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.
91728 Commits
32 Branches
312MB
Tree: aeb73c73e1
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'aeb73c73e1'
${ noResults }
FFmpeg/libavcodec/dnxhdenc.c

1406 lines
50KB
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 "profiles.h"

  38. #include "dnxhdenc.h"

  39. 

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

  41. #define DNX10BIT_QMAT_SHIFT 18

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

  43. #define LAMBDA_FRAC_BITS 10

  44. 

  45. #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM

  46. static const AVOption options[] = {

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

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

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

  50.         offsetof(DNXHDEncContext, intra_quant_bias), AV_OPT_TYPE_INT,

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

  52.     { "profile",       NULL, offsetof(DNXHDEncContext, profile), AV_OPT_TYPE_INT,

  53.         { .i64 = FF_PROFILE_DNXHD },

  54.         FF_PROFILE_DNXHD, FF_PROFILE_DNXHR_444, VE, "profile" },

  55.     { "dnxhd",     NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_PROFILE_DNXHD },

  56.         0, 0, VE, "profile" },

  57.     { "dnxhr_444", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_PROFILE_DNXHR_444 },

  58.         0, 0, VE, "profile" },

  59.     { "dnxhr_hqx", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_PROFILE_DNXHR_HQX },

  60.         0, 0, VE, "profile" },

  61.     { "dnxhr_hq",  NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_PROFILE_DNXHR_HQ },

  62.         0, 0, VE, "profile" },

  63.     { "dnxhr_sq",  NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_PROFILE_DNXHR_SQ },

  64.         0, 0, VE, "profile" },

  65.     { "dnxhr_lb",  NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_PROFILE_DNXHR_LB },

  66.         0, 0, VE, "profile" },

  67.     { NULL }

  68. };

  69. 

  70. static const AVClass dnxhd_class = {

  71.     .class_name = "dnxhd",

  72.     .item_name  = av_default_item_name,

  73.     .option     = options,

  74.     .version    = LIBAVUTIL_VERSION_INT,

  75. };

  76. 

  77. static void dnxhd_8bit_get_pixels_8x4_sym(int16_t *av_restrict block,

  78.                                           const uint8_t *pixels,

  79.                                           ptrdiff_t line_size)

  80. {

  81.     int i;

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

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

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

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

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

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

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

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

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

  91.         pixels  += line_size;

  92.         block   += 8;

  93.     }

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

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

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

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

  98. }

  99. 

  100. static av_always_inline

  101. void dnxhd_10bit_get_pixels_8x4_sym(int16_t *av_restrict block,

  102.                                     const uint8_t *pixels,

  103.                                     ptrdiff_t line_size)

  104. {

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

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

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

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

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

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

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

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

  113. }

  114. 

  115. static int dnxhd_10bit_dct_quantize_444(MpegEncContext *ctx, int16_t *block,

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

  117. {

  118.     int i, j, level, last_non_zero, start_i;

  119.     const int *qmat;

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

  121.     int bias;

  122.     int max = 0;

  123.     unsigned int threshold1, threshold2;

  124. 

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

  126. 

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

  128.     start_i = 1;

  129.     last_non_zero = 0;

  130.     qmat = n < 4 ? ctx->q_intra_matrix[qscale] : ctx->q_chroma_intra_matrix[qscale];

  131.     bias= ctx->intra_quant_bias * (1 << (16 - 8));

  132.     threshold1 = (1 << 16) - bias - 1;

  133.     threshold2 = (threshold1 << 1);

  134. 

  135.     for (i = 63; i >= start_i; i--) {

  136.         j = scantable[i];

  137.         level = block[j] * qmat[j];

  138. 

  139.         if (((unsigned)(level + threshold1)) > threshold2) {

  140.             last_non_zero = i;

  141.             break;

  142.         } else{

  143.             block[j]=0;

  144.         }

  145.     }

  146. 

  147.     for (i = start_i; i <= last_non_zero; i++) {

  148.         j = scantable[i];

  149.         level = block[j] * qmat[j];

  150. 

  151.         if (((unsigned)(level + threshold1)) > threshold2) {

  152.             if (level > 0) {

  153.                 level = (bias + level) >> 16;

  154.                 block[j] = level;

  155.             } else{

  156.                 level = (bias - level) >> 16;

  157.                 block[j] = -level;

  158.             }

  159.             max |= level;

  160.         } else {

  161.             block[j] = 0;

  162.         }

  163.     }

  164.     *overflow = ctx->max_qcoeff < max; //overflow might have happened

  165. 

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

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

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

  169.                          scantable, last_non_zero);

  170. 

  171.     return last_non_zero;

  172. }

  173. 

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

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

  176. {

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

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

  179.     int last_non_zero = 0;

  180.     int i;

  181. 

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

  183. 

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

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

  186. 

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

  188.         int j = scantable[i];

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

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

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

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

  193.         if (level)

  194.             last_non_zero = i;

  195.     }

  196. 

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

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

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

  200.                          scantable, last_non_zero);

  201. 

  202.     return last_non_zero;

  203. }

  204. 

  205. static av_cold int dnxhd_init_vlc(DNXHDEncContext *ctx)

  206. {

  207.     int i, j, level, run;

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

  209. 

  210.     FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->orig_vlc_codes,

  211.                       max_level, 4 * sizeof(*ctx->orig_vlc_codes), fail);

  212.     FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->orig_vlc_bits,

  213.                       max_level, 4 * sizeof(*ctx->orig_vlc_bits), fail);

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

  215.                       63 * 2, fail);

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

  217.                       63, fail);

  218. 

  219.     ctx->vlc_codes = ctx->orig_vlc_codes + max_level * 2;

  220.     ctx->vlc_bits  = ctx->orig_vlc_bits + max_level * 2;

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

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

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

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

  225. 

  226.             MASK_ABS(sign, alevel);

  227.             if (alevel > 64) {

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

  229.                 alevel -= offset << 6;

  230.             }

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

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

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

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

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

  236.                     if (alevel) {

  237.                         ctx->vlc_codes[index] =

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

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

  240.                     } else {

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

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

  243.                     }

  244.                     break;

  245.                 }

  246.             }

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

  248.             if (offset) {

  249.                 ctx->vlc_codes[index] =

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

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

  252.             }

  253.         }

  254.     }

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

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

  257.         av_assert0(run < 63);

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

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

  260.     }

  261.     return 0;

  262. fail:

  263.     return AVERROR(ENOMEM);

  264. }

  265. 

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

  267. {

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

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

  270.     int qscale, i;

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

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

  273. 

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

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

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

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

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

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

  280.                       fail);

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

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

  283.                       fail);

  284. 

  285.     if (ctx->bit_depth == 8) {

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

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

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

  289.         }

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

  291.                           weight_matrix, ctx->intra_quant_bias, 1,

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

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

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

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

  296.         }

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

  298.                           weight_matrix, ctx->intra_quant_bias, 1,

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

  300. 

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

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

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

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

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

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

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

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

  309.             }

  310.         }

  311.     } else {

  312.         // 10-bit

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

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

  315.                 int j = ff_zigzag_direct[i];

  316. 

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

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

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

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

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

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

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

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

  325.                  *     ((1 << DNX10BIT_QMAT_SHIFT) * (p / s)) /

  326.                  *     (qscale * weight_table[i])

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

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

  329.                                             (qscale * luma_weight_table[i]);

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

  331.                                             (qscale * chroma_weight_table[i]);

  332.             }

  333.         }

  334.     }

  335. 

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

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

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

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

  340. 

  341.     return 0;

  342. fail:

  343.     return AVERROR(ENOMEM);

  344. }

  345. 

  346. static av_cold int dnxhd_init_rc(DNXHDEncContext *ctx)

  347. {

  348.     FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->mb_rc, (ctx->m.avctx->qmax + 1),

  349.                           ctx->m.mb_num * sizeof(RCEntry), fail);

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

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

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

  353.         FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->mb_cmp_tmp,

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

  355.     }

  356.     ctx->frame_bits = (ctx->coding_unit_size -

  357.                        ctx->data_offset - 4 - ctx->min_padding) * 8;

  358.     ctx->qscale = 1;

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

  360.     return 0;

  361. fail:

  362.     return AVERROR(ENOMEM);

  363. }

  364. 

  365. static av_cold int dnxhd_encode_init(AVCodecContext *avctx)

  366. {

  367.     DNXHDEncContext *ctx = avctx->priv_data;

  368.     int i, index, ret;

  369. 

  370.     switch (avctx->pix_fmt) {

  371.     case AV_PIX_FMT_YUV422P:

  372.         ctx->bit_depth = 8;

  373.         break;

  374.     case AV_PIX_FMT_YUV422P10:

  375.     case AV_PIX_FMT_YUV444P10:

  376.     case AV_PIX_FMT_GBRP10:

  377.         ctx->bit_depth = 10;

  378.         break;

  379.     default:

  380.         av_log(avctx, AV_LOG_ERROR,

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

  382.         return AVERROR(EINVAL);

  383.     }

  384. 

  385.     if ((ctx->profile == FF_PROFILE_DNXHR_444 && (avctx->pix_fmt != AV_PIX_FMT_YUV444P10 &&

  386.                                                   avctx->pix_fmt != AV_PIX_FMT_GBRP10)) ||

  387.         (ctx->profile != FF_PROFILE_DNXHR_444 && (avctx->pix_fmt == AV_PIX_FMT_YUV444P10 ||

  388.                                                   avctx->pix_fmt == AV_PIX_FMT_GBRP10))) {

  389.         av_log(avctx, AV_LOG_ERROR,

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

  391.         return AVERROR(EINVAL);

  392.     }

  393. 

  394.     if (ctx->profile == FF_PROFILE_DNXHR_HQX && avctx->pix_fmt != AV_PIX_FMT_YUV422P10) {

  395.         av_log(avctx, AV_LOG_ERROR,

  396.                "pixel format is incompatible with DNxHR HQX profile\n");

  397.         return AVERROR(EINVAL);

  398.     }

  399. 

  400.     if ((ctx->profile == FF_PROFILE_DNXHR_LB ||

  401.          ctx->profile == FF_PROFILE_DNXHR_SQ ||

  402.          ctx->profile == FF_PROFILE_DNXHR_HQ) && avctx->pix_fmt != AV_PIX_FMT_YUV422P) {

  403.         av_log(avctx, AV_LOG_ERROR,

  404.                "pixel format is incompatible with DNxHR LB/SQ/HQ profile\n");

  405.         return AVERROR(EINVAL);

  406.     }

  407. 

  408.     ctx->is_444 = ctx->profile == FF_PROFILE_DNXHR_444;

  409.     avctx->profile = ctx->profile;

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

  411.     if (!ctx->cid) {

  412.         av_log(avctx, AV_LOG_ERROR,

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

  414.         ff_dnxhd_print_profiles(avctx, AV_LOG_ERROR);

  415.         return AVERROR(EINVAL);

  416.     }

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

  418. 

  419.     if (ctx->cid >= 1270 && ctx->cid <= 1274)

  420.         avctx->codec_tag = MKTAG('A','V','d','h');

  421. 

  422.     if (avctx->width < 256 || avctx->height < 120) {

  423.         av_log(avctx, AV_LOG_ERROR,

  424.                "Input dimensions too small, input must be at least 256x120\n");

  425.         return AVERROR(EINVAL);

  426.     }

  427. 

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

  429.     av_assert0(index >= 0);

  430. 

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

  432. 

  433.     ctx->m.avctx    = avctx;

  434.     ctx->m.mb_intra = 1;

  435.     ctx->m.h263_aic = 1;

  436. 

  437.     avctx->bits_per_raw_sample = ctx->bit_depth;

  438. 

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

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

  441.     ff_mpv_idct_init(&ctx->m);

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

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

  444.     ff_dct_encode_init(&ctx->m);

  445. 

  446.     if (ctx->profile != FF_PROFILE_DNXHD)

  447.         ff_videodsp_init(&ctx->m.vdsp, ctx->bit_depth);

  448. 

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

  450.         ctx->m.dct_quantize = ff_dct_quantize_c;

  451. 

  452.     if (ctx->is_444 || ctx->profile == FF_PROFILE_DNXHR_HQX) {

  453.         ctx->m.dct_quantize     = dnxhd_10bit_dct_quantize_444;

  454.         ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym;

  455.         ctx->block_width_l2     = 4;

  456.     } else if (ctx->bit_depth == 10) {

  457.         ctx->m.dct_quantize     = dnxhd_10bit_dct_quantize;

  458.         ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym;

  459.         ctx->block_width_l2     = 4;

  460.     } else {

  461.         ctx->get_pixels_8x4_sym = dnxhd_8bit_get_pixels_8x4_sym;

  462.         ctx->block_width_l2     = 3;

  463.     }

  464. 

  465.     if (ARCH_X86)

  466.         ff_dnxhdenc_init_x86(ctx);

  467. 

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

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

  470. 

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

  472.         ctx->interlaced   = 1;

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

  474.     }

  475. 

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

  477. 

  478.     if (ctx->cid_table->frame_size == DNXHD_VARIABLE) {

  479.         ctx->frame_size = ff_dnxhd_get_hr_frame_size(ctx->cid,

  480.                                                      avctx->width, avctx->height);

  481.         av_assert0(ctx->frame_size >= 0);

  482.         ctx->coding_unit_size = ctx->frame_size;

  483.     } else {

  484.         ctx->frame_size = ctx->cid_table->frame_size;

  485.         ctx->coding_unit_size = ctx->cid_table->coding_unit_size;

  486.     }

  487. 

  488.     if (ctx->m.mb_height > 68)

  489.         ctx->data_offset = 0x170 + (ctx->m.mb_height << 2);

  490.     else

  491.         ctx->data_offset = 0x280;

  492. 

  493.     // XXX tune lbias/cbias

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

  495.         return ret;

  496. 

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

  498.      * in the coding unit payload */

  499.     if (ctx->nitris_compat)

  500.         ctx->min_padding = 1600;

  501. 

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

  503.         return ret;

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

  505.         return ret;

  506. 

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

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

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

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

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

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

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

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

  515. 

  516. #if FF_API_CODED_FRAME

  517. FF_DISABLE_DEPRECATION_WARNINGS

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

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

  520. FF_ENABLE_DEPRECATION_WARNINGS

  521. #endif

  522. 

  523.     if (avctx->active_thread_type == FF_THREAD_SLICE) {

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

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

  526.             return AVERROR(EINVAL);

  527.         }

  528.     }

  529. 

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

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

  532.         return AVERROR(EINVAL);

  533.     }

  534. 

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

  536.     if (avctx->active_thread_type == FF_THREAD_SLICE) {

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

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

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

  540.         }

  541.     }

  542. 

  543.     return 0;

  544. fail:  // for FF_ALLOCZ_OR_GOTO

  545.     return AVERROR(ENOMEM);

  546. }

  547. 

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

  549. {

  550.     DNXHDEncContext *ctx = avctx->priv_data;

  551. 

  552.     memset(buf, 0, ctx->data_offset);

  553. 

  554.     // * write prefix */

  555.     AV_WB16(buf + 0x02, ctx->data_offset);

  556.     if (ctx->cid >= 1270 && ctx->cid <= 1274)

  557.         buf[4] = 0x03;

  558.     else

  559.         buf[4] = 0x01;

  560. 

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

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

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

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

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

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

  567. 

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

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

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

  571.     buf[0x2c] = (!ctx->interlaced << 7) | (ctx->is_444 << 6) | (avctx->pix_fmt == AV_PIX_FMT_YUV444P10);

  572. 

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

  574. 

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

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

  577.     AV_WB16(buf + 0x16c, ctx->m.mb_height); // Ns

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

  579. 

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

  581.     return 0;

  582. }

  583. 

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

  585. {

  586.     int nbits;

  587.     if (diff < 0) {

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

  589.         diff--;

  590.     } else {

  591.         nbits = av_log2_16bit(2 * diff);

  592.     }

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

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

  595.              av_mod_uintp2(diff, nbits));

  596. }

  597. 

  598. static av_always_inline

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

  600.                         int last_index, int n)

  601. {

  602.     int last_non_zero = 0;

  603.     int slevel, i, j;

  604. 

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

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

  607. 

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

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

  610.         slevel = block[j];

  611.         if (slevel) {

  612.             int run_level = i - last_non_zero - 1;

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

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

  615.             if (run_level)

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

  617.                          ctx->run_codes[run_level]);

  618.             last_non_zero = i;

  619.         }

  620.     }

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

  622. }

  623. 

  624. static av_always_inline

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

  626.                         int qscale, int last_index)

  627. {

  628.     const uint8_t *weight_matrix;

  629.     int level;

  630.     int i;

  631. 

  632.     if (ctx->is_444) {

  633.         weight_matrix = ((n % 6) < 2) ? ctx->cid_table->luma_weight

  634.                                       : ctx->cid_table->chroma_weight;

  635.     } else {

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

  637.                                 : ctx->cid_table->luma_weight;

  638.     }

  639. 

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

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

  642.         level = block[j];

  643.         if (level) {

  644.             if (level < 0) {

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

  646.                 if (ctx->bit_depth == 10) {

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

  648.                         level += 8;

  649.                     level >>= 4;

  650.                 } else {

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

  652.                         level += 32;

  653.                     level >>= 6;

  654.                 }

  655.                 level = -level;

  656.             } else {

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

  658.                 if (ctx->bit_depth == 10) {

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

  660.                         level += 8;

  661.                     level >>= 4;

  662.                 } else {

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

  664.                         level += 32;

  665.                     level >>= 6;

  666.                 }

  667.             }

  668.             block[j] = level;

  669.         }

  670.     }

  671. }

  672. 

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

  674. {

  675.     int score = 0;

  676.     int i;

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

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

  679.     return score;

  680. }

  681. 

  682. static av_always_inline

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

  684. {

  685.     int last_non_zero = 0;

  686.     int bits = 0;

  687.     int i, j, level;

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

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

  690.         level = block[j];

  691.         if (level) {

  692.             int run_level = i - last_non_zero - 1;

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

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

  695.             last_non_zero = i;

  696.         }

  697.     }

  698.     return bits;

  699. }

  700. 

  701. static av_always_inline

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

  703. {

  704.     const int bs = ctx->block_width_l2;

  705.     const int bw = 1 << bs;

  706.     int dct_y_offset = ctx->dct_y_offset;

  707.     int dct_uv_offset = ctx->dct_uv_offset;

  708.     int linesize = ctx->m.linesize;

  709.     int uvlinesize = ctx->m.uvlinesize;

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

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

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

  713.                            ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs + ctx->is_444);

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

  715.                            ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs + ctx->is_444);

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

  717.     VideoDSPContext *vdsp = &ctx->m.vdsp;

  718. 

  719.     if (ctx->bit_depth != 10 && vdsp->emulated_edge_mc && ((mb_x << 4) + 16 > ctx->m.avctx->width ||

  720.                                                            (mb_y << 4) + 16 > ctx->m.avctx->height)) {

  721.         int y_w = ctx->m.avctx->width  - (mb_x << 4);

  722.         int y_h = ctx->m.avctx->height - (mb_y << 4);

  723.         int uv_w = (y_w + 1) / 2;

  724.         int uv_h = y_h;

  725.         linesize = 16;

  726.         uvlinesize = 8;

  727. 

  728.         vdsp->emulated_edge_mc(&ctx->edge_buf_y[0], ptr_y,

  729.                                linesize, ctx->m.linesize,

  730.                                linesize, 16,

  731.                                0, 0, y_w, y_h);

  732.         vdsp->emulated_edge_mc(&ctx->edge_buf_uv[0][0], ptr_u,

  733.                                uvlinesize, ctx->m.uvlinesize,

  734.                                uvlinesize, 16,

  735.                                0, 0, uv_w, uv_h);

  736.         vdsp->emulated_edge_mc(&ctx->edge_buf_uv[1][0], ptr_v,

  737.                                uvlinesize, ctx->m.uvlinesize,

  738.                                uvlinesize, 16,

  739.                                0, 0, uv_w, uv_h);

  740. 

  741.         dct_y_offset =  bw * linesize;

  742.         dct_uv_offset = bw * uvlinesize;

  743.         ptr_y = &ctx->edge_buf_y[0];

  744.         ptr_u = &ctx->edge_buf_uv[0][0];

  745.         ptr_v = &ctx->edge_buf_uv[1][0];

  746.     } else if (ctx->bit_depth == 10 && vdsp->emulated_edge_mc && ((mb_x << 4) + 16 > ctx->m.avctx->width ||

  747.                                                                   (mb_y << 4) + 16 > ctx->m.avctx->height)) {

  748.         int y_w = ctx->m.avctx->width  - (mb_x << 4);

  749.         int y_h = ctx->m.avctx->height - (mb_y << 4);

  750.         int uv_w = ctx->is_444 ? y_w : (y_w + 1) / 2;

  751.         int uv_h = y_h;

  752.         linesize = 32;

  753.         uvlinesize = 16 + 16 * ctx->is_444;

  754. 

  755.         vdsp->emulated_edge_mc(&ctx->edge_buf_y[0], ptr_y,

  756.                                linesize, ctx->m.linesize,

  757.                                linesize / 2, 16,

  758.                                0, 0, y_w, y_h);

  759.         vdsp->emulated_edge_mc(&ctx->edge_buf_uv[0][0], ptr_u,

  760.                                uvlinesize, ctx->m.uvlinesize,

  761.                                uvlinesize / 2, 16,

  762.                                0, 0, uv_w, uv_h);

  763.         vdsp->emulated_edge_mc(&ctx->edge_buf_uv[1][0], ptr_v,

  764.                                uvlinesize, ctx->m.uvlinesize,

  765.                                uvlinesize / 2, 16,

  766.                                0, 0, uv_w, uv_h);

  767. 

  768.         dct_y_offset =  bw * linesize / 2;

  769.         dct_uv_offset = bw * uvlinesize / 2;

  770.         ptr_y = &ctx->edge_buf_y[0];

  771.         ptr_u = &ctx->edge_buf_uv[0][0];

  772.         ptr_v = &ctx->edge_buf_uv[1][0];

  773.     }

  774. 

  775.     if (!ctx->is_444) {

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

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

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

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

  780. 

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

  782.             if (ctx->interlaced) {

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

  784.                                         ptr_y + dct_y_offset,

  785.                                         linesize);

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

  787.                                         ptr_y + dct_y_offset + bw,

  788.                                         linesize);

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

  790.                                         ptr_u + dct_uv_offset,

  791.                                         uvlinesize);

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

  793.                                         ptr_v + dct_uv_offset,

  794.                                         uvlinesize);

  795.             } else {

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

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

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

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

  800.             }

  801.         } else {

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

  803.                              ptr_y + dct_y_offset, linesize);

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

  805.                              ptr_y + dct_y_offset + bw, linesize);

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

  807.                              ptr_u + dct_uv_offset, uvlinesize);

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

  809.                              ptr_v + dct_uv_offset, uvlinesize);

  810.         }

  811.     } else {

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

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

  814.         pdsp->get_pixels(ctx->blocks[6], ptr_y + dct_y_offset, linesize);

  815.         pdsp->get_pixels(ctx->blocks[7], ptr_y + dct_y_offset + bw, linesize);

  816. 

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

  818.         pdsp->get_pixels(ctx->blocks[3], ptr_u + bw, uvlinesize);

  819.         pdsp->get_pixels(ctx->blocks[8], ptr_u + dct_uv_offset, uvlinesize);

  820.         pdsp->get_pixels(ctx->blocks[9], ptr_u + dct_uv_offset + bw, uvlinesize);

  821. 

  822.         pdsp->get_pixels(ctx->blocks[4], ptr_v,      uvlinesize);

  823.         pdsp->get_pixels(ctx->blocks[5], ptr_v + bw, uvlinesize);

  824.         pdsp->get_pixels(ctx->blocks[10], ptr_v + dct_uv_offset, uvlinesize);

  825.         pdsp->get_pixels(ctx->blocks[11], ptr_v + dct_uv_offset + bw, uvlinesize);

  826.     }

  827. }

  828. 

  829. static av_always_inline

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

  831. {

  832.     int x;

  833. 

  834.     if (ctx->is_444) {

  835.         x = (i >> 1) % 3;

  836.     } else {

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

  838.         x = component[i];

  839.     }

  840.     return x;

  841. }

  842. 

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

  844.                                   int jobnr, int threadnr)

  845. {

  846.     DNXHDEncContext *ctx = avctx->priv_data;

  847.     int mb_y = jobnr, mb_x;

  848.     int qscale = ctx->qscale;

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

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

  851. 

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

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

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

  855. 

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

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

  858.         int ssd     = 0;

  859.         int ac_bits = 0;

  860.         int dc_bits = 0;

  861.         int i;

  862. 

  863.         dnxhd_get_blocks(ctx, mb_x, mb_y);

  864. 

  865.         for (i = 0; i < 8 + 4 * ctx->is_444; i++) {

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

  867.             int overflow, nbits, diff, last_index;

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

  869. 

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

  871.             last_index = ctx->m.dct_quantize(&ctx->m, block,

  872.                                              ctx->is_444 ? 4 * (n > 0): 4 & (2*i),

  873.                                              qscale, &overflow);

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

  875. 

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

  877.             if (diff < 0)

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

  879.             else

  880.                 nbits = av_log2_16bit(2 * diff);

  881. 

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

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

  884. 

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

  886. 

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

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

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

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

  891.             }

  892.         }

  893.         ctx->mb_rc[(qscale * ctx->m.mb_num) + mb].ssd  = ssd;

  894.         ctx->mb_rc[(qscale * ctx->m.mb_num) + mb].bits = ac_bits + dc_bits + 12 +

  895.                                      (1 + ctx->is_444) * 8 * ctx->vlc_bits[0];

  896.     }

  897.     return 0;

  898. }

  899. 

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

  901.                                int jobnr, int threadnr)

  902. {

  903.     DNXHDEncContext *ctx = avctx->priv_data;

  904.     int mb_y = jobnr, mb_x;

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

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

  907.                   ctx->slice_size[jobnr]);

  908. 

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

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

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

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

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

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

  915.         int i;

  916. 

  917.         put_bits(&ctx->m.pb, 11, qscale);

  918.         put_bits(&ctx->m.pb, 1, avctx->pix_fmt == AV_PIX_FMT_YUV444P10);

  919. 

  920.         dnxhd_get_blocks(ctx, mb_x, mb_y);

  921. 

  922.         for (i = 0; i < 8 + 4 * ctx->is_444; i++) {

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

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

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

  926.                                                  ctx->is_444 ? (((i >> 1) % 3) < 1 ? 0 : 4): 4 & (2*i),

  927.                                                  qscale, &overflow);

  928.             // START_TIMER;

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

  930.             // STOP_TIMER("encode_block");

  931.         }

  932.     }

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

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

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

  936.     return 0;

  937. }

  938. 

  939. static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx)

  940. {

  941.     int mb_y, mb_x;

  942.     int offset = 0;

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

  944.         int thread_size;

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

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

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

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

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

  950.         }

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

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

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

  954.         offset += thread_size;

  955.     }

  956. }

  957. 

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

  959.                                int jobnr, int threadnr)

  960. {

  961.     DNXHDEncContext *ctx = avctx->priv_data;

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

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

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

  965. 

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

  967.     if (ctx->bit_depth == 8) {

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

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

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

  971.             int sum;

  972.             int varc;

  973. 

  974.             if (!partial_last_row && mb_x * 16 <= avctx->width - 16 && (avctx->width % 16) == 0) {

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

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

  977.             } else {

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

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

  980.                 sum = varc = 0;

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

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

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

  984.                         sum  += val;

  985.                         varc += val * val;

  986.                     }

  987.                 }

  988.             }

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

  990. 

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

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

  993.         }

  994.     } else { // 10-bit

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

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

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

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

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

  1000.             int sum = 0;

  1001.             int sqsum = 0;

  1002.             int bw = FFMIN(avctx->width - 16 * mb_x, 16);

  1003.             int bh = FFMIN((avctx->height >> ctx->interlaced) - 16 * mb_y, 16);

  1004.             int mean, sqmean;

  1005.             int i, j;

  1006.             // Macroblocks are 16x16 pixels, unlike DCT blocks which are 8x8.

  1007.             for (i = 0; i < bh; ++i) {

  1008.                 for (j = 0; j < bw; ++j) {

  1009.                     // Turn 16-bit pixels into 10-bit ones.

  1010.                     const int sample = (unsigned) pix[j] >> 6;

  1011.                     sum   += sample;

  1012.                     sqsum += sample * sample;

  1013.                     // 2^10 * 2^10 * 16 * 16 = 2^28, which is less than INT_MAX

  1014.                 }

  1015.                 pix += linesize;

  1016.             }

  1017.             mean = sum >> 8; // 16*16 == 2^8

  1018.             sqmean = sqsum >> 8;

  1019.             ctx->mb_cmp[mb].value = sqmean - mean * mean;

  1020.             ctx->mb_cmp[mb].mb    = mb;

  1021.         }

  1022.     }

  1023.     return 0;

  1024. }

  1025. 

  1026. static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx)

  1027. {

  1028.     int lambda, up_step, down_step;

  1029.     int last_lower = INT_MAX, last_higher = 0;

  1030.     int x, y, q;

  1031. 

  1032.     for (q = 1; q < avctx->qmax; q++) {

  1033.         ctx->qscale = q;

  1034.         avctx->execute2(avctx, dnxhd_calc_bits_thread,

  1035.                         NULL, NULL, ctx->m.mb_height);

  1036.     }

  1037.     up_step = down_step = 2 << LAMBDA_FRAC_BITS;

  1038.     lambda  = ctx->lambda;

  1039. 

  1040.     for (;;) {

  1041.         int bits = 0;

  1042.         int end  = 0;

  1043.         if (lambda == last_higher) {

  1044.             lambda++;

  1045.             end = 1; // need to set final qscales/bits

  1046.         }

  1047.         for (y = 0; y < ctx->m.mb_height; y++) {

  1048.             for (x = 0; x < ctx->m.mb_width; x++) {

  1049.                 unsigned min = UINT_MAX;

  1050.                 int qscale = 1;

  1051.                 int mb     = y * ctx->m.mb_width + x;

  1052.                 int rc = 0;

  1053.                 for (q = 1; q < avctx->qmax; q++) {

  1054.                     int i = (q*ctx->m.mb_num) + mb;

  1055.                     unsigned score = ctx->mb_rc[i].bits * lambda +

  1056.                                      ((unsigned) ctx->mb_rc[i].ssd << LAMBDA_FRAC_BITS);

  1057.                     if (score < min) {

  1058.                         min    = score;

  1059.                         qscale = q;

  1060.                         rc = i;

  1061.                     }

  1062.                 }

  1063.                 bits += ctx->mb_rc[rc].bits;

  1064.                 ctx->mb_qscale[mb] = qscale;

  1065.                 ctx->mb_bits[mb]   = ctx->mb_rc[rc].bits;

  1066.             }

  1067.             bits = (bits + 31) & ~31; // padding

  1068.             if (bits > ctx->frame_bits)

  1069.                 break;

  1070.         }

  1071.         if (end) {

  1072.             if (bits > ctx->frame_bits)

  1073.                 return AVERROR(EINVAL);

  1074.             break;

  1075.         }

  1076.         if (bits < ctx->frame_bits) {

  1077.             last_lower = FFMIN(lambda, last_lower);

  1078.             if (last_higher != 0)

  1079.                 lambda = (lambda+last_higher)>>1;

  1080.             else

  1081.                 lambda -= down_step;

  1082.             down_step = FFMIN((int64_t)down_step*5, INT_MAX);

  1083.             up_step = 1<<LAMBDA_FRAC_BITS;

  1084.             lambda = FFMAX(1, lambda);

  1085.             if (lambda == last_lower)

  1086.                 break;

  1087.         } else {

  1088.             last_higher = FFMAX(lambda, last_higher);

  1089.             if (last_lower != INT_MAX)

  1090.                 lambda = (lambda+last_lower)>>1;

  1091.             else if ((int64_t)lambda + up_step > INT_MAX)

  1092.                 return AVERROR(EINVAL);

  1093.             else

  1094.                 lambda += up_step;

  1095.             up_step = FFMIN((int64_t)up_step*5, INT_MAX);

  1096.             down_step = 1<<LAMBDA_FRAC_BITS;

  1097.         }

  1098.     }

  1099.     ctx->lambda = lambda;

  1100.     return 0;

  1101. }

  1102. 

  1103. static int dnxhd_find_qscale(DNXHDEncContext *ctx)

  1104. {

  1105.     int bits = 0;

  1106.     int up_step = 1;

  1107.     int down_step = 1;

  1108.     int last_higher = 0;

  1109.     int last_lower = INT_MAX;

  1110.     int qscale;

  1111.     int x, y;

  1112. 

  1113.     qscale = ctx->qscale;

  1114.     for (;;) {

  1115.         bits = 0;

  1116.         ctx->qscale = qscale;

  1117.         // XXX avoid recalculating bits

  1118.         ctx->m.avctx->execute2(ctx->m.avctx, dnxhd_calc_bits_thread,

  1119.                                NULL, NULL, ctx->m.mb_height);

  1120.         for (y = 0; y < ctx->m.mb_height; y++) {

  1121.             for (x = 0; x < ctx->m.mb_width; x++)

  1122.                 bits += ctx->mb_rc[(qscale*ctx->m.mb_num) + (y*ctx->m.mb_width+x)].bits;

  1123.             bits = (bits+31)&~31; // padding

  1124.             if (bits > ctx->frame_bits)

  1125.                 break;

  1126.         }

  1127.         if (bits < ctx->frame_bits) {

  1128.             if (qscale == 1)

  1129.                 return 1;

  1130.             if (last_higher == qscale - 1) {

  1131.                 qscale = last_higher;

  1132.                 break;

  1133.             }

  1134.             last_lower = FFMIN(qscale, last_lower);

  1135.             if (last_higher != 0)

  1136.                 qscale = (qscale + last_higher) >> 1;

  1137.             else

  1138.                 qscale -= down_step++;

  1139.             if (qscale < 1)

  1140.                 qscale = 1;

  1141.             up_step = 1;

  1142.         } else {

  1143.             if (last_lower == qscale + 1)

  1144.                 break;

  1145.             last_higher = FFMAX(qscale, last_higher);

  1146.             if (last_lower != INT_MAX)

  1147.                 qscale = (qscale + last_lower) >> 1;

  1148.             else

  1149.                 qscale += up_step++;

  1150.             down_step = 1;

  1151.             if (qscale >= ctx->m.avctx->qmax)

  1152.                 return AVERROR(EINVAL);

  1153.         }

  1154.     }

  1155.     ctx->qscale = qscale;

  1156.     return 0;

  1157. }

  1158. 

  1159. #define BUCKET_BITS 8

  1160. #define RADIX_PASSES 4

  1161. #define NBUCKETS (1 << BUCKET_BITS)

  1162. 

  1163. static inline int get_bucket(int value, int shift)

  1164. {

  1165.     value >>= shift;

  1166.     value  &= NBUCKETS - 1;

  1167.     return NBUCKETS - 1 - value;

  1168. }

  1169. 

  1170. static void radix_count(const RCCMPEntry *data, int size,

  1171.                         int buckets[RADIX_PASSES][NBUCKETS])

  1172. {

  1173.     int i, j;

  1174.     memset(buckets, 0, sizeof(buckets[0][0]) * RADIX_PASSES * NBUCKETS);

  1175.     for (i = 0; i < size; i++) {

  1176.         int v = data[i].value;

  1177.         for (j = 0; j < RADIX_PASSES; j++) {

  1178.             buckets[j][get_bucket(v, 0)]++;

  1179.             v >>= BUCKET_BITS;

  1180.         }

  1181.         av_assert1(!v);

  1182.     }

  1183.     for (j = 0; j < RADIX_PASSES; j++) {

  1184.         int offset = size;

  1185.         for (i = NBUCKETS - 1; i >= 0; i--)

  1186.             buckets[j][i] = offset -= buckets[j][i];

  1187.         av_assert1(!buckets[j][0]);

  1188.     }

  1189. }

  1190. 

  1191. static void radix_sort_pass(RCCMPEntry *dst, const RCCMPEntry *data,

  1192.                             int size, int buckets[NBUCKETS], int pass)

  1193. {

  1194.     int shift = pass * BUCKET_BITS;

  1195.     int i;

  1196.     for (i = 0; i < size; i++) {

  1197.         int v   = get_bucket(data[i].value, shift);

  1198.         int pos = buckets[v]++;

  1199.         dst[pos] = data[i];

  1200.     }

  1201. }

  1202. 

  1203. static void radix_sort(RCCMPEntry *data, RCCMPEntry *tmp, int size)

  1204. {

  1205.     int buckets[RADIX_PASSES][NBUCKETS];

  1206.     radix_count(data, size, buckets);

  1207.     radix_sort_pass(tmp, data, size, buckets[0], 0);

  1208.     radix_sort_pass(data, tmp, size, buckets[1], 1);

  1209.     if (buckets[2][NBUCKETS - 1] || buckets[3][NBUCKETS - 1]) {

  1210.         radix_sort_pass(tmp, data, size, buckets[2], 2);

  1211.         radix_sort_pass(data, tmp, size, buckets[3], 3);

  1212.     }

  1213. }

  1214. 

  1215. static int dnxhd_encode_fast(AVCodecContext *avctx, DNXHDEncContext *ctx)

  1216. {

  1217.     int max_bits = 0;

  1218.     int ret, x, y;

  1219.     if ((ret = dnxhd_find_qscale(ctx)) < 0)

  1220.         return ret;

  1221.     for (y = 0; y < ctx->m.mb_height; y++) {

  1222.         for (x = 0; x < ctx->m.mb_width; x++) {

  1223.             int mb = y * ctx->m.mb_width + x;

  1224.             int rc = (ctx->qscale * ctx->m.mb_num ) + mb;

  1225.             int delta_bits;

  1226.             ctx->mb_qscale[mb] = ctx->qscale;

  1227.             ctx->mb_bits[mb] = ctx->mb_rc[rc].bits;

  1228.             max_bits += ctx->mb_rc[rc].bits;

  1229.             if (!RC_VARIANCE) {

  1230.                 delta_bits = ctx->mb_rc[rc].bits -

  1231.                              ctx->mb_rc[rc + ctx->m.mb_num].bits;

  1232.                 ctx->mb_cmp[mb].mb = mb;

  1233.                 ctx->mb_cmp[mb].value =

  1234.                     delta_bits ? ((ctx->mb_rc[rc].ssd -

  1235.                                    ctx->mb_rc[rc + ctx->m.mb_num].ssd) * 100) /

  1236.                                   delta_bits

  1237.                                : INT_MIN; // avoid increasing qscale

  1238.             }

  1239.         }

  1240.         max_bits += 31; // worst padding

  1241.     }

  1242.     if (!ret) {

  1243.         if (RC_VARIANCE)

  1244.             avctx->execute2(avctx, dnxhd_mb_var_thread,

  1245.                             NULL, NULL, ctx->m.mb_height);

  1246.         radix_sort(ctx->mb_cmp, ctx->mb_cmp_tmp, ctx->m.mb_num);

  1247.         for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) {

  1248.             int mb = ctx->mb_cmp[x].mb;

  1249.             int rc = (ctx->qscale * ctx->m.mb_num ) + mb;

  1250.             max_bits -= ctx->mb_rc[rc].bits -

  1251.                         ctx->mb_rc[rc + ctx->m.mb_num].bits;

  1252.             ctx->mb_qscale[mb] = ctx->qscale + 1;

  1253.             ctx->mb_bits[mb]   = ctx->mb_rc[rc + ctx->m.mb_num].bits;

  1254.         }

  1255.     }

  1256.     return 0;

  1257. }

  1258. 

  1259. static void dnxhd_load_picture(DNXHDEncContext *ctx, const AVFrame *frame)

  1260. {

  1261.     int i;

  1262. 

  1263.     for (i = 0; i < ctx->m.avctx->thread_count; i++) {

  1264.         ctx->thread[i]->m.linesize    = frame->linesize[0] << ctx->interlaced;

  1265.         ctx->thread[i]->m.uvlinesize  = frame->linesize[1] << ctx->interlaced;

  1266.         ctx->thread[i]->dct_y_offset  = ctx->m.linesize  *8;

  1267.         ctx->thread[i]->dct_uv_offset = ctx->m.uvlinesize*8;

  1268.     }

  1269. 

  1270. #if FF_API_CODED_FRAME

  1271. FF_DISABLE_DEPRECATION_WARNINGS

  1272.     ctx->m.avctx->coded_frame->interlaced_frame = frame->interlaced_frame;

  1273. FF_ENABLE_DEPRECATION_WARNINGS

  1274. #endif

  1275.     ctx->cur_field = frame->interlaced_frame && !frame->top_field_first;

  1276. }

  1277. 

  1278. static int dnxhd_encode_picture(AVCodecContext *avctx, AVPacket *pkt,

  1279.                                 const AVFrame *frame, int *got_packet)

  1280. {

  1281.     DNXHDEncContext *ctx = avctx->priv_data;

  1282.     int first_field = 1;

  1283.     int offset, i, ret;

  1284.     uint8_t *buf;

  1285. 

  1286.     if ((ret = ff_alloc_packet2(avctx, pkt, ctx->frame_size, 0)) < 0)

  1287.         return ret;

  1288.     buf = pkt->data;

  1289. 

  1290.     dnxhd_load_picture(ctx, frame);

  1291. 

  1292. encode_coding_unit:

  1293.     for (i = 0; i < 3; i++) {

  1294.         ctx->src[i] = frame->data[i];

  1295.         if (ctx->interlaced && ctx->cur_field)

  1296.             ctx->src[i] += frame->linesize[i];

  1297.     }

  1298. 

  1299.     dnxhd_write_header(avctx, buf);

  1300. 

  1301.     if (avctx->mb_decision == FF_MB_DECISION_RD)

  1302.         ret = dnxhd_encode_rdo(avctx, ctx);

  1303.     else

  1304.         ret = dnxhd_encode_fast(avctx, ctx);

  1305.     if (ret < 0) {

  1306.         av_log(avctx, AV_LOG_ERROR,

  1307.                "picture could not fit ratecontrol constraints, increase qmax\n");

  1308.         return ret;

  1309.     }

  1310. 

  1311.     dnxhd_setup_threads_slices(ctx);

  1312. 

  1313.     offset = 0;

  1314.     for (i = 0; i < ctx->m.mb_height; i++) {

  1315.         AV_WB32(ctx->msip + i * 4, offset);

  1316.         offset += ctx->slice_size[i];

  1317.         av_assert1(!(ctx->slice_size[i] & 3));

  1318.     }

  1319. 

  1320.     avctx->execute2(avctx, dnxhd_encode_thread, buf, NULL, ctx->m.mb_height);

  1321. 

  1322.     av_assert1(ctx->data_offset + offset + 4 <= ctx->coding_unit_size);

  1323.     memset(buf + ctx->data_offset + offset, 0,

  1324.            ctx->coding_unit_size - 4 - offset - ctx->data_offset);

  1325. 

  1326.     AV_WB32(buf + ctx->coding_unit_size - 4, 0x600DC0DE); // EOF

  1327. 

  1328.     if (ctx->interlaced && first_field) {

  1329.         first_field     = 0;

  1330.         ctx->cur_field ^= 1;

  1331.         buf            += ctx->coding_unit_size;

  1332.         goto encode_coding_unit;

  1333.     }

  1334. 

  1335. #if FF_API_CODED_FRAME

  1336. FF_DISABLE_DEPRECATION_WARNINGS

  1337.     avctx->coded_frame->quality = ctx->qscale * FF_QP2LAMBDA;

  1338. FF_ENABLE_DEPRECATION_WARNINGS

  1339. #endif

  1340. 

  1341.     ff_side_data_set_encoder_stats(pkt, ctx->qscale * FF_QP2LAMBDA, NULL, 0, AV_PICTURE_TYPE_I);

  1342. 

  1343.     pkt->flags |= AV_PKT_FLAG_KEY;

  1344.     *got_packet = 1;

  1345.     return 0;

  1346. }

  1347. 

  1348. static av_cold int dnxhd_encode_end(AVCodecContext *avctx)

  1349. {

  1350.     DNXHDEncContext *ctx = avctx->priv_data;

  1351.     int i;

  1352. 

  1353.     av_freep(&ctx->orig_vlc_codes);

  1354.     av_freep(&ctx->orig_vlc_bits);

  1355.     av_freep(&ctx->run_codes);

  1356.     av_freep(&ctx->run_bits);

  1357. 

  1358.     av_freep(&ctx->mb_bits);

  1359.     av_freep(&ctx->mb_qscale);

  1360.     av_freep(&ctx->mb_rc);

  1361.     av_freep(&ctx->mb_cmp);

  1362.     av_freep(&ctx->mb_cmp_tmp);

  1363.     av_freep(&ctx->slice_size);

  1364.     av_freep(&ctx->slice_offs);

  1365. 

  1366.     av_freep(&ctx->qmatrix_c);

  1367.     av_freep(&ctx->qmatrix_l);

  1368.     av_freep(&ctx->qmatrix_c16);

  1369.     av_freep(&ctx->qmatrix_l16);

  1370. 

  1371.     if (avctx->active_thread_type == FF_THREAD_SLICE) {

  1372.         for (i = 1; i < avctx->thread_count; i++)

  1373.             av_freep(&ctx->thread[i]);

  1374.     }

  1375. 

  1376.     return 0;

  1377. }

  1378. 

  1379. static const AVCodecDefault dnxhd_defaults[] = {

  1380.     { "qmax", "1024" }, /* Maximum quantization scale factor allowed for VC-3 */

  1381.     { NULL },

  1382. };

  1383. 

  1384. AVCodec ff_dnxhd_encoder = {

  1385.     .name           = "dnxhd",

  1386.     .long_name      = NULL_IF_CONFIG_SMALL("VC3/DNxHD"),

  1387.     .type           = AVMEDIA_TYPE_VIDEO,

  1388.     .id             = AV_CODEC_ID_DNXHD,

  1389.     .priv_data_size = sizeof(DNXHDEncContext),

  1390.     .init           = dnxhd_encode_init,

  1391.     .encode2        = dnxhd_encode_picture,

  1392.     .close          = dnxhd_encode_end,

  1393.     .capabilities   = AV_CODEC_CAP_SLICE_THREADS | AV_CODEC_CAP_FRAME_THREADS | AV_CODEC_CAP_INTRA_ONLY,

  1394.     .caps_internal  = FF_CODEC_CAP_INIT_CLEANUP,

  1395.     .pix_fmts       = (const enum AVPixelFormat[]) {

  1396.         AV_PIX_FMT_YUV422P,

  1397.         AV_PIX_FMT_YUV422P10,

  1398.         AV_PIX_FMT_YUV444P10,

  1399.         AV_PIX_FMT_GBRP10,

  1400.         AV_PIX_FMT_NONE

  1401.     },

  1402.     .priv_class     = &dnxhd_class,

  1403.     .defaults       = dnxhd_defaults,

  1404.     .profiles       = NULL_IF_CONFIG_SMALL(ff_dnxhd_profiles),

  1405. };