 0)

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

  716.             else

  717.                 lambda -= down_step;

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

  719.             up_step = 1<<LAMBDA_FRAC_BITS;

  720.             lambda = FFMAX(1, lambda);

  721.             if (lambda == last_lower)

  722.                 break;

  723.         } else {

  724.             last_higher = FFMAX(lambda, last_higher);

  725.             if (last_lower != INT_MAX)

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

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

  728.                 return -1;

  729.             else

  730.                 lambda += up_step;

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

  732.             down_step = 1<<LAMBDA_FRAC_BITS;

  733.         }

  734.     }

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

  736.     ctx->lambda = lambda;

  737.     return 0;

  738. }

  739. 

  740. static int dnxhd_find_qscale(DNXHDEncContext *ctx)

  741. {

  742.     int bits = 0;

  743.     int up_step = 1;

  744.     int down_step = 1;

  745.     int last_higher = 0;

  746.     int last_lower = INT_MAX;

  747.     int qscale;

  748.     int x, y;

  749. 

  750.     qscale = ctx->qscale;

  751.     for (;;) {

  752.         bits = 0;

  753.         ctx->qscale = qscale;

  754.         // XXX avoid recalculating bits

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

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

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

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

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

  760.             if (bits > ctx->frame_bits)

  761.                 break;

  762.         }

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

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

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

  766.             if (qscale == 1)

  767.                 return 1;

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

  769.                 qscale = last_higher;

  770.                 break;

  771.             }

  772.             last_lower = FFMIN(qscale, last_lower);

  773.             if (last_higher != 0)

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

  775.             else

  776.                 qscale -= down_step++;

  777.             if (qscale < 1)

  778.                 qscale = 1;

  779.             up_step = 1;

  780.         } else {

  781.             if (last_lower == qscale + 1)

  782.                 break;

  783.             last_higher = FFMAX(qscale, last_higher);

  784.             if (last_lower != INT_MAX)

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

  786.             else

  787.                 qscale += up_step++;

  788.             down_step = 1;

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

  790.                 return -1;

  791.         }

  792.     }

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

  794.     ctx->qscale = qscale;

  795.     return 0;

  796. }

  797. 

  798. #define BUCKET_BITS 8

  799. #define RADIX_PASSES 4

  800. #define NBUCKETS (1 << BUCKET_BITS)

  801. 

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

  803. {

  804.     value >>= shift;

  805.     value &= NBUCKETS - 1;

  806.     return NBUCKETS - 1 - value;

  807. }

  808. 

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

  810. {

  811.     int i, j;

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

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

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

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

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

  817.             v >>= BUCKET_BITS;

  818.         }

  819.         av_assert1(!v);

  820.     }

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

  822.         int offset = size;

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

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

  825.         av_assert1(!buckets[j][0]);

  826.     }

  827. }

  828. 

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

  830. {

  831.     int shift = pass * BUCKET_BITS;

  832.     int i;

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

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

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

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

  837.     }

  838. }

  839. 

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

  841. {

  842.     int buckets[RADIX_PASSES][NBUCKETS];

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

  844.     radix_count(data, size, buckets);

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

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

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

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

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

  850.     }

  851.     av_free(tmp);

  852. }

  853. 

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

  855. {

  856.     int max_bits = 0;

  857.     int ret, x, y;

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

  859.         return -1;

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

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

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

  863.             int delta_bits;

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

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

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

  867.             if (!RC_VARIANCE) {

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

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

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

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

  872.                     : INT_MIN; //avoid increasing qscale

  873.             }

  874.         }

  875.         max_bits += 31; //worst padding

  876.     }

  877.     if (!ret) {

  878.         if (RC_VARIANCE)

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

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

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

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

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

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

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

  886.         }

  887.     }

  888.     return 0;

  889. }

  890. 

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

  892. {

  893.     int i;

  894. 

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

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

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

  898.     }

  899. 

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

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

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

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

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

  905.     }

  906. 

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

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

  909. }

  910. 

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

  912.                                 const AVFrame *frame, int *got_packet)

  913. {

  914.     DNXHDEncContext *ctx = avctx->priv_data;

  915.     int first_field = 1;

  916.     int offset, i, ret;

  917.     uint8_t *buf;

  918. 

  919.     if ((ret = ff_alloc_packet2(avctx, pkt, ctx->cid_table->frame_size)) < 0)

  920.         return ret;

  921.     buf = pkt->data;

  922. 

  923.     dnxhd_load_picture(ctx, frame);

  924. 

  925.  encode_coding_unit:

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

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

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

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

  930.     }

  931. 

  932.     dnxhd_write_header(avctx, buf);

  933. 

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

  935.         ret = dnxhd_encode_rdo(avctx, ctx);

  936.     else

  937.         ret = dnxhd_encode_fast(avctx, ctx);

  938.     if (ret < 0) {

  939.         av_log(avctx, AV_LOG_ERROR,

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

  941.         return -1;

  942.     }

  943. 

  944.     dnxhd_setup_threads_slices(ctx);

  945. 

  946.     offset = 0;

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

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

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

  950.         av_assert1(!(ctx->slice_size[i] & 3));

  951.     }

  952. 

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

  954. 

  955.     av_assert1(640 + offset + 4 <= ctx->cid_table->coding_unit_size);

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

  957. 

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

  959. 

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

  961.         first_field     = 0;

  962.         ctx->cur_field ^= 1;

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

  964.         goto encode_coding_unit;

  965.     }

  966. 

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

  968. 

  969.     pkt->flags |= AV_PKT_FLAG_KEY;

  970.     *got_packet = 1;

  971.     return 0;

  972. }

  973. 

  974. static int dnxhd_encode_end(AVCodecContext *avctx)

  975. {

  976.     DNXHDEncContext *ctx = avctx->priv_data;

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

  978.     int i;

  979. 

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

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

  982.     av_freep(&ctx->run_codes);

  983.     av_freep(&ctx->run_bits);

  984. 

  985.     av_freep(&ctx->mb_bits);

  986.     av_freep(&ctx->mb_qscale);

  987.     av_freep(&ctx->mb_rc);

  988.     av_freep(&ctx->mb_cmp);

  989.     av_freep(&ctx->slice_size);

  990.     av_freep(&ctx->slice_offs);

  991. 

  992.     av_freep(&ctx->qmatrix_c);

  993.     av_freep(&ctx->qmatrix_l);

  994.     av_freep(&ctx->qmatrix_c16);

  995.     av_freep(&ctx->qmatrix_l16);

  996. 

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

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

  999. 

  1000.     return 0;

  1001. }

  1002. 

  1003. static const AVCodecDefault dnxhd_defaults[] = {

  1004.     { "qmax", "1024" }, /* Maximum quantization scale factor allowed for VC-3 */

  1005.     { NULL },

  1006. };

  1007. 

  1008. AVCodec ff_dnxhd_encoder = {

  1009.     .name           = "dnxhd",

  1010.     .type           = AVMEDIA_TYPE_VIDEO,

  1011.     .id             = AV_CODEC_ID_DNXHD,

  1012.     .priv_data_size = sizeof(DNXHDEncContext),

  1013.     .init           = dnxhd_encode_init,

  1014.     .encode2        = dnxhd_encode_picture,

  1015.     .close          = dnxhd_encode_end,

  1016.     .capabilities   = CODEC_CAP_SLICE_THREADS,

  1017.     .pix_fmts       = (const enum PixelFormat[]){ PIX_FMT_YUV422P,

  1018.                                                   PIX_FMT_YUV422P10,

  1019.                                                   PIX_FMT_NONE },

  1020.     .long_name      = NULL_IF_CONFIG_SMALL("VC3/DNxHD"),

  1021.     .priv_class     = &class,

  1022.     .defaults       = dnxhd_defaults,

  1023. };