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.
29556 Commits
32 Branches
555MB
Tree: 1073823984
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '1073823984'
${ noResults }
FFmpeg/libavcodec/dnxhdenc.c

878 lines
30KB
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 Libav.

  8.  *

  9.  * Libav 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.  * Libav 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 Libav; 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 "libavutil/opt.h"

  28. #include "avcodec.h"

  29. #include "dsputil.h"

  30. #include "mpegvideo.h"

  31. #include "mpegvideo_common.h"

  32. #include "dnxhdenc.h"

  33. 

  34. #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM

  35. 

  36. static const AVOption options[]={

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

  38. {NULL}

  39. };

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

  41. 

  42. #define LAMBDA_FRAC_BITS 10

  43. 

  44. static av_always_inline void dnxhd_get_pixels_8x4(DCTELEM *restrict block, const uint8_t *pixels, int line_size)

  45. {

  46.     int i;

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

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

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

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

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

  52.         pixels += line_size;

  53.         block += 8;

  54.     }

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

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

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

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

  59. }

  60. 

  61. static int dnxhd_init_vlc(DNXHDEncContext *ctx)

  62. {

  63.     int i, j, level, run;

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

  65. 

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

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

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

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

  70. 

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

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

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

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

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

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

  77. 

  78.             MASK_ABS(sign, alevel);

  79.             if (alevel > 64) {

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

  81.                 alevel -= offset<<6;

  82.             }

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

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

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

  86.                     (!run    || (ctx->cid_table->ac_run_flag  [j] && run))) {

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

  88.                     if (alevel) {

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

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

  91.                     } else {

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

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

  94.                     }

  95.                     break;

  96.                 }

  97.             }

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

  99.             if (offset) {

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

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

  102.             }

  103.         }

  104.     }

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

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

  107.         assert(run < 63);

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

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

  110.     }

  111.     return 0;

  112.  fail:

  113.     return -1;

  114. }

  115. 

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

  117. {

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

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

  120.     int qscale, i;

  121. 

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

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

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

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

  126. 

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

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

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

  130.     }

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

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

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

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

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

  136.     }

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

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

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

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

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

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

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

  144.         }

  145.     }

  146.     return 0;

  147.  fail:

  148.     return -1;

  149. }

  150. 

  151. static int dnxhd_init_rc(DNXHDEncContext *ctx)

  152. {

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

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

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

  156. 

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

  158.     ctx->qscale = 1;

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

  160.     return 0;

  161.  fail:

  162.     return -1;

  163. }

  164. 

  165. static int dnxhd_encode_init(AVCodecContext *avctx)

  166. {

  167.     DNXHDEncContext *ctx = avctx->priv_data;

  168.     int i, index;

  169. 

  170.     ctx->cid = ff_dnxhd_find_cid(avctx);

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

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

  173.         return -1;

  174.     }

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

  176. 

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

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

  179. 

  180.     ctx->m.avctx = avctx;

  181.     ctx->m.mb_intra = 1;

  182.     ctx->m.h263_aic = 1;

  183. 

  184.     ctx->get_pixels_8x4_sym = dnxhd_get_pixels_8x4;

  185. 

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

  187.     ff_dct_common_init(&ctx->m);

  188. #if HAVE_MMX

  189.     ff_dnxhd_init_mmx(ctx);

  190. #endif

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

  192.         ctx->m.dct_quantize = dct_quantize_c;

  193. 

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

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

  196. 

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

  198.         ctx->interlaced = 1;

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

  200.     }

  201. 

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

  203. 

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

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

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

  207.         return -1;

  208. 

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

  210.     if (ctx->nitris_compat)

  211.         ctx->min_padding = 1600;

  212. 

  213.     if (dnxhd_init_vlc(ctx) < 0)

  214.         return -1;

  215.     if (dnxhd_init_rc(ctx) < 0)

  216.         return -1;

  217. 

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

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

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

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

  222. 

  223.     ctx->frame.key_frame = 1;

  224.     ctx->frame.pict_type = AV_PICTURE_TYPE_I;

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

  226. 

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

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

  229.         return -1;

  230.     }

  231. 

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

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

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

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

  236.     }

  237. 

  238.     return 0;

  239.  fail: //for FF_ALLOCZ_OR_GOTO

  240.     return -1;

  241. }

  242. 

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

  244. {

  245.     DNXHDEncContext *ctx = avctx->priv_data;

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

  247. 

  248.     memset(buf, 0, 640);

  249. 

  250.     memcpy(buf, header_prefix, 5);

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

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

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

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

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

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

  257. 

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

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

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

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

  262. 

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

  264. 

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

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

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

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

  269. 

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

  271.     return 0;

  272. }

  273. 

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

  275. {

  276.     int nbits;

  277.     if (diff < 0) {

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

  279.         diff--;

  280.     } else {

  281.         nbits = av_log2_16bit(2*diff);

  282.     }

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

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

  285. }

  286. 

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

  288. {

  289.     int last_non_zero = 0;

  290.     int slevel, i, j;

  291. 

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

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

  294. 

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

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

  297.         slevel = block[j];

  298.         if (slevel) {

  299.             int run_level = i - last_non_zero - 1;

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

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

  302.             if (run_level)

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

  304.             last_non_zero = i;

  305.         }

  306.     }

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

  308. }

  309. 

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

  311. {

  312.     const uint8_t *weight_matrix;

  313.     int level;

  314.     int i;

  315. 

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

  317. 

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

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

  320.         level = block[j];

  321.         if (level) {

  322.             if (level < 0) {

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

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

  325.                     level += 32;

  326.                 level >>= 6;

  327.                 level = -level;

  328.             } else {

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

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

  331.                     level += 32;

  332.                 level >>= 6;

  333.             }

  334.             block[j] = level;

  335.         }

  336.     }

  337. }

  338. 

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

  340. {

  341.     int score = 0;

  342.     int i;

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

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

  345.     return score;

  346. }

  347. 

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

  349. {

  350.     int last_non_zero = 0;

  351.     int bits = 0;

  352.     int i, j, level;

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

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

  355.         level = block[j];

  356.         if (level) {

  357.             int run_level = i - last_non_zero - 1;

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

  359.             last_non_zero = i;

  360.         }

  361.     }

  362.     return bits;

  363. }

  364. 

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

  366. {

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

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

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

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

  371. 

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

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

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

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

  376. 

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

  378.         if (ctx->interlaced) {

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

  380.             ctx->get_pixels_8x4_sym(ctx->blocks[5], ptr_y + ctx->dct_y_offset + 8, ctx->m.linesize);

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

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

  383.         } else {

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

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

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

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

  388.         }

  389.     } else {

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

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

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

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

  394.     }

  395. }

  396. 

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

  398. {

  399.     if (i&2) {

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

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

  402.         return 1 + (i&1);

  403.     } else {

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

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

  406.         return 0;

  407.     }

  408. }

  409. 

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

  411. {

  412.     DNXHDEncContext *ctx = avctx->priv_data;

  413.     int mb_y = jobnr, mb_x;

  414.     int qscale = ctx->qscale;

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

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

  417. 

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

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

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

  421. 

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

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

  424.         int ssd     = 0;

  425.         int ac_bits = 0;

  426.         int dc_bits = 0;

  427.         int i;

  428. 

  429.         dnxhd_get_blocks(ctx, mb_x, mb_y);

  430. 

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

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

  433.             int overflow, nbits, diff, last_index;

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

  435. 

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

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

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

  439. 

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

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

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

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

  444. 

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

  446. 

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

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

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

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

  451.             }

  452.         }

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

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

  455.     }

  456.     return 0;

  457. }

  458. 

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

  460. {

  461.     DNXHDEncContext *ctx = avctx->priv_data;

  462.     int mb_y = jobnr, mb_x;

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

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

  465. 

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

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

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

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

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

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

  472.         int i;

  473. 

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

  475. 

  476.         dnxhd_get_blocks(ctx, mb_x, mb_y);

  477. 

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

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

  480.             int last_index, overflow;

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

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

  483.             //START_TIMER;

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

  485.             //STOP_TIMER("encode_block");

  486.         }

  487.     }

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

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

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

  491.     return 0;

  492. }

  493. 

  494. static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx)

  495. {

  496.     int mb_y, mb_x;

  497.     int offset = 0;

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

  499.         int thread_size;

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

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

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

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

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

  505.         }

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

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

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

  509.         offset += thread_size;

  510.     }

  511. }

  512. 

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

  514. {

  515.     DNXHDEncContext *ctx = avctx->priv_data;

  516.     int mb_y = jobnr, mb_x;

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

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

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

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

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

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

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

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

  525.     }

  526.     return 0;

  527. }

  528. 

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

  530. {

  531.     int lambda, up_step, down_step;

  532.     int last_lower = INT_MAX, last_higher = 0;

  533.     int x, y, q;

  534. 

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

  536.         ctx->qscale = q;

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

  538.     }

  539.     up_step = down_step = 2<<LAMBDA_FRAC_BITS;

  540.     lambda = ctx->lambda;

  541. 

  542.     for (;;) {

  543.         int bits = 0;

  544.         int end = 0;

  545.         if (lambda == last_higher) {

  546.             lambda++;

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

  548.         }

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

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

  551.                 unsigned min = UINT_MAX;

  552.                 int qscale = 1;

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

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

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

  556.                     if (score < min) {

  557.                         min = score;

  558.                         qscale = q;

  559.                     }

  560.                 }

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

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

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

  564.             }

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

  566.             if (bits > ctx->frame_bits)

  567.                 break;

  568.         }

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

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

  571.         if (end) {

  572.             if (bits > ctx->frame_bits)

  573.                 return -1;

  574.             break;

  575.         }

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

  577.             last_lower = FFMIN(lambda, last_lower);

  578.             if (last_higher != 0)

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

  580.             else

  581.                 lambda -= down_step;

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

  583.             up_step = 1<<LAMBDA_FRAC_BITS;

  584.             lambda = FFMAX(1, lambda);

  585.             if (lambda == last_lower)

  586.                 break;

  587.         } else {

  588.             last_higher = FFMAX(lambda, last_higher);

  589.             if (last_lower != INT_MAX)

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

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

  592.                 return -1;

  593.             else

  594.                 lambda += up_step;

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

  596.             down_step = 1<<LAMBDA_FRAC_BITS;

  597.         }

  598.     }

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

  600.     ctx->lambda = lambda;

  601.     return 0;

  602. }

  603. 

  604. static int dnxhd_find_qscale(DNXHDEncContext *ctx)

  605. {

  606.     int bits = 0;

  607.     int up_step = 1;

  608.     int down_step = 1;

  609.     int last_higher = 0;

  610.     int last_lower = INT_MAX;

  611.     int qscale;

  612.     int x, y;

  613. 

  614.     qscale = ctx->qscale;

  615.     for (;;) {

  616.         bits = 0;

  617.         ctx->qscale = qscale;

  618.         // XXX avoid recalculating bits

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

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

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

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

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

  624.             if (bits > ctx->frame_bits)

  625.                 break;

  626.         }

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

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

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

  630.             if (qscale == 1)

  631.                 return 1;

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

  633.                 qscale = last_higher;

  634.                 break;

  635.             }

  636.             last_lower = FFMIN(qscale, last_lower);

  637.             if (last_higher != 0)

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

  639.             else

  640.                 qscale -= down_step++;

  641.             if (qscale < 1)

  642.                 qscale = 1;

  643.             up_step = 1;

  644.         } else {

  645.             if (last_lower == qscale + 1)

  646.                 break;

  647.             last_higher = FFMAX(qscale, last_higher);

  648.             if (last_lower != INT_MAX)

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

  650.             else

  651.                 qscale += up_step++;

  652.             down_step = 1;

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

  654.                 return -1;

  655.         }

  656.     }

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

  658.     ctx->qscale = qscale;

  659.     return 0;

  660. }

  661. 

  662. #define BUCKET_BITS 8

  663. #define RADIX_PASSES 4

  664. #define NBUCKETS (1 << BUCKET_BITS)

  665. 

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

  667. {

  668.     value >>= shift;

  669.     value &= NBUCKETS - 1;

  670.     return NBUCKETS - 1 - value;

  671. }

  672. 

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

  674. {

  675.     int i, j;

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

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

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

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

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

  681.             v >>= BUCKET_BITS;

  682.         }

  683.         assert(!v);

  684.     }

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

  686.         int offset = size;

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

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

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

  690.     }

  691. }

  692. 

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

  694. {

  695.     int shift = pass * BUCKET_BITS;

  696.     int i;

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

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

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

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

  701.     }

  702. }

  703. 

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

  705. {

  706.     int buckets[RADIX_PASSES][NBUCKETS];

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

  708.     radix_count(data, size, buckets);

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

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

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

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

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

  714.     }

  715.     av_free(tmp);

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

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

  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, const 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,

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

  804.         return -1;

  805.     }

  806. 

  807.     dnxhd_setup_threads_slices(ctx);

  808. 

  809.     offset = 0;

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

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

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

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

  814.     }

  815. 

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

  817. 

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

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

  820. 

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

  822. 

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

  824.         first_field     = 0;

  825.         ctx->cur_field ^= 1;

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

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

  828.         goto encode_coding_unit;

  829.     }

  830. 

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

  832. 

  833.     return ctx->cid_table->frame_size;

  834. }

  835. 

  836. static int dnxhd_encode_end(AVCodecContext *avctx)

  837. {

  838.     DNXHDEncContext *ctx = avctx->priv_data;

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

  840.     int i;

  841. 

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

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

  844.     av_freep(&ctx->run_codes);

  845.     av_freep(&ctx->run_bits);

  846. 

  847.     av_freep(&ctx->mb_bits);

  848.     av_freep(&ctx->mb_qscale);

  849.     av_freep(&ctx->mb_rc);

  850.     av_freep(&ctx->mb_cmp);

  851.     av_freep(&ctx->slice_size);

  852.     av_freep(&ctx->slice_offs);

  853. 

  854.     av_freep(&ctx->qmatrix_c);

  855.     av_freep(&ctx->qmatrix_l);

  856.     av_freep(&ctx->qmatrix_c16);

  857.     av_freep(&ctx->qmatrix_l16);

  858. 

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

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

  861. 

  862.     return 0;

  863. }

  864. 

  865. AVCodec ff_dnxhd_encoder = {

  866.     "dnxhd",

  867.     AVMEDIA_TYPE_VIDEO,

  868.     CODEC_ID_DNXHD,

  869.     sizeof(DNXHDEncContext),

  870.     dnxhd_encode_init,

  871.     dnxhd_encode_picture,

  872.     dnxhd_encode_end,

  873.     .capabilities = CODEC_CAP_SLICE_THREADS,

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

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

  876.     .priv_class = &class,

  877. };