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.
11648 Commits
32 Branches
498MB
Tree: bc2bc41bb8
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'bc2bc41bb8'
${ noResults }
FFmpeg/libavcodec/dnxhdenc.c

867 lines
29KB
Raw Blame History 

  1. /*

  2.  * VC3/DNxHD encoder

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

  4.  *

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

  6.  *

  7.  * This file is part of FFmpeg.

  8.  *

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

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

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

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

  13.  *

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

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

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

  17.  * Lesser General Public License for more details.

  18.  *

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

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

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

  22.  */

  23. 

  24. //#define DEBUG

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

  26. 

  27. #include "avcodec.h"

  28. #include "dsputil.h"

  29. #include "mpegvideo.h"

  30. #include "dnxhddata.h"

  31. 

  32. typedef struct {

  33.     uint16_t mb;

  34.     int value;

  35. } RCCMPEntry;

  36. 

  37. typedef struct {

  38.     int ssd;

  39.     int bits;

  40. } RCEntry;

  41. 

  42. int dct_quantize_c(MpegEncContext *s, DCTELEM *block, int n, int qscale, int *overflow);

  43. 

  44. typedef struct DNXHDEncContext {

  45.     MpegEncContext m; ///< Used for quantization dsp functions

  46. 

  47.     AVFrame frame;

  48.     int cid;

  49.     const CIDEntry *cid_table;

  50.     uint8_t *msip; ///< Macroblock Scan Indices Payload

  51.     uint32_t *slice_size;

  52. 

  53.     struct DNXHDEncContext *thread[MAX_THREADS];

  54. 

  55.     unsigned dct_y_offset;

  56.     unsigned dct_uv_offset;

  57.     int interlaced;

  58.     int cur_field;

  59. 

  60.     DECLARE_ALIGNED_16(DCTELEM, blocks[8][64]);

  61. 

  62.     int      (*qmatrix_c)     [64];

  63.     int      (*qmatrix_l)     [64];

  64.     uint16_t (*qmatrix_l16)[2][64];

  65.     uint16_t (*qmatrix_c16)[2][64];

  66. 

  67.     unsigned frame_bits;

  68.     uint8_t *src[3];

  69. 

  70.     uint16_t *table_vlc_codes;

  71.     uint8_t  *table_vlc_bits;

  72.     uint16_t *table_run_codes;

  73.     uint8_t  *table_run_bits;

  74. 

  75.     /** Rate control */

  76.     unsigned slice_bits;

  77.     unsigned qscale;

  78.     unsigned lambda;

  79. 

  80.     unsigned thread_size;

  81. 

  82.     uint16_t *mb_bits;

  83.     uint8_t  *mb_qscale;

  84. 

  85.     RCCMPEntry *mb_cmp;

  86.     RCEntry   (*mb_rc)[8160];

  87. } DNXHDEncContext;

  88. 

  89. #define LAMBDA_FRAC_BITS 10

  90. 

  91. static int dnxhd_init_vlc(DNXHDEncContext *ctx)

  92. {

  93.     int i;

  94. 

  95.     CHECKED_ALLOCZ(ctx->table_vlc_codes, 449*2);

  96.     CHECKED_ALLOCZ(ctx->table_vlc_bits,    449);

  97.     CHECKED_ALLOCZ(ctx->table_run_codes,  63*2);

  98.     CHECKED_ALLOCZ(ctx->table_run_bits,     63);

  99. 

  100.     for (i = 0; i < 257; i++) {

  101.         int level = ctx->cid_table->ac_level[i] +

  102.             (ctx->cid_table->ac_run_flag[i] << 7) + (ctx->cid_table->ac_index_flag[i] << 8);

  103.         assert(level < 449);

  104.         if (ctx->cid_table->ac_level[i] == 64 && ctx->cid_table->ac_index_flag[i])

  105.             level -= 64; // use 0+(1<<8) level

  106.         ctx->table_vlc_codes[level] = ctx->cid_table->ac_codes[i];

  107.         ctx->table_vlc_bits [level] = ctx->cid_table->ac_bits[i];

  108.     }

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

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

  111.         assert(run < 63);

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

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

  114.     }

  115.     return 0;

  116.  fail:

  117.     return -1;

  118. }

  119. 

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

  121. {

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

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

  124.     int qscale, i;

  125. 

  126.     CHECKED_ALLOCZ(ctx->qmatrix_l,   (ctx->m.avctx->qmax+1) * 64 * sizeof(int));

  127.     CHECKED_ALLOCZ(ctx->qmatrix_c,   (ctx->m.avctx->qmax+1) * 64 * sizeof(int));

  128.     CHECKED_ALLOCZ(ctx->qmatrix_l16, (ctx->m.avctx->qmax+1) * 64 * 2 * sizeof(uint16_t));

  129.     CHECKED_ALLOCZ(ctx->qmatrix_c16, (ctx->m.avctx->qmax+1) * 64 * 2 * sizeof(uint16_t));

  130. 

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

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

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

  134.     }

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

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

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

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

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

  140.     }

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

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

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

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

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

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

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

  148.         }

  149.     }

  150.     return 0;

  151.  fail:

  152.     return -1;

  153. }

  154. 

  155. static int dnxhd_init_rc(DNXHDEncContext *ctx)

  156. {

  157.     CHECKED_ALLOCZ(ctx->mb_rc, 8160*ctx->m.avctx->qmax*sizeof(RCEntry));

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

  159.         CHECKED_ALLOCZ(ctx->mb_cmp, ctx->m.mb_num*sizeof(RCCMPEntry));

  160. 

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

  162.     ctx->qscale = 1;

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

  164.     return 0;

  165.  fail:

  166.     return -1;

  167. }

  168. 

  169. static int dnxhd_encode_init(AVCodecContext *avctx)

  170. {

  171.     DNXHDEncContext *ctx = avctx->priv_data;

  172.     int i, index;

  173. 

  174.     if (avctx->width == 1920 && avctx->height == 1080) {

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

  176.             if      (avctx->bit_rate == 120000000)

  177.                 ctx->cid = 1242;

  178.             else if (avctx->bit_rate == 185000000)

  179.                 ctx->cid = 1243;

  180.         } else {

  181.             if      (avctx->bit_rate == 120000000)

  182.                 ctx->cid = 1237;

  183.             else if (avctx->bit_rate == 185000000)

  184.                 ctx->cid = 1238;

  185.             else if (avctx->bit_rate ==  36000000)

  186.                 ctx->cid = 1253;

  187.         }

  188.     } else if (avctx->width == 1280 && avctx->height == 720 &&

  189.                !(avctx->flags & CODEC_FLAG_INTERLACED_DCT)) {

  190.             if      (avctx->bit_rate ==  90000000)

  191.                 ctx->cid = 1251;

  192.             else if (avctx->bit_rate ==  60000000)

  193.                 ctx->cid = 1252;

  194.     }

  195.     if (!ctx->cid || avctx->pix_fmt != PIX_FMT_YUV422P) {

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

  197.         return -1;

  198.     }

  199. 

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

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

  202. 

  203.     ctx->m.avctx = avctx;

  204.     ctx->m.mb_intra = 1;

  205.     ctx->m.h263_aic = 1;

  206. 

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

  208.     ff_dct_common_init(&ctx->m);

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

  210.         ctx->m.dct_quantize = dct_quantize_c;

  211. 

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

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

  214. 

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

  216.         ctx->interlaced = 1;

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

  218.     }

  219. 

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

  221. 

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

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

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

  225.         return -1;

  226. 

  227.     if (dnxhd_init_vlc(ctx) < 0)

  228.         return -1;

  229.     if (dnxhd_init_rc(ctx) < 0)

  230.         return -1;

  231. 

  232.     CHECKED_ALLOCZ(ctx->slice_size, ctx->m.mb_height*sizeof(uint32_t));

  233.     CHECKED_ALLOCZ(ctx->mb_bits,    ctx->m.mb_num   *sizeof(uint16_t));

  234.     CHECKED_ALLOCZ(ctx->mb_qscale,  ctx->m.mb_num   *sizeof(uint8_t));

  235. 

  236.     ctx->frame.key_frame = 1;

  237.     ctx->frame.pict_type = FF_I_TYPE;

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

  239. 

  240.     if (avctx->thread_count > MAX_THREADS || (avctx->thread_count > ctx->m.mb_height)) {

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

  242.         return -1;

  243.     }

  244. 

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

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

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

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

  249.     }

  250. 

  251.     for (i = 0; i < avctx->thread_count; i++) {

  252.         ctx->thread[i]->m.start_mb_y = (ctx->m.mb_height*(i  ) + avctx->thread_count/2) / avctx->thread_count;

  253.         ctx->thread[i]->m.end_mb_y   = (ctx->m.mb_height*(i+1) + avctx->thread_count/2) / avctx->thread_count;

  254.     }

  255. 

  256.     return 0;

  257.  fail: //for CHECKED_ALLOCZ

  258.     return -1;

  259. }

  260. 

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

  262. {

  263.     DNXHDEncContext *ctx = avctx->priv_data;

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

  265. 

  266.     memcpy(buf, header_prefix, 5);

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

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

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

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

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

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

  273. 

  274.     buf[0x21] = 0x38; // FIXME 8 bit per comp

  275.     buf[0x22] = 0x88 + (ctx->frame.interlaced_frame<<2);

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

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

  278. 

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

  280. 

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

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

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

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

  285. 

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

  287.     return 0;

  288. }

  289. 

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

  291. {

  292.     int nbits;

  293.     if (diff < 0) {

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

  295.         diff--;

  296.     } else {

  297.         nbits = av_log2_16bit(2*diff);

  298.     }

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

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

  301. }

  302. 

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

  304. {

  305.     int last_non_zero = 0;

  306.     int offset = 0;

  307.     int slevel, i, j;

  308. 

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

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

  311. 

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

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

  314.         slevel = block[j];

  315.         if (slevel) {

  316.             int run_level = i - last_non_zero - 1;

  317.             int sign;

  318.             MASK_ABS(sign, slevel);

  319.             if (slevel > 64) {

  320.                 offset = (slevel-1) >> 6;

  321.                 slevel = 256 | (slevel & 63); // level 64 is treated as 0

  322.             }

  323.             if (run_level)

  324.                 slevel |= 128;

  325.             put_bits(&ctx->m.pb, ctx->table_vlc_bits[slevel]+1, (ctx->table_vlc_codes[slevel]<<1)|(sign&1));

  326.             if (offset) {

  327.                 put_bits(&ctx->m.pb, 4, offset);

  328.                 offset = 0;

  329.             }

  330.             if (run_level)

  331.                 put_bits(&ctx->m.pb, ctx->table_run_bits[run_level], ctx->table_run_codes[run_level]);

  332.             last_non_zero = i;

  333.         }

  334.     }

  335.     put_bits(&ctx->m.pb, ctx->table_vlc_bits[0], ctx->table_vlc_codes[0]); // EOB

  336. }

  337. 

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

  339. {

  340.     const uint8_t *weight_matrix;

  341.     int level;

  342.     int i;

  343. 

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

  345. 

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

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

  348.         level = block[j];

  349.         if (level) {

  350.             if (level < 0) {

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

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

  353.                     level += 32;

  354.                 level >>= 6;

  355.                 level = -level;

  356.             } else {

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

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

  359.                     level += 32;

  360.                 level >>= 6;

  361.             }

  362.             block[j] = level;

  363.         }

  364.     }

  365. }

  366. 

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

  368. {

  369.     int score = 0;

  370.     int i;

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

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

  373.     return score;

  374. }

  375. 

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

  377. {

  378.     int last_non_zero = 0;

  379.     int bits = 0;

  380.     int i, j, level;

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

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

  383.         level = block[j];

  384.         if (level) {

  385.             int run_level = i - last_non_zero - 1;

  386.             level = FFABS(level);

  387.             if (level > 64) {

  388.                 level = 256 | (level & 63); // level 64 is treated as 0

  389.                 bits += 4;

  390.             }

  391.             level |= (!!run_level)<<7;

  392.             bits += ctx->table_vlc_bits[level]+1 + ctx->table_run_bits[run_level];

  393.             last_non_zero = i;

  394.         }

  395.     }

  396.     return bits;

  397. }

  398. 

  399. static av_always_inline void dnxhd_get_pixels_4x8(DCTELEM *restrict block, const uint8_t *pixels, int line_size)

  400. {

  401.     int i;

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

  403.         block[0] = pixels[0];

  404.         block[1] = pixels[1];

  405.         block[2] = pixels[2];

  406.         block[3] = pixels[3];

  407.         block[4] = pixels[4];

  408.         block[5] = pixels[5];

  409.         block[6] = pixels[6];

  410.         block[7] = pixels[7];

  411.         pixels += line_size;

  412.         block += 8;

  413.     }

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

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

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

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

  418. }

  419. 

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

  421. {

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

  423.     const uint8_t *ptr_u = ctx->thread[0]->src[1] + ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << 3);

  424.     const uint8_t *ptr_v = ctx->thread[0]->src[2] + ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << 3);

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

  426. 

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

  428.     dsp->get_pixels(ctx->blocks[1], ptr_y + 8, ctx->m.linesize);

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

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

  431. 

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

  433.         if (ctx->interlaced) {

  434.             dnxhd_get_pixels_4x8(ctx->blocks[4], ptr_y + ctx->dct_y_offset    , ctx->m.linesize);

  435.             dnxhd_get_pixels_4x8(ctx->blocks[5], ptr_y + ctx->dct_y_offset + 8, ctx->m.linesize);

  436.             dnxhd_get_pixels_4x8(ctx->blocks[6], ptr_u + ctx->dct_uv_offset   , ctx->m.uvlinesize);

  437.             dnxhd_get_pixels_4x8(ctx->blocks[7], ptr_v + ctx->dct_uv_offset   , ctx->m.uvlinesize);

  438.         } else

  439.             memset(ctx->blocks[4], 0, 4*64*sizeof(DCTELEM));

  440.     } else {

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

  442.         dsp->get_pixels(ctx->blocks[5], ptr_y + ctx->dct_y_offset + 8, ctx->m.linesize);

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

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

  445.     }

  446. }

  447. 

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

  449. {

  450.     if (i&2) {

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

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

  453.         return 1 + (i&1);

  454.     } else {

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

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

  457.         return 0;

  458.     }

  459. }

  460. 

  461. static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg)

  462. {

  463.     DNXHDEncContext *ctx = arg;

  464.     int mb_y, mb_x;

  465.     int qscale = ctx->thread[0]->qscale;

  466. 

  467.     for (mb_y = ctx->m.start_mb_y; mb_y < ctx->m.end_mb_y; mb_y++) {

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

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

  470.         ctx->m.last_dc[2] = 1024;

  471. 

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

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

  474.             int ssd     = 0;

  475.             int ac_bits = 0;

  476.             int dc_bits = 0;

  477.             int i;

  478. 

  479.             dnxhd_get_blocks(ctx, mb_x, mb_y);

  480. 

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

  482.                 DECLARE_ALIGNED_16(DCTELEM, block[64]);

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

  484.                 int overflow, nbits, diff, last_index;

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

  486. 

  487.                 memcpy(block, src_block, sizeof(block));

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

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

  490. 

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

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

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

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

  495. 

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

  497. 

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

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

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

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

  502.                 }

  503.             }

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

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

  506.         }

  507.     }

  508.     return 0;

  509. }

  510. 

  511. static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg)

  512. {

  513.     DNXHDEncContext *ctx = arg;

  514.     int mb_y, mb_x;

  515. 

  516.     for (mb_y = ctx->m.start_mb_y; mb_y < ctx->m.end_mb_y; mb_y++) {

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

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

  519.         ctx->m.last_dc[2] = 1024;

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

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

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

  523.             int i;

  524. 

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

  526. 

  527.             dnxhd_get_blocks(ctx, mb_x, mb_y);

  528. 

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

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

  531.                 int last_index, overflow;

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

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

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

  535.             }

  536.         }

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

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

  539.     }

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

  541.     return 0;

  542. }

  543. 

  544. static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx, uint8_t *buf)

  545. {

  546.     int mb_y, mb_x;

  547.     int i, offset = 0;

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

  549.         int thread_size = 0;

  550.         for (mb_y = ctx->thread[i]->m.start_mb_y; mb_y < ctx->thread[i]->m.end_mb_y; mb_y++) {

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

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

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

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

  555.             }

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

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

  558.             thread_size += ctx->slice_size[mb_y];

  559.         }

  560.         init_put_bits(&ctx->thread[i]->m.pb, buf + 640 + offset, thread_size);

  561.         offset += thread_size;

  562.     }

  563. }

  564. 

  565. static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg)

  566. {

  567.     DNXHDEncContext *ctx = arg;

  568.     int mb_y, mb_x;

  569.     for (mb_y = ctx->m.start_mb_y; mb_y < ctx->m.end_mb_y; mb_y++) {

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

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

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

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

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

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

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

  577.         }

  578.     }

  579.     return 0;

  580. }

  581. 

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

  583. {

  584.     int lambda, up_step, down_step;

  585.     int last_lower = INT_MAX, last_higher = 0;

  586.     int x, y, q;

  587. 

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

  589.         ctx->qscale = q;

  590.         avctx->execute(avctx, dnxhd_calc_bits_thread, (void**)&ctx->thread[0], NULL, avctx->thread_count);

  591.     }

  592.     up_step = down_step = 2<<LAMBDA_FRAC_BITS;

  593.     lambda = ctx->lambda;

  594. 

  595.     for (;;) {

  596.         int bits = 0;

  597.         int end = 0;

  598.         if (lambda == last_higher) {

  599.             lambda++;

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

  601.         }

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

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

  604.                 unsigned min = UINT_MAX;

  605.                 int qscale = 1;

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

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

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

  609.                     if (score < min) {

  610.                         min = score;

  611.                         qscale = q;

  612.                     }

  613.                 }

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

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

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

  617.             }

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

  619.             if (bits > ctx->frame_bits)

  620.                 break;

  621.         }

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

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

  624.         if (end) {

  625.             if (bits > ctx->frame_bits)

  626.                 return -1;

  627.             break;

  628.         }

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

  630.             last_lower = FFMIN(lambda, last_lower);

  631.             if (last_higher != 0)

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

  633.             else

  634.                 lambda -= down_step;

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

  636.             up_step = 1<<LAMBDA_FRAC_BITS;

  637.             lambda = FFMAX(1, lambda);

  638.             if (lambda == last_lower)

  639.                 break;

  640.         } else {

  641.             last_higher = FFMAX(lambda, last_higher);

  642.             if (last_lower != INT_MAX)

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

  644.             else

  645.                 lambda += up_step;

  646.             up_step *= 5;

  647.             down_step = 1<<LAMBDA_FRAC_BITS;

  648.         }

  649.     }

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

  651.     ctx->lambda = lambda;

  652.     return 0;

  653. }

  654. 

  655. static int dnxhd_find_qscale(DNXHDEncContext *ctx)

  656. {

  657.     int bits = 0;

  658.     int up_step = 1;

  659.     int down_step = 1;

  660.     int last_higher = 0;

  661.     int last_lower = INT_MAX;

  662.     int qscale;

  663.     int x, y;

  664. 

  665.     qscale = ctx->qscale;

  666.     for (;;) {

  667.         bits = 0;

  668.         ctx->qscale = qscale;

  669.         // XXX avoid recalculating bits

  670.         ctx->m.avctx->execute(ctx->m.avctx, dnxhd_calc_bits_thread, (void**)&ctx->thread[0], NULL, ctx->m.avctx->thread_count);

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

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

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

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

  675.             if (bits > ctx->frame_bits)

  676.                 break;

  677.         }

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

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

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

  681.             if (qscale == 1)

  682.                 return 1;

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

  684.                 qscale = last_higher;

  685.                 break;

  686.             }

  687.             last_lower = FFMIN(qscale, last_lower);

  688.             if (last_higher != 0)

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

  690.             else

  691.                 qscale -= down_step++;

  692.             if (qscale < 1)

  693.                 qscale = 1;

  694.             up_step = 1;

  695.         } else {

  696.             if (last_lower == qscale + 1)

  697.                 break;

  698.             last_higher = FFMAX(qscale, last_higher);

  699.             if (last_lower != INT_MAX)

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

  701.             else

  702.                 qscale += up_step++;

  703.             down_step = 1;

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

  705.                 return -1;

  706.         }

  707.     }

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

  709.     ctx->qscale = qscale;

  710.     return 0;

  711. }

  712. 

  713. static int dnxhd_rc_cmp(const void *a, const void *b)

  714. {

  715.     return ((RCCMPEntry *)b)->value - ((RCCMPEntry *)a)->value;

  716. }

  717. 

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

  719. {

  720.     int max_bits = 0;

  721.     int ret, x, y;

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

  723.         return -1;

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

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

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

  727.             int delta_bits;

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

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

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

  731.             if (!RC_VARIANCE) {

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

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

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

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

  736.                     : INT_MIN; //avoid increasing qscale

  737.             }

  738.         }

  739.         max_bits += 31; //worst padding

  740.     }

  741.     if (!ret) {

  742.         if (RC_VARIANCE)

  743.             avctx->execute(avctx, dnxhd_mb_var_thread, (void**)&ctx->thread[0], NULL, avctx->thread_count);

  744.         qsort(ctx->mb_cmp, ctx->m.mb_num, sizeof(RCEntry), dnxhd_rc_cmp);

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

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

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

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

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

  750.         }

  751.     }

  752.     return 0;

  753. }

  754. 

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

  756. {

  757.     int i;

  758. 

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

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

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

  762.     }

  763. 

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

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

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

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

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

  769.     }

  770. 

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

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

  773. }

  774. 

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

  776. {

  777.     DNXHDEncContext *ctx = avctx->priv_data;

  778.     int first_field = 1;

  779.     int offset, i, ret;

  780. 

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

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

  783.         return -1;

  784.     }

  785. 

  786.     dnxhd_load_picture(ctx, data);

  787. 

  788.  encode_coding_unit:

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

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

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

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

  793.     }

  794. 

  795.     dnxhd_write_header(avctx, buf);

  796. 

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

  798.         ret = dnxhd_encode_rdo(avctx, ctx);

  799.     else

  800.         ret = dnxhd_encode_fast(avctx, ctx);

  801.     if (ret < 0) {

  802.         av_log(avctx, AV_LOG_ERROR, "picture could not fit ratecontrol constraints\n");

  803.         return -1;

  804.     }

  805. 

  806.     dnxhd_setup_threads_slices(ctx, buf);

  807. 

  808.     offset = 0;

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

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

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

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

  813.     }

  814. 

  815.     avctx->execute(avctx, dnxhd_encode_thread, (void**)&ctx->thread[0], NULL, avctx->thread_count);

  816. 

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

  818. 

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

  820.         first_field     = 0;

  821.         ctx->cur_field ^= 1;

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

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

  824.         goto encode_coding_unit;

  825.     }

  826. 

  827.     return ctx->cid_table->frame_size;

  828. }

  829. 

  830. static int dnxhd_encode_end(AVCodecContext *avctx)

  831. {

  832.     DNXHDEncContext *ctx = avctx->priv_data;

  833.     int i;

  834. 

  835.     av_freep(&ctx->table_vlc_codes);

  836.     av_freep(&ctx->table_vlc_bits);

  837.     av_freep(&ctx->table_run_codes);

  838.     av_freep(&ctx->table_run_bits);

  839. 

  840.     av_freep(&ctx->mb_bits);

  841.     av_freep(&ctx->mb_qscale);

  842.     av_freep(&ctx->mb_rc);

  843.     av_freep(&ctx->mb_cmp);

  844.     av_freep(&ctx->slice_size);

  845. 

  846.     av_freep(&ctx->qmatrix_c);

  847.     av_freep(&ctx->qmatrix_l);

  848.     av_freep(&ctx->qmatrix_c16);

  849.     av_freep(&ctx->qmatrix_l16);

  850. 

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

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

  853. 

  854.     return 0;

  855. }

  856. 

  857. AVCodec dnxhd_encoder = {

  858.     "dnxhd",

  859.     CODEC_TYPE_VIDEO,

  860.     CODEC_ID_DNXHD,

  861.     sizeof(DNXHDEncContext),

  862.     dnxhd_encode_init,

  863.     dnxhd_encode_picture,

  864.     dnxhd_encode_end,

  865.     .pix_fmts = (enum PixelFormat[]){PIX_FMT_YUV422P, -1},

  866. };