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.
80498 Commits
32 Branches
424MB
Tree: 8ef57a0d61
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '8ef57a0d61'
${ noResults }
FFmpeg/libavcodec/dnxhdenc.c

1174 lines
40KB
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. #include "libavutil/attributes.h"

  27. #include "libavutil/internal.h"

  28. #include "libavutil/opt.h"

  29. #include "libavutil/timer.h"

  30. 

  31. #include "avcodec.h"

  32. #include "blockdsp.h"

  33. #include "fdctdsp.h"

  34. #include "internal.h"

  35. #include "mpegvideo.h"

  36. #include "pixblockdsp.h"

  37. #include "dnxhdenc.h"

  38. 

  39. // The largest value that will not lead to overflow for 10-bit samples.

  40. #define DNX10BIT_QMAT_SHIFT 18

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

  42. #define LAMBDA_FRAC_BITS 10

  43. 

  44. #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM

  45. static const AVOption options[] = {

  46.     { "nitris_compat", "encode with Avid Nitris compatibility",

  47.         offsetof(DNXHDEncContext, nitris_compat), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, VE },

  48.     { "ibias", "intra quant bias",

  49.         offsetof(DNXHDEncContext, intra_quant_bias), AV_OPT_TYPE_INT,

  50.         { .i64 = 0 }, INT_MIN, INT_MAX, VE },

  51.     { NULL }

  52. };

  53. 

  54. static const AVClass dnxhd_class = {

  55.     .class_name = "dnxhd",

  56.     .item_name  = av_default_item_name,

  57.     .option     = options,

  58.     .version    = LIBAVUTIL_VERSION_INT,

  59. };

  60. 

  61. static void dnxhd_8bit_get_pixels_8x4_sym(int16_t *av_restrict block,

  62.                                           const uint8_t *pixels,

  63.                                           ptrdiff_t line_size)

  64. {

  65.     int i;

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

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

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

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

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

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

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

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

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

  75.         pixels  += line_size;

  76.         block   += 8;

  77.     }

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

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

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

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

  82. }

  83. 

  84. static av_always_inline

  85. void dnxhd_10bit_get_pixels_8x4_sym(int16_t *av_restrict block,

  86.                                     const uint8_t *pixels,

  87.                                     ptrdiff_t line_size)

  88. {

  89.     memcpy(block + 0 * 8, pixels + 0 * line_size, 8 * sizeof(*block));

  90.     memcpy(block + 7 * 8, pixels + 0 * line_size, 8 * sizeof(*block));

  91.     memcpy(block + 1 * 8, pixels + 1 * line_size, 8 * sizeof(*block));

  92.     memcpy(block + 6 * 8, pixels + 1 * line_size, 8 * sizeof(*block));

  93.     memcpy(block + 2 * 8, pixels + 2 * line_size, 8 * sizeof(*block));

  94.     memcpy(block + 5 * 8, pixels + 2 * line_size, 8 * sizeof(*block));

  95.     memcpy(block + 3 * 8, pixels + 3 * line_size, 8 * sizeof(*block));

  96.     memcpy(block + 4 * 8, pixels + 3 * line_size, 8 * sizeof(*block));

  97. }

  98. 

  99. static int dnxhd_10bit_dct_quantize(MpegEncContext *ctx, int16_t *block,

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

  101. {

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

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

  104.     int last_non_zero = 0;

  105.     int i;

  106. 

  107.     ctx->fdsp.fdct(block);

  108. 

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

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

  111. 

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

  113.         int j = scantable[i];

  114.         int sign = FF_SIGNBIT(block[j]);

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

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

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

  118.         if (level)

  119.             last_non_zero = i;

  120.     }

  121. 

  122.     /* we need this permutation so that we correct the IDCT, we only permute the !=0 elements */

  123.     if (ctx->idsp.perm_type != FF_IDCT_PERM_NONE)

  124.         ff_block_permute(block, ctx->idsp.idct_permutation,

  125.                          scantable, last_non_zero);

  126. 

  127.     return last_non_zero;

  128. }

  129. 

  130. static av_cold int dnxhd_init_vlc(DNXHDEncContext *ctx)

  131. {

  132.     int i, j, level, run;

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

  134. 

  135.     FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->vlc_codes,

  136.                       max_level, 4 * sizeof(*ctx->vlc_codes), fail);

  137.     FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->vlc_bits,

  138.                       max_level, 4 * sizeof(*ctx->vlc_bits), fail);

  139.     FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_codes,

  140.                       63 * 2, fail);

  141.     FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_bits,

  142.                       63, fail);

  143. 

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

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

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

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

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

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

  150. 

  151.             MASK_ABS(sign, alevel);

  152.             if (alevel > 64) {

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

  154.                 alevel -= offset << 6;

  155.             }

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

  157.                 if (ctx->cid_table->ac_info[2*j+0] >> 1 == alevel &&

  158.                     (!offset || (ctx->cid_table->ac_info[2*j+1] & 1) && offset) &&

  159.                     (!run    || (ctx->cid_table->ac_info[2*j+1] & 2) && run)) {

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

  161.                     if (alevel) {

  162.                         ctx->vlc_codes[index] =

  163.                             (ctx->cid_table->ac_codes[j] << 1) | (sign & 1);

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

  165.                     } else {

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

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

  168.                     }

  169.                     break;

  170.                 }

  171.             }

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

  173.             if (offset) {

  174.                 ctx->vlc_codes[index] =

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

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

  177.             }

  178.         }

  179.     }

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

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

  182.         av_assert0(run < 63);

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

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

  185.     }

  186.     return 0;

  187. fail:

  188.     return AVERROR(ENOMEM);

  189. }

  190. 

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

  192. {

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

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

  195.     int qscale, i;

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

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

  198. 

  199.     FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l,

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

  201.     FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c,

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

  203.     FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l16,

  204.                       (ctx->m.avctx->qmax + 1), 64 * 2 * sizeof(uint16_t),

  205.                       fail);

  206.     FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c16,

  207.                       (ctx->m.avctx->qmax + 1), 64 * 2 * sizeof(uint16_t),

  208.                       fail);

  209. 

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

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

  212.             int j = ctx->m.idsp.idct_permutation[ff_zigzag_direct[i]];

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

  214.         }

  215.         ff_convert_matrix(&ctx->m, ctx->qmatrix_l, ctx->qmatrix_l16,

  216.                           weight_matrix, ctx->intra_quant_bias, 1,

  217.                           ctx->m.avctx->qmax, 1);

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

  219.             int j = ctx->m.idsp.idct_permutation[ff_zigzag_direct[i]];

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

  221.         }

  222.         ff_convert_matrix(&ctx->m, ctx->qmatrix_c, ctx->qmatrix_c16,

  223.                           weight_matrix, ctx->intra_quant_bias, 1,

  224.                           ctx->m.avctx->qmax, 1);

  225. 

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

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

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

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

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

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

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

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

  234.             }

  235.         }

  236.     } else {

  237.         // 10-bit

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

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

  240.                 int j = ff_zigzag_direct[i];

  241. 

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

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

  244.                  *             (qscale * weight_table[i]))

  245.                  * Where p is 32 for 8-bit samples and 8 for 10-bit ones.

  246.                  * The s factor compensates scaling of DCT coefficients done by

  247.                  * the DCT routines, and therefore is not present in standard.

  248.                  * It's 8 for 8-bit samples and 4 for 10-bit ones.

  249.                  * We want values of ctx->qtmatrix_l and ctx->qtmatrix_r to be:

  250.                  *     ((1 << DNX10BIT_QMAT_SHIFT) * (p / s)) /

  251.                  *     (qscale * weight_table[i])

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

  253.                 ctx->qmatrix_l[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) /

  254.                                             (qscale * luma_weight_table[i]);

  255.                 ctx->qmatrix_c[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) /

  256.                                             (qscale * chroma_weight_table[i]);

  257.             }

  258.         }

  259.     }

  260. 

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

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

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

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

  265. 

  266.     return 0;

  267. fail:

  268.     return AVERROR(ENOMEM);

  269. }

  270. 

  271. static av_cold int dnxhd_init_rc(DNXHDEncContext *ctx)

  272. {

  273.     FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->mb_rc, (ctx->m.avctx->qmax + 1), 8160 * sizeof(RCEntry), fail);

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

  275.         FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->mb_cmp,

  276.                           ctx->m.mb_num, sizeof(RCCMPEntry), fail);

  277. 

  278.     ctx->frame_bits = (ctx->cid_table->coding_unit_size -

  279.                        640 - 4 - ctx->min_padding) * 8;

  280.     ctx->qscale = 1;

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

  282.     return 0;

  283. fail:

  284.     return AVERROR(ENOMEM);

  285. }

  286. 

  287. static av_cold int dnxhd_encode_init(AVCodecContext *avctx)

  288. {

  289.     DNXHDEncContext *ctx = avctx->priv_data;

  290.     int i, index, bit_depth, ret;

  291. 

  292.     switch (avctx->pix_fmt) {

  293.     case AV_PIX_FMT_YUV422P:

  294.         bit_depth = 8;

  295.         break;

  296.     case AV_PIX_FMT_YUV422P10:

  297.         bit_depth = 10;

  298.         break;

  299.     default:

  300.         av_log(avctx, AV_LOG_ERROR,

  301.                "pixel format is incompatible with DNxHD\n");

  302.         return AVERROR(EINVAL);

  303.     }

  304. 

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

  306.     if (!ctx->cid) {

  307.         av_log(avctx, AV_LOG_ERROR,

  308.                "video parameters incompatible with DNxHD. Valid DNxHD profiles:\n");

  309.         ff_dnxhd_print_profiles(avctx, AV_LOG_ERROR);

  310.         return AVERROR(EINVAL);

  311.     }

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

  313. 

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

  315.     av_assert0(index >= 0);

  316. 

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

  318. 

  319.     ctx->m.avctx    = avctx;

  320.     ctx->m.mb_intra = 1;

  321.     ctx->m.h263_aic = 1;

  322. 

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

  324. 

  325.     ff_blockdsp_init(&ctx->bdsp, avctx);

  326.     ff_fdctdsp_init(&ctx->m.fdsp, avctx);

  327.     ff_mpv_idct_init(&ctx->m);

  328.     ff_mpegvideoencdsp_init(&ctx->m.mpvencdsp, avctx);

  329.     ff_pixblockdsp_init(&ctx->m.pdsp, avctx);

  330.     ff_dct_encode_init(&ctx->m);

  331. 

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

  333.         ctx->m.dct_quantize = ff_dct_quantize_c;

  334. 

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

  336.         ctx->m.dct_quantize     = dnxhd_10bit_dct_quantize;

  337.         ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym;

  338.         ctx->block_width_l2     = 4;

  339.     } else {

  340.         ctx->get_pixels_8x4_sym = dnxhd_8bit_get_pixels_8x4_sym;

  341.         ctx->block_width_l2     = 3;

  342.     }

  343. 

  344.     if (ARCH_X86)

  345.         ff_dnxhdenc_init_x86(ctx);

  346. 

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

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

  349. 

  350.     if (avctx->flags & AV_CODEC_FLAG_INTERLACED_DCT) {

  351.         ctx->interlaced   = 1;

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

  353.     }

  354. 

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

  356. 

  357. #if FF_API_QUANT_BIAS

  358. FF_DISABLE_DEPRECATION_WARNINGS

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

  360.         ctx->intra_quant_bias = avctx->intra_quant_bias;

  361. FF_ENABLE_DEPRECATION_WARNINGS

  362. #endif

  363.     // XXX tune lbias/cbias

  364.     if ((ret = dnxhd_init_qmat(ctx, ctx->intra_quant_bias, 0)) < 0)

  365.         return ret;

  366. 

  367.     /* Avid Nitris hardware decoder requires a minimum amount of padding

  368.      * in the coding unit payload */

  369.     if (ctx->nitris_compat)

  370.         ctx->min_padding = 1600;

  371. 

  372.     if ((ret = dnxhd_init_vlc(ctx)) < 0)

  373.         return ret;

  374.     if ((ret = dnxhd_init_rc(ctx)) < 0)

  375.         return ret;

  376. 

  377.     FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_size,

  378.                       ctx->m.mb_height * sizeof(uint32_t), fail);

  379.     FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_offs,

  380.                       ctx->m.mb_height * sizeof(uint32_t), fail);

  381.     FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_bits,

  382.                       ctx->m.mb_num * sizeof(uint16_t), fail);

  383.     FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_qscale,

  384.                       ctx->m.mb_num * sizeof(uint8_t), fail);

  385. 

  386. #if FF_API_CODED_FRAME

  387. FF_DISABLE_DEPRECATION_WARNINGS

  388.     avctx->coded_frame->key_frame = 1;

  389.     avctx->coded_frame->pict_type = AV_PICTURE_TYPE_I;

  390. FF_ENABLE_DEPRECATION_WARNINGS

  391. #endif

  392. 

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

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

  395.         return AVERROR(EINVAL);

  396.     }

  397. 

  398.     if (avctx->qmax <= 1) {

  399.         av_log(avctx, AV_LOG_ERROR, "qmax must be at least 2\n");

  400.         return AVERROR(EINVAL);

  401.     }

  402. 

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

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

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

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

  407.     }

  408. 

  409.     return 0;

  410. fail:  // for FF_ALLOCZ_OR_GOTO

  411.     return AVERROR(ENOMEM);

  412. }

  413. 

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

  415. {

  416.     DNXHDEncContext *ctx = avctx->priv_data;

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

  418. 

  419.     memset(buf, 0, 640);

  420. 

  421.     memcpy(buf, header_prefix, 5);

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

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

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

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

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

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

  428. 

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

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

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

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

  433. 

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

  435. 

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

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

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

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

  440. 

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

  442.     return 0;

  443. }

  444. 

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

  446. {

  447.     int nbits;

  448.     if (diff < 0) {

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

  450.         diff--;

  451.     } else {

  452.         nbits = av_log2_16bit(2 * diff);

  453.     }

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

  455.              (ctx->cid_table->dc_codes[nbits] << nbits) +

  456.              av_mod_uintp2(diff, nbits));

  457. }

  458. 

  459. static av_always_inline

  460. void dnxhd_encode_block(DNXHDEncContext *ctx, int16_t *block,

  461.                         int last_index, int n)

  462. {

  463.     int last_non_zero = 0;

  464.     int slevel, i, j;

  465. 

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

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

  468. 

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

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

  471.         slevel = block[j];

  472.         if (slevel) {

  473.             int run_level = i - last_non_zero - 1;

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

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

  476.             if (run_level)

  477.                 put_bits(&ctx->m.pb, ctx->run_bits[run_level],

  478.                          ctx->run_codes[run_level]);

  479.             last_non_zero = i;

  480.         }

  481.     }

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

  483. }

  484. 

  485. static av_always_inline

  486. void dnxhd_unquantize_c(DNXHDEncContext *ctx, int16_t *block, int n,

  487.                         int qscale, int last_index)

  488. {

  489.     const uint8_t *weight_matrix;

  490.     int level;

  491.     int i;

  492. 

  493.     weight_matrix = (n & 2) ? ctx->cid_table->chroma_weight

  494.                             : ctx->cid_table->luma_weight;

  495. 

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

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

  498.         level = block[j];

  499.         if (level) {

  500.             if (level < 0) {

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

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

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

  504.                         level += 8;

  505.                     level >>= 4;

  506.                 } else {

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

  508.                         level += 32;

  509.                     level >>= 6;

  510.                 }

  511.                 level = -level;

  512.             } else {

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

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

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

  516.                         level += 8;

  517.                     level >>= 4;

  518.                 } else {

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

  520.                         level += 32;

  521.                     level >>= 6;

  522.                 }

  523.             }

  524.             block[j] = level;

  525.         }

  526.     }

  527. }

  528. 

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

  530. {

  531.     int score = 0;

  532.     int i;

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

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

  535.     return score;

  536. }

  537. 

  538. static av_always_inline

  539. int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, int16_t *block, int last_index)

  540. {

  541.     int last_non_zero = 0;

  542.     int bits = 0;

  543.     int i, j, level;

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

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

  546.         level = block[j];

  547.         if (level) {

  548.             int run_level = i - last_non_zero - 1;

  549.             bits += ctx->vlc_bits[(level << 1) |

  550.                     !!run_level] + ctx->run_bits[run_level];

  551.             last_non_zero = i;

  552.         }

  553.     }

  554.     return bits;

  555. }

  556. 

  557. static av_always_inline

  558. void dnxhd_get_blocks(DNXHDEncContext *ctx, int mb_x, int mb_y)

  559. {

  560.     const int bs = ctx->block_width_l2;

  561.     const int bw = 1 << bs;

  562.     const uint8_t *ptr_y = ctx->thread[0]->src[0] +

  563.                            ((mb_y << 4) * ctx->m.linesize) + (mb_x << bs + 1);

  564.     const uint8_t *ptr_u = ctx->thread[0]->src[1] +

  565.                            ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs);

  566.     const uint8_t *ptr_v = ctx->thread[0]->src[2] +

  567.                            ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs);

  568.     PixblockDSPContext *pdsp = &ctx->m.pdsp;

  569. 

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

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

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

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

  574. 

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

  576.         if (ctx->interlaced) {

  577.             ctx->get_pixels_8x4_sym(ctx->blocks[4],

  578.                                     ptr_y + ctx->dct_y_offset,

  579.                                     ctx->m.linesize);

  580.             ctx->get_pixels_8x4_sym(ctx->blocks[5],

  581.                                     ptr_y + ctx->dct_y_offset + bw,

  582.                                     ctx->m.linesize);

  583.             ctx->get_pixels_8x4_sym(ctx->blocks[6],

  584.                                     ptr_u + ctx->dct_uv_offset,

  585.                                     ctx->m.uvlinesize);

  586.             ctx->get_pixels_8x4_sym(ctx->blocks[7],

  587.                                     ptr_v + ctx->dct_uv_offset,

  588.                                     ctx->m.uvlinesize);

  589.         } else {

  590.             ctx->bdsp.clear_block(ctx->blocks[4]);

  591.             ctx->bdsp.clear_block(ctx->blocks[5]);

  592.             ctx->bdsp.clear_block(ctx->blocks[6]);

  593.             ctx->bdsp.clear_block(ctx->blocks[7]);

  594.         }

  595.     } else {

  596.         pdsp->get_pixels(ctx->blocks[4],

  597.                          ptr_y + ctx->dct_y_offset, ctx->m.linesize);

  598.         pdsp->get_pixels(ctx->blocks[5],

  599.                          ptr_y + ctx->dct_y_offset + bw, ctx->m.linesize);

  600.         pdsp->get_pixels(ctx->blocks[6],

  601.                          ptr_u + ctx->dct_uv_offset, ctx->m.uvlinesize);

  602.         pdsp->get_pixels(ctx->blocks[7],

  603.                          ptr_v + ctx->dct_uv_offset, ctx->m.uvlinesize);

  604.     }

  605. }

  606. 

  607. static av_always_inline

  608. int dnxhd_switch_matrix(DNXHDEncContext *ctx, int i)

  609. {

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

  611.     return component[i];

  612. }

  613. 

  614. static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg,

  615.                                   int jobnr, int threadnr)

  616. {

  617.     DNXHDEncContext *ctx = avctx->priv_data;

  618.     int mb_y = jobnr, mb_x;

  619.     int qscale = ctx->qscale;

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

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

  622. 

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

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

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

  626. 

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

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

  629.         int ssd     = 0;

  630.         int ac_bits = 0;

  631.         int dc_bits = 0;

  632.         int i;

  633. 

  634.         dnxhd_get_blocks(ctx, mb_x, mb_y);

  635. 

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

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

  638.             int overflow, nbits, diff, last_index;

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

  640. 

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

  642.             last_index = ctx->m.dct_quantize(&ctx->m, block, 4 & (2*i),

  643.                                              qscale, &overflow);

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

  645. 

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

  647.             if (diff < 0)

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

  649.             else

  650.                 nbits = av_log2_16bit(2 * diff);

  651. 

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

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

  654. 

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

  656. 

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

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

  659.                 ctx->m.idsp.idct(block);

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

  661.             }

  662.         }

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

  664.         ctx->mb_rc[qscale][mb].bits = ac_bits + dc_bits + 12 +

  665.                                       8 * ctx->vlc_bits[0];

  666.     }

  667.     return 0;

  668. }

  669. 

  670. static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg,

  671.                                int jobnr, int threadnr)

  672. {

  673.     DNXHDEncContext *ctx = avctx->priv_data;

  674.     int mb_y = jobnr, mb_x;

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

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

  677.                   ctx->slice_size[jobnr]);

  678. 

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

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

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

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

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

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

  685.         int i;

  686. 

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

  688. 

  689.         dnxhd_get_blocks(ctx, mb_x, mb_y);

  690. 

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

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

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

  694.             int last_index = ctx->m.dct_quantize(&ctx->m, block, 4 & (2*i),

  695.                                                  qscale, &overflow);

  696.             // START_TIMER;

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

  698.             // STOP_TIMER("encode_block");

  699.         }

  700.     }

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

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

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

  704.     return 0;

  705. }

  706. 

  707. static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx)

  708. {

  709.     int mb_y, mb_x;

  710.     int offset = 0;

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

  712.         int thread_size;

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

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

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

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

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

  718.         }

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

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

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

  722.         offset += thread_size;

  723.     }

  724. }

  725. 

  726. static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg,

  727.                                int jobnr, int threadnr)

  728. {

  729.     DNXHDEncContext *ctx = avctx->priv_data;

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

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

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

  733. 

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

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

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

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

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

  739.             int sum;

  740.             int varc;

  741. 

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

  743.                 sum  = ctx->m.mpvencdsp.pix_sum(pix, ctx->m.linesize);

  744.                 varc = ctx->m.mpvencdsp.pix_norm1(pix, ctx->m.linesize);

  745.             } else {

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

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

  748.                 sum = varc = 0;

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

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

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

  752.                         sum  += val;

  753.                         varc += val * val;

  754.                     }

  755.                 }

  756.             }

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

  758. 

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

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

  761.         }

  762.     } else { // 10-bit

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

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

  765.             uint16_t *pix = (uint16_t *)ctx->thread[0]->src[0] +

  766.                             ((mb_y << 4) * linesize) + (mb_x << 4);

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

  768.             int sum = 0;

  769.             int sqsum = 0;

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

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

  772.             int mean, sqmean;

  773.             int i, j;

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

  775.             for (i = 0; i < bh; ++i) {

  776.                 for (j = 0; j < bw; ++j) {

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

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

  779.                     sum   += sample;

  780.                     sqsum += sample * sample;

  781.                     // 2^10 * 2^10 * 16 * 16 = 2^28, which is less than INT_MAX

  782.                 }

  783.                 pix += linesize;

  784.             }

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

  786.             sqmean = sqsum >> 8;

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

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

  789.         }

  790.     }

  791.     return 0;

  792. }

  793. 

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

  795. {

  796.     int lambda, up_step, down_step;

  797.     int last_lower = INT_MAX, last_higher = 0;

  798.     int x, y, q;

  799. 

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

  801.         ctx->qscale = q;

  802.         avctx->execute2(avctx, dnxhd_calc_bits_thread,

  803.                         NULL, NULL, ctx->m.mb_height);

  804.     }

  805.     up_step = down_step = 2 << LAMBDA_FRAC_BITS;

  806.     lambda  = ctx->lambda;

  807. 

  808.     for (;;) {

  809.         int bits = 0;

  810.         int end  = 0;

  811.         if (lambda == last_higher) {

  812.             lambda++;

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

  814.         }

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

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

  817.                 unsigned min = UINT_MAX;

  818.                 int qscale = 1;

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

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

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

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

  823.                     if (score < min) {

  824.                         min    = score;

  825.                         qscale = q;

  826.                     }

  827.                 }

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

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

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

  831.             }

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

  833.             if (bits > ctx->frame_bits)

  834.                 break;

  835.         }

  836.         // ff_dlog(ctx->m.avctx,

  837.         //         "lambda %d, up %u, down %u, bits %d, frame %d\n",

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

  839.         if (end) {

  840.             if (bits > ctx->frame_bits)

  841.                 return AVERROR(EINVAL);

  842.             break;

  843.         }

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

  845.             last_lower = FFMIN(lambda, last_lower);

  846.             if (last_higher != 0)

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

  848.             else

  849.                 lambda -= down_step;

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

  851.             up_step = 1<<LAMBDA_FRAC_BITS;

  852.             lambda = FFMAX(1, lambda);

  853.             if (lambda == last_lower)

  854.                 break;

  855.         } else {

  856.             last_higher = FFMAX(lambda, last_higher);

  857.             if (last_lower != INT_MAX)

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

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

  860.                 return AVERROR(EINVAL);

  861.             else

  862.                 lambda += up_step;

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

  864.             down_step = 1<<LAMBDA_FRAC_BITS;

  865.         }

  866.     }

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

  868.     ctx->lambda = lambda;

  869.     return 0;

  870. }

  871. 

  872. static int dnxhd_find_qscale(DNXHDEncContext *ctx)

  873. {

  874.     int bits = 0;

  875.     int up_step = 1;

  876.     int down_step = 1;

  877.     int last_higher = 0;

  878.     int last_lower = INT_MAX;

  879.     int qscale;

  880.     int x, y;

  881. 

  882.     qscale = ctx->qscale;

  883.     for (;;) {

  884.         bits = 0;

  885.         ctx->qscale = qscale;

  886.         // XXX avoid recalculating bits

  887.         ctx->m.avctx->execute2(ctx->m.avctx, dnxhd_calc_bits_thread,

  888.                                NULL, NULL, ctx->m.mb_height);

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

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

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

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

  893.             if (bits > ctx->frame_bits)

  894.                 break;

  895.         }

  896.         // ff_dlog(ctx->m.avctx,

  897.         //         "%d, qscale %d, bits %d, frame %d, higher %d, lower %d\n",

  898.         //         ctx->m.avctx->frame_number, qscale, bits, ctx->frame_bits,

  899.         //         last_higher, last_lower);

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

  901.             if (qscale == 1)

  902.                 return 1;

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

  904.                 qscale = last_higher;

  905.                 break;

  906.             }

  907.             last_lower = FFMIN(qscale, last_lower);

  908.             if (last_higher != 0)

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

  910.             else

  911.                 qscale -= down_step++;

  912.             if (qscale < 1)

  913.                 qscale = 1;

  914.             up_step = 1;

  915.         } else {

  916.             if (last_lower == qscale + 1)

  917.                 break;

  918.             last_higher = FFMAX(qscale, last_higher);

  919.             if (last_lower != INT_MAX)

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

  921.             else

  922.                 qscale += up_step++;

  923.             down_step = 1;

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

  925.                 return AVERROR(EINVAL);

  926.         }

  927.     }

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

  929.     ctx->qscale = qscale;

  930.     return 0;

  931. }

  932. 

  933. #define BUCKET_BITS 8

  934. #define RADIX_PASSES 4

  935. #define NBUCKETS (1 << BUCKET_BITS)

  936. 

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

  938. {

  939.     value >>= shift;

  940.     value  &= NBUCKETS - 1;

  941.     return NBUCKETS - 1 - value;

  942. }

  943. 

  944. static void radix_count(const RCCMPEntry *data, int size,

  945.                         int buckets[RADIX_PASSES][NBUCKETS])

  946. {

  947.     int i, j;

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

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

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

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

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

  953.             v >>= BUCKET_BITS;

  954.         }

  955.         av_assert1(!v);

  956.     }

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

  958.         int offset = size;

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

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

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

  962.     }

  963. }

  964. 

  965. static void radix_sort_pass(RCCMPEntry *dst, const RCCMPEntry *data,

  966.                             int size, int buckets[NBUCKETS], int pass)

  967. {

  968.     int shift = pass * BUCKET_BITS;

  969.     int i;

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

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

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

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

  974.     }

  975. }

  976. 

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

  978. {

  979.     int buckets[RADIX_PASSES][NBUCKETS];

  980.     RCCMPEntry *tmp = av_malloc_array(size, sizeof(*tmp));

  981.     radix_count(data, size, buckets);

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

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

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

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

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

  987.     }

  988.     av_free(tmp);

  989. }

  990. 

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

  992. {

  993.     int max_bits = 0;

  994.     int ret, x, y;

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

  996.         return ret;

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

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

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

  1000.             int delta_bits;

  1001.             ctx->mb_qscale[mb] = ctx->qscale;

  1002.             ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale][mb].bits;

  1003.             max_bits += ctx->mb_rc[ctx->qscale][mb].bits;

  1004.             if (!RC_VARIANCE) {

  1005.                 delta_bits = ctx->mb_rc[ctx->qscale][mb].bits -

  1006.                              ctx->mb_rc[ctx->qscale + 1][mb].bits;

  1007.                 ctx->mb_cmp[mb].mb = mb;

  1008.                 ctx->mb_cmp[mb].value =

  1009.                     delta_bits ? ((ctx->mb_rc[ctx->qscale][mb].ssd -

  1010.                                    ctx->mb_rc[ctx->qscale + 1][mb].ssd) * 100) /

  1011.                                   delta_bits

  1012.                                : INT_MIN; // avoid increasing qscale

  1013.             }

  1014.         }

  1015.         max_bits += 31; // worst padding

  1016.     }

  1017.     if (!ret) {

  1018.         if (RC_VARIANCE)

  1019.             avctx->execute2(avctx, dnxhd_mb_var_thread,

  1020.                             NULL, NULL, ctx->m.mb_height);

  1021.         radix_sort(ctx->mb_cmp, ctx->m.mb_num);

  1022.         for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) {

  1023.             int mb = ctx->mb_cmp[x].mb;

  1024.             max_bits -= ctx->mb_rc[ctx->qscale][mb].bits -

  1025.                         ctx->mb_rc[ctx->qscale + 1][mb].bits;

  1026.             ctx->mb_qscale[mb] = ctx->qscale + 1;

  1027.             ctx->mb_bits[mb]   = ctx->mb_rc[ctx->qscale + 1][mb].bits;

  1028.         }

  1029.     }

  1030.     return 0;

  1031. }

  1032. 

  1033. static void dnxhd_load_picture(DNXHDEncContext *ctx, const AVFrame *frame)

  1034. {

  1035.     int i;

  1036. 

  1037.     for (i = 0; i < ctx->m.avctx->thread_count; i++) {

  1038.         ctx->thread[i]->m.linesize    = frame->linesize[0] << ctx->interlaced;

  1039.         ctx->thread[i]->m.uvlinesize  = frame->linesize[1] << ctx->interlaced;

  1040.         ctx->thread[i]->dct_y_offset  = ctx->m.linesize  *8;

  1041.         ctx->thread[i]->dct_uv_offset = ctx->m.uvlinesize*8;

  1042.     }

  1043. 

  1044. #if FF_API_CODED_FRAME

  1045. FF_DISABLE_DEPRECATION_WARNINGS

  1046.     ctx->m.avctx->coded_frame->interlaced_frame = frame->interlaced_frame;

  1047. FF_ENABLE_DEPRECATION_WARNINGS

  1048. #endif

  1049.     ctx->cur_field = frame->interlaced_frame && !frame->top_field_first;

  1050. }

  1051. 

  1052. static int dnxhd_encode_picture(AVCodecContext *avctx, AVPacket *pkt,

  1053.                                 const AVFrame *frame, int *got_packet)

  1054. {

  1055.     DNXHDEncContext *ctx = avctx->priv_data;

  1056.     int first_field = 1;

  1057.     int offset, i, ret;

  1058.     uint8_t *buf;

  1059. 

  1060.     if ((ret = ff_alloc_packet2(avctx, pkt, ctx->cid_table->frame_size, 0)) < 0)

  1061.         return ret;

  1062.     buf = pkt->data;

  1063. 

  1064.     dnxhd_load_picture(ctx, frame);

  1065. 

  1066. encode_coding_unit:

  1067.     for (i = 0; i < 3; i++) {

  1068.         ctx->src[i] = frame->data[i];

  1069.         if (ctx->interlaced && ctx->cur_field)

  1070.             ctx->src[i] += frame->linesize[i];

  1071.     }

  1072. 

  1073.     dnxhd_write_header(avctx, buf);

  1074. 

  1075.     if (avctx->mb_decision == FF_MB_DECISION_RD)

  1076.         ret = dnxhd_encode_rdo(avctx, ctx);

  1077.     else

  1078.         ret = dnxhd_encode_fast(avctx, ctx);

  1079.     if (ret < 0) {

  1080.         av_log(avctx, AV_LOG_ERROR,

  1081.                "picture could not fit ratecontrol constraints, increase qmax\n");

  1082.         return ret;

  1083.     }

  1084. 

  1085.     dnxhd_setup_threads_slices(ctx);

  1086. 

  1087.     offset = 0;

  1088.     for (i = 0; i < ctx->m.mb_height; i++) {

  1089.         AV_WB32(ctx->msip + i * 4, offset);

  1090.         offset += ctx->slice_size[i];

  1091.         av_assert1(!(ctx->slice_size[i] & 3));

  1092.     }

  1093. 

  1094.     avctx->execute2(avctx, dnxhd_encode_thread, buf, NULL, ctx->m.mb_height);

  1095. 

  1096.     av_assert1(640 + offset + 4 <= ctx->cid_table->coding_unit_size);

  1097.     memset(buf + 640 + offset, 0,

  1098.            ctx->cid_table->coding_unit_size - 4 - offset - 640);

  1099. 

  1100.     AV_WB32(buf + ctx->cid_table->coding_unit_size - 4, 0x600DC0DE); // EOF

  1101. 

  1102.     if (ctx->interlaced && first_field) {

  1103.         first_field     = 0;

  1104.         ctx->cur_field ^= 1;

  1105.         buf            += ctx->cid_table->coding_unit_size;

  1106.         goto encode_coding_unit;

  1107.     }

  1108. 

  1109. #if FF_API_CODED_FRAME

  1110. FF_DISABLE_DEPRECATION_WARNINGS

  1111.     avctx->coded_frame->quality = ctx->qscale * FF_QP2LAMBDA;

  1112. FF_ENABLE_DEPRECATION_WARNINGS

  1113. #endif

  1114. 

  1115.     ff_side_data_set_encoder_stats(pkt, ctx->qscale * FF_QP2LAMBDA, NULL, 0, AV_PICTURE_TYPE_I);

  1116. 

  1117.     pkt->flags |= AV_PKT_FLAG_KEY;

  1118.     *got_packet = 1;

  1119.     return 0;

  1120. }

  1121. 

  1122. static av_cold int dnxhd_encode_end(AVCodecContext *avctx)

  1123. {

  1124.     DNXHDEncContext *ctx = avctx->priv_data;

  1125.     int max_level        = 1 << (ctx->cid_table->bit_depth + 2);

  1126.     int i;

  1127. 

  1128.     av_free(ctx->vlc_codes - max_level * 2);

  1129.     av_free(ctx->vlc_bits - max_level * 2);

  1130.     av_freep(&ctx->run_codes);

  1131.     av_freep(&ctx->run_bits);

  1132. 

  1133.     av_freep(&ctx->mb_bits);

  1134.     av_freep(&ctx->mb_qscale);

  1135.     av_freep(&ctx->mb_rc);

  1136.     av_freep(&ctx->mb_cmp);

  1137.     av_freep(&ctx->slice_size);

  1138.     av_freep(&ctx->slice_offs);

  1139. 

  1140.     av_freep(&ctx->qmatrix_c);

  1141.     av_freep(&ctx->qmatrix_l);

  1142.     av_freep(&ctx->qmatrix_c16);

  1143.     av_freep(&ctx->qmatrix_l16);

  1144. 

  1145.     for (i = 1; i < avctx->thread_count; i++)

  1146.         av_freep(&ctx->thread[i]);

  1147. 

  1148.     return 0;

  1149. }

  1150. 

  1151. static const AVCodecDefault dnxhd_defaults[] = {

  1152.     { "qmax", "1024" }, /* Maximum quantization scale factor allowed for VC-3 */

  1153.     { NULL },

  1154. };

  1155. 

  1156. AVCodec ff_dnxhd_encoder = {

  1157.     .name           = "dnxhd",

  1158.     .long_name      = NULL_IF_CONFIG_SMALL("VC3/DNxHD"),

  1159.     .type           = AVMEDIA_TYPE_VIDEO,

  1160.     .id             = AV_CODEC_ID_DNXHD,

  1161.     .priv_data_size = sizeof(DNXHDEncContext),

  1162.     .init           = dnxhd_encode_init,

  1163.     .encode2        = dnxhd_encode_picture,

  1164.     .close          = dnxhd_encode_end,

  1165.     .capabilities   = AV_CODEC_CAP_SLICE_THREADS,

  1166.     .pix_fmts       = (const enum AVPixelFormat[]) {

  1167.         AV_PIX_FMT_YUV422P,

  1168.         AV_PIX_FMT_YUV422P10,

  1169.         AV_PIX_FMT_NONE

  1170.     },

  1171.     .priv_class     = &dnxhd_class,

  1172.     .defaults       = dnxhd_defaults,

  1173. };