PageRenderTime 206ms CodeModel.GetById 74ms app.highlight 120ms RepoModel.GetById 1ms app.codeStats 0ms

/libavcodec/dnxhdenc.c

http://github.com/FFmpeg/FFmpeg
C | 1411 lines | 1209 code | 153 blank | 49 comment | 244 complexity | d6780bcd77606d7ab96536dc485f9d50 MD5 | raw file

Large files files are truncated, but you can click here to view the full file

   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
  30#include "avcodec.h"
  31#include "blockdsp.h"
  32#include "fdctdsp.h"
  33#include "internal.h"
  34#include "mpegvideo.h"
  35#include "pixblockdsp.h"
  36#include "profiles.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
  45static 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    { "profile",       NULL, offsetof(DNXHDEncContext, profile), AV_OPT_TYPE_INT,
  52        { .i64 = FF_PROFILE_DNXHD },
  53        FF_PROFILE_DNXHD, FF_PROFILE_DNXHR_444, VE, "profile" },
  54    { "dnxhd",     NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_PROFILE_DNXHD },
  55        0, 0, VE, "profile" },
  56    { "dnxhr_444", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_PROFILE_DNXHR_444 },
  57        0, 0, VE, "profile" },
  58    { "dnxhr_hqx", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_PROFILE_DNXHR_HQX },
  59        0, 0, VE, "profile" },
  60    { "dnxhr_hq",  NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_PROFILE_DNXHR_HQ },
  61        0, 0, VE, "profile" },
  62    { "dnxhr_sq",  NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_PROFILE_DNXHR_SQ },
  63        0, 0, VE, "profile" },
  64    { "dnxhr_lb",  NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_PROFILE_DNXHR_LB },
  65        0, 0, VE, "profile" },
  66    { NULL }
  67};
  68
  69static const AVClass dnxhd_class = {
  70    .class_name = "dnxhd",
  71    .item_name  = av_default_item_name,
  72    .option     = options,
  73    .version    = LIBAVUTIL_VERSION_INT,
  74};
  75
  76static void dnxhd_8bit_get_pixels_8x4_sym(int16_t *av_restrict block,
  77                                          const uint8_t *pixels,
  78                                          ptrdiff_t line_size)
  79{
  80    int i;
  81    for (i = 0; i < 4; i++) {
  82        block[0] = pixels[0];
  83        block[1] = pixels[1];
  84        block[2] = pixels[2];
  85        block[3] = pixels[3];
  86        block[4] = pixels[4];
  87        block[5] = pixels[5];
  88        block[6] = pixels[6];
  89        block[7] = pixels[7];
  90        pixels  += line_size;
  91        block   += 8;
  92    }
  93    memcpy(block,      block -  8, sizeof(*block) * 8);
  94    memcpy(block +  8, block - 16, sizeof(*block) * 8);
  95    memcpy(block + 16, block - 24, sizeof(*block) * 8);
  96    memcpy(block + 24, block - 32, sizeof(*block) * 8);
  97}
  98
  99static av_always_inline
 100void dnxhd_10bit_get_pixels_8x4_sym(int16_t *av_restrict block,
 101                                    const uint8_t *pixels,
 102                                    ptrdiff_t line_size)
 103{
 104    memcpy(block + 0 * 8, pixels + 0 * line_size, 8 * sizeof(*block));
 105    memcpy(block + 7 * 8, pixels + 0 * line_size, 8 * sizeof(*block));
 106    memcpy(block + 1 * 8, pixels + 1 * line_size, 8 * sizeof(*block));
 107    memcpy(block + 6 * 8, pixels + 1 * line_size, 8 * sizeof(*block));
 108    memcpy(block + 2 * 8, pixels + 2 * line_size, 8 * sizeof(*block));
 109    memcpy(block + 5 * 8, pixels + 2 * line_size, 8 * sizeof(*block));
 110    memcpy(block + 3 * 8, pixels + 3 * line_size, 8 * sizeof(*block));
 111    memcpy(block + 4 * 8, pixels + 3 * line_size, 8 * sizeof(*block));
 112}
 113
 114static int dnxhd_10bit_dct_quantize_444(MpegEncContext *ctx, int16_t *block,
 115                                        int n, int qscale, int *overflow)
 116{
 117    int i, j, level, last_non_zero, start_i;
 118    const int *qmat;
 119    const uint8_t *scantable= ctx->intra_scantable.scantable;
 120    int bias;
 121    int max = 0;
 122    unsigned int threshold1, threshold2;
 123
 124    ctx->fdsp.fdct(block);
 125
 126    block[0] = (block[0] + 2) >> 2;
 127    start_i = 1;
 128    last_non_zero = 0;
 129    qmat = n < 4 ? ctx->q_intra_matrix[qscale] : ctx->q_chroma_intra_matrix[qscale];
 130    bias= ctx->intra_quant_bias * (1 << (16 - 8));
 131    threshold1 = (1 << 16) - bias - 1;
 132    threshold2 = (threshold1 << 1);
 133
 134    for (i = 63; i >= start_i; i--) {
 135        j = scantable[i];
 136        level = block[j] * qmat[j];
 137
 138        if (((unsigned)(level + threshold1)) > threshold2) {
 139            last_non_zero = i;
 140            break;
 141        } else{
 142            block[j]=0;
 143        }
 144    }
 145
 146    for (i = start_i; i <= last_non_zero; i++) {
 147        j = scantable[i];
 148        level = block[j] * qmat[j];
 149
 150        if (((unsigned)(level + threshold1)) > threshold2) {
 151            if (level > 0) {
 152                level = (bias + level) >> 16;
 153                block[j] = level;
 154            } else{
 155                level = (bias - level) >> 16;
 156                block[j] = -level;
 157            }
 158            max |= level;
 159        } else {
 160            block[j] = 0;
 161        }
 162    }
 163    *overflow = ctx->max_qcoeff < max; //overflow might have happened
 164
 165    /* we need this permutation so that we correct the IDCT, we only permute the !=0 elements */
 166    if (ctx->idsp.perm_type != FF_IDCT_PERM_NONE)
 167        ff_block_permute(block, ctx->idsp.idct_permutation,
 168                         scantable, last_non_zero);
 169
 170    return last_non_zero;
 171}
 172
 173static int dnxhd_10bit_dct_quantize(MpegEncContext *ctx, int16_t *block,
 174                                    int n, int qscale, int *overflow)
 175{
 176    const uint8_t *scantable= ctx->intra_scantable.scantable;
 177    const int *qmat = n<4 ? ctx->q_intra_matrix[qscale] : ctx->q_chroma_intra_matrix[qscale];
 178    int last_non_zero = 0;
 179    int i;
 180
 181    ctx->fdsp.fdct(block);
 182
 183    // Divide by 4 with rounding, to compensate scaling of DCT coefficients
 184    block[0] = (block[0] + 2) >> 2;
 185
 186    for (i = 1; i < 64; ++i) {
 187        int j = scantable[i];
 188        int sign = FF_SIGNBIT(block[j]);
 189        int level = (block[j] ^ sign) - sign;
 190        level = level * qmat[j] >> DNX10BIT_QMAT_SHIFT;
 191        block[j] = (level ^ sign) - sign;
 192        if (level)
 193            last_non_zero = i;
 194    }
 195
 196    /* we need this permutation so that we correct the IDCT, we only permute the !=0 elements */
 197    if (ctx->idsp.perm_type != FF_IDCT_PERM_NONE)
 198        ff_block_permute(block, ctx->idsp.idct_permutation,
 199                         scantable, last_non_zero);
 200
 201    return last_non_zero;
 202}
 203
 204static av_cold int dnxhd_init_vlc(DNXHDEncContext *ctx)
 205{
 206    int i, j, level, run;
 207    int max_level = 1 << (ctx->bit_depth + 2);
 208
 209    FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->orig_vlc_codes,
 210                      max_level, 4 * sizeof(*ctx->orig_vlc_codes), fail);
 211    FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->orig_vlc_bits,
 212                      max_level, 4 * sizeof(*ctx->orig_vlc_bits), fail);
 213    FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_codes,
 214                      63 * 2, fail);
 215    FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_bits,
 216                      63, fail);
 217
 218    ctx->vlc_codes = ctx->orig_vlc_codes + max_level * 2;
 219    ctx->vlc_bits  = ctx->orig_vlc_bits + max_level * 2;
 220    for (level = -max_level; level < max_level; level++) {
 221        for (run = 0; run < 2; run++) {
 222            int index = level * (1 << 1) | run;
 223            int sign, offset = 0, alevel = level;
 224
 225            MASK_ABS(sign, alevel);
 226            if (alevel > 64) {
 227                offset  = (alevel - 1) >> 6;
 228                alevel -= offset << 6;
 229            }
 230            for (j = 0; j < 257; j++) {
 231                if (ctx->cid_table->ac_info[2*j+0] >> 1 == alevel &&
 232                    (!offset || (ctx->cid_table->ac_info[2*j+1] & 1) && offset) &&
 233                    (!run    || (ctx->cid_table->ac_info[2*j+1] & 2) && run)) {
 234                    av_assert1(!ctx->vlc_codes[index]);
 235                    if (alevel) {
 236                        ctx->vlc_codes[index] =
 237                            (ctx->cid_table->ac_codes[j] << 1) | (sign & 1);
 238                        ctx->vlc_bits[index] = ctx->cid_table->ac_bits[j] + 1;
 239                    } else {
 240                        ctx->vlc_codes[index] = ctx->cid_table->ac_codes[j];
 241                        ctx->vlc_bits[index]  = ctx->cid_table->ac_bits[j];
 242                    }
 243                    break;
 244                }
 245            }
 246            av_assert0(!alevel || j < 257);
 247            if (offset) {
 248                ctx->vlc_codes[index] =
 249                    (ctx->vlc_codes[index] << ctx->cid_table->index_bits) | offset;
 250                ctx->vlc_bits[index] += ctx->cid_table->index_bits;
 251            }
 252        }
 253    }
 254    for (i = 0; i < 62; i++) {
 255        int run = ctx->cid_table->run[i];
 256        av_assert0(run < 63);
 257        ctx->run_codes[run] = ctx->cid_table->run_codes[i];
 258        ctx->run_bits[run]  = ctx->cid_table->run_bits[i];
 259    }
 260    return 0;
 261fail:
 262    return AVERROR(ENOMEM);
 263}
 264
 265static av_cold int dnxhd_init_qmat(DNXHDEncContext *ctx, int lbias, int cbias)
 266{
 267    // init first elem to 1 to avoid div by 0 in convert_matrix
 268    uint16_t weight_matrix[64] = { 1, }; // convert_matrix needs uint16_t*
 269    int qscale, i;
 270    const uint8_t *luma_weight_table   = ctx->cid_table->luma_weight;
 271    const uint8_t *chroma_weight_table = ctx->cid_table->chroma_weight;
 272
 273    FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l,
 274                      (ctx->m.avctx->qmax + 1), 64 * sizeof(int), fail);
 275    FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c,
 276                      (ctx->m.avctx->qmax + 1), 64 * sizeof(int), fail);
 277    FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l16,
 278                      (ctx->m.avctx->qmax + 1), 64 * 2 * sizeof(uint16_t),
 279                      fail);
 280    FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c16,
 281                      (ctx->m.avctx->qmax + 1), 64 * 2 * sizeof(uint16_t),
 282                      fail);
 283
 284    if (ctx->bit_depth == 8) {
 285        for (i = 1; i < 64; i++) {
 286            int j = ctx->m.idsp.idct_permutation[ff_zigzag_direct[i]];
 287            weight_matrix[j] = ctx->cid_table->luma_weight[i];
 288        }
 289        ff_convert_matrix(&ctx->m, ctx->qmatrix_l, ctx->qmatrix_l16,
 290                          weight_matrix, ctx->intra_quant_bias, 1,
 291                          ctx->m.avctx->qmax, 1);
 292        for (i = 1; i < 64; i++) {
 293            int j = ctx->m.idsp.idct_permutation[ff_zigzag_direct[i]];
 294            weight_matrix[j] = ctx->cid_table->chroma_weight[i];
 295        }
 296        ff_convert_matrix(&ctx->m, ctx->qmatrix_c, ctx->qmatrix_c16,
 297                          weight_matrix, ctx->intra_quant_bias, 1,
 298                          ctx->m.avctx->qmax, 1);
 299
 300        for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
 301            for (i = 0; i < 64; i++) {
 302                ctx->qmatrix_l[qscale][i]      <<= 2;
 303                ctx->qmatrix_c[qscale][i]      <<= 2;
 304                ctx->qmatrix_l16[qscale][0][i] <<= 2;
 305                ctx->qmatrix_l16[qscale][1][i] <<= 2;
 306                ctx->qmatrix_c16[qscale][0][i] <<= 2;
 307                ctx->qmatrix_c16[qscale][1][i] <<= 2;
 308            }
 309        }
 310    } else {
 311        // 10-bit
 312        for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
 313            for (i = 1; i < 64; i++) {
 314                int j = ff_zigzag_direct[i];
 315
 316                /* The quantization formula from the VC-3 standard is:
 317                 * quantized = sign(block[i]) * floor(abs(block[i]/s) * p /
 318                 *             (qscale * weight_table[i]))
 319                 * Where p is 32 for 8-bit samples and 8 for 10-bit ones.
 320                 * The s factor compensates scaling of DCT coefficients done by
 321                 * the DCT routines, and therefore is not present in standard.
 322                 * It's 8 for 8-bit samples and 4 for 10-bit ones.
 323                 * We want values of ctx->qtmatrix_l and ctx->qtmatrix_r to be:
 324                 *     ((1 << DNX10BIT_QMAT_SHIFT) * (p / s)) /
 325                 *     (qscale * weight_table[i])
 326                 * For 10-bit samples, p / s == 2 */
 327                ctx->qmatrix_l[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) /
 328                                            (qscale * luma_weight_table[i]);
 329                ctx->qmatrix_c[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) /
 330                                            (qscale * chroma_weight_table[i]);
 331            }
 332        }
 333    }
 334
 335    ctx->m.q_chroma_intra_matrix16 = ctx->qmatrix_c16;
 336    ctx->m.q_chroma_intra_matrix   = ctx->qmatrix_c;
 337    ctx->m.q_intra_matrix16        = ctx->qmatrix_l16;
 338    ctx->m.q_intra_matrix          = ctx->qmatrix_l;
 339
 340    return 0;
 341fail:
 342    return AVERROR(ENOMEM);
 343}
 344
 345static av_cold int dnxhd_init_rc(DNXHDEncContext *ctx)
 346{
 347    FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->mb_rc, (ctx->m.avctx->qmax + 1),
 348                          ctx->m.mb_num * sizeof(RCEntry), fail);
 349    if (ctx->m.avctx->mb_decision != FF_MB_DECISION_RD) {
 350        FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->mb_cmp,
 351                          ctx->m.mb_num, sizeof(RCCMPEntry), fail);
 352        FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->mb_cmp_tmp,
 353                          ctx->m.mb_num, sizeof(RCCMPEntry), fail);
 354    }
 355    ctx->frame_bits = (ctx->coding_unit_size -
 356                       ctx->data_offset - 4 - ctx->min_padding) * 8;
 357    ctx->qscale = 1;
 358    ctx->lambda = 2 << LAMBDA_FRAC_BITS; // qscale 2
 359    return 0;
 360fail:
 361    return AVERROR(ENOMEM);
 362}
 363
 364static av_cold int dnxhd_encode_init(AVCodecContext *avctx)
 365{
 366    DNXHDEncContext *ctx = avctx->priv_data;
 367    int i, index, ret;
 368
 369    switch (avctx->pix_fmt) {
 370    case AV_PIX_FMT_YUV422P:
 371        ctx->bit_depth = 8;
 372        break;
 373    case AV_PIX_FMT_YUV422P10:
 374    case AV_PIX_FMT_YUV444P10:
 375    case AV_PIX_FMT_GBRP10:
 376        ctx->bit_depth = 10;
 377        break;
 378    default:
 379        av_log(avctx, AV_LOG_ERROR,
 380               "pixel format is incompatible with DNxHD\n");
 381        return AVERROR(EINVAL);
 382    }
 383
 384    if ((ctx->profile == FF_PROFILE_DNXHR_444 && (avctx->pix_fmt != AV_PIX_FMT_YUV444P10 &&
 385                                                  avctx->pix_fmt != AV_PIX_FMT_GBRP10)) ||
 386        (ctx->profile != FF_PROFILE_DNXHR_444 && (avctx->pix_fmt == AV_PIX_FMT_YUV444P10 ||
 387                                                  avctx->pix_fmt == AV_PIX_FMT_GBRP10))) {
 388        av_log(avctx, AV_LOG_ERROR,
 389               "pixel format is incompatible with DNxHD profile\n");
 390        return AVERROR(EINVAL);
 391    }
 392
 393    if (ctx->profile == FF_PROFILE_DNXHR_HQX && avctx->pix_fmt != AV_PIX_FMT_YUV422P10) {
 394        av_log(avctx, AV_LOG_ERROR,
 395               "pixel format is incompatible with DNxHR HQX profile\n");
 396        return AVERROR(EINVAL);
 397    }
 398
 399    if ((ctx->profile == FF_PROFILE_DNXHR_LB ||
 400         ctx->profile == FF_PROFILE_DNXHR_SQ ||
 401         ctx->profile == FF_PROFILE_DNXHR_HQ) && avctx->pix_fmt != AV_PIX_FMT_YUV422P) {
 402        av_log(avctx, AV_LOG_ERROR,
 403               "pixel format is incompatible with DNxHR LB/SQ/HQ profile\n");
 404        return AVERROR(EINVAL);
 405    }
 406
 407    ctx->is_444 = ctx->profile == FF_PROFILE_DNXHR_444;
 408    avctx->profile = ctx->profile;
 409    ctx->cid = ff_dnxhd_find_cid(avctx, ctx->bit_depth);
 410    if (!ctx->cid) {
 411        av_log(avctx, AV_LOG_ERROR,
 412               "video parameters incompatible with DNxHD. Valid DNxHD profiles:\n");
 413        ff_dnxhd_print_profiles(avctx, AV_LOG_ERROR);
 414        return AVERROR(EINVAL);
 415    }
 416    av_log(avctx, AV_LOG_DEBUG, "cid %d\n", ctx->cid);
 417
 418    if (ctx->cid >= 1270 && ctx->cid <= 1274)
 419        avctx->codec_tag = MKTAG('A','V','d','h');
 420
 421    if (avctx->width < 256 || avctx->height < 120) {
 422        av_log(avctx, AV_LOG_ERROR,
 423               "Input dimensions too small, input must be at least 256x120\n");
 424        return AVERROR(EINVAL);
 425    }
 426
 427    index = ff_dnxhd_get_cid_table(ctx->cid);
 428    av_assert0(index >= 0);
 429
 430    ctx->cid_table = &ff_dnxhd_cid_table[index];
 431
 432    ctx->m.avctx    = avctx;
 433    ctx->m.mb_intra = 1;
 434    ctx->m.h263_aic = 1;
 435
 436    avctx->bits_per_raw_sample = ctx->bit_depth;
 437
 438    ff_blockdsp_init(&ctx->bdsp, avctx);
 439    ff_fdctdsp_init(&ctx->m.fdsp, avctx);
 440    ff_mpv_idct_init(&ctx->m);
 441    ff_mpegvideoencdsp_init(&ctx->m.mpvencdsp, avctx);
 442    ff_pixblockdsp_init(&ctx->m.pdsp, avctx);
 443    ff_dct_encode_init(&ctx->m);
 444
 445    if (ctx->profile != FF_PROFILE_DNXHD)
 446        ff_videodsp_init(&ctx->m.vdsp, ctx->bit_depth);
 447
 448    if (!ctx->m.dct_quantize)
 449        ctx->m.dct_quantize = ff_dct_quantize_c;
 450
 451    if (ctx->is_444 || ctx->profile == FF_PROFILE_DNXHR_HQX) {
 452        ctx->m.dct_quantize     = dnxhd_10bit_dct_quantize_444;
 453        ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym;
 454        ctx->block_width_l2     = 4;
 455    } else if (ctx->bit_depth == 10) {
 456        ctx->m.dct_quantize     = dnxhd_10bit_dct_quantize;
 457        ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym;
 458        ctx->block_width_l2     = 4;
 459    } else {
 460        ctx->get_pixels_8x4_sym = dnxhd_8bit_get_pixels_8x4_sym;
 461        ctx->block_width_l2     = 3;
 462    }
 463
 464    if (ARCH_X86)
 465        ff_dnxhdenc_init_x86(ctx);
 466
 467    ctx->m.mb_height = (avctx->height + 15) / 16;
 468    ctx->m.mb_width  = (avctx->width  + 15) / 16;
 469
 470    if (avctx->flags & AV_CODEC_FLAG_INTERLACED_DCT) {
 471        ctx->interlaced   = 1;
 472        ctx->m.mb_height /= 2;
 473    }
 474
 475    if (ctx->interlaced && ctx->profile != FF_PROFILE_DNXHD) {
 476        av_log(avctx, AV_LOG_ERROR,
 477               "Interlaced encoding is not supported for DNxHR profiles.\n");
 478        return AVERROR(EINVAL);
 479    }
 480
 481    ctx->m.mb_num = ctx->m.mb_height * ctx->m.mb_width;
 482
 483    if (ctx->cid_table->frame_size == DNXHD_VARIABLE) {
 484        ctx->frame_size = avpriv_dnxhd_get_hr_frame_size(ctx->cid,
 485                                                     avctx->width, avctx->height);
 486        av_assert0(ctx->frame_size >= 0);
 487        ctx->coding_unit_size = ctx->frame_size;
 488    } else {
 489        ctx->frame_size = ctx->cid_table->frame_size;
 490        ctx->coding_unit_size = ctx->cid_table->coding_unit_size;
 491    }
 492
 493    if (ctx->m.mb_height > 68)
 494        ctx->data_offset = 0x170 + (ctx->m.mb_height << 2);
 495    else
 496        ctx->data_offset = 0x280;
 497
 498    // XXX tune lbias/cbias
 499    if ((ret = dnxhd_init_qmat(ctx, ctx->intra_quant_bias, 0)) < 0)
 500        return ret;
 501
 502    /* Avid Nitris hardware decoder requires a minimum amount of padding
 503     * in the coding unit payload */
 504    if (ctx->nitris_compat)
 505        ctx->min_padding = 1600;
 506
 507    if ((ret = dnxhd_init_vlc(ctx)) < 0)
 508        return ret;
 509    if ((ret = dnxhd_init_rc(ctx)) < 0)
 510        return ret;
 511
 512    FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_size,
 513                      ctx->m.mb_height * sizeof(uint32_t), fail);
 514    FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_offs,
 515                      ctx->m.mb_height * sizeof(uint32_t), fail);
 516    FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_bits,
 517                      ctx->m.mb_num * sizeof(uint16_t), fail);
 518    FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_qscale,
 519                      ctx->m.mb_num * sizeof(uint8_t), fail);
 520
 521#if FF_API_CODED_FRAME
 522FF_DISABLE_DEPRECATION_WARNINGS
 523    avctx->coded_frame->key_frame = 1;
 524    avctx->coded_frame->pict_type = AV_PICTURE_TYPE_I;
 525FF_ENABLE_DEPRECATION_WARNINGS
 526#endif
 527
 528    if (avctx->active_thread_type == FF_THREAD_SLICE) {
 529        if (avctx->thread_count > MAX_THREADS) {
 530            av_log(avctx, AV_LOG_ERROR, "too many threads\n");
 531            return AVERROR(EINVAL);
 532        }
 533    }
 534
 535    if (avctx->qmax <= 1) {
 536        av_log(avctx, AV_LOG_ERROR, "qmax must be at least 2\n");
 537        return AVERROR(EINVAL);
 538    }
 539
 540    ctx->thread[0] = ctx;
 541    if (avctx->active_thread_type == FF_THREAD_SLICE) {
 542        for (i = 1; i < avctx->thread_count; i++) {
 543            ctx->thread[i] = av_malloc(sizeof(DNXHDEncContext));
 544            if (!ctx->thread[i])
 545                goto fail;
 546            memcpy(ctx->thread[i], ctx, sizeof(DNXHDEncContext));
 547        }
 548    }
 549
 550    return 0;
 551fail:  // for FF_ALLOCZ_OR_GOTO
 552    return AVERROR(ENOMEM);
 553}
 554
 555static int dnxhd_write_header(AVCodecContext *avctx, uint8_t *buf)
 556{
 557    DNXHDEncContext *ctx = avctx->priv_data;
 558
 559    memset(buf, 0, ctx->data_offset);
 560
 561    // * write prefix */
 562    AV_WB16(buf + 0x02, ctx->data_offset);
 563    if (ctx->cid >= 1270 && ctx->cid <= 1274)
 564        buf[4] = 0x03;
 565    else
 566        buf[4] = 0x01;
 567
 568    buf[5] = ctx->interlaced ? ctx->cur_field + 2 : 0x01;
 569    buf[6] = 0x80; // crc flag off
 570    buf[7] = 0xa0; // reserved
 571    AV_WB16(buf + 0x18, avctx->height >> ctx->interlaced); // ALPF
 572    AV_WB16(buf + 0x1a, avctx->width);  // SPL
 573    AV_WB16(buf + 0x1d, avctx->height >> ctx->interlaced); // NAL
 574
 575    buf[0x21] = ctx->bit_depth == 10 ? 0x58 : 0x38;
 576    buf[0x22] = 0x88 + (ctx->interlaced << 2);
 577    AV_WB32(buf + 0x28, ctx->cid); // CID
 578    buf[0x2c] = (!ctx->interlaced << 7) | (ctx->is_444 << 6) | (avctx->pix_fmt == AV_PIX_FMT_YUV444P10);
 579
 580    buf[0x5f] = 0x01; // UDL
 581
 582    buf[0x167] = 0x02; // reserved
 583    AV_WB16(buf + 0x16a, ctx->m.mb_height * 4 + 4); // MSIPS
 584    AV_WB16(buf + 0x16c, ctx->m.mb_height); // Ns
 585    buf[0x16f] = 0x10; // reserved
 586
 587    ctx->msip = buf + 0x170;
 588    return 0;
 589}
 590
 591static av_always_inline void dnxhd_encode_dc(DNXHDEncContext *ctx, int diff)
 592{
 593    int nbits;
 594    if (diff < 0) {
 595        nbits = av_log2_16bit(-2 * diff);
 596        diff--;
 597    } else {
 598        nbits = av_log2_16bit(2 * diff);
 599    }
 600    put_bits(&ctx->m.pb, ctx->cid_table->dc_bits[nbits] + nbits,
 601             (ctx->cid_table->dc_codes[nbits] << nbits) +
 602             av_mod_uintp2(diff, nbits));
 603}
 604
 605static av_always_inline
 606void dnxhd_encode_block(DNXHDEncContext *ctx, int16_t *block,
 607                        int last_index, int n)
 608{
 609    int last_non_zero = 0;
 610    int slevel, i, j;
 611
 612    dnxhd_encode_dc(ctx, block[0] - ctx->m.last_dc[n]);
 613    ctx->m.last_dc[n] = block[0];
 614
 615    for (i = 1; i <= last_index; i++) {
 616        j = ctx->m.intra_scantable.permutated[i];
 617        slevel = block[j];
 618        if (slevel) {
 619            int run_level = i - last_non_zero - 1;
 620            int rlevel = slevel * (1 << 1) | !!run_level;
 621            put_bits(&ctx->m.pb, ctx->vlc_bits[rlevel], ctx->vlc_codes[rlevel]);
 622            if (run_level)
 623                put_bits(&ctx->m.pb, ctx->run_bits[run_level],
 624                         ctx->run_codes[run_level]);
 625            last_non_zero = i;
 626        }
 627    }
 628    put_bits(&ctx->m.pb, ctx->vlc_bits[0], ctx->vlc_codes[0]); // EOB
 629}
 630
 631static av_always_inline
 632void dnxhd_unquantize_c(DNXHDEncContext *ctx, int16_t *block, int n,
 633                        int qscale, int last_index)
 634{
 635    const uint8_t *weight_matrix;
 636    int level;
 637    int i;
 638
 639    if (ctx->is_444) {
 640        weight_matrix = ((n % 6) < 2) ? ctx->cid_table->luma_weight
 641                                      : ctx->cid_table->chroma_weight;
 642    } else {
 643        weight_matrix = (n & 2) ? ctx->cid_table->chroma_weight
 644                                : ctx->cid_table->luma_weight;
 645    }
 646
 647    for (i = 1; i <= last_index; i++) {
 648        int j = ctx->m.intra_scantable.permutated[i];
 649        level = block[j];
 650        if (level) {
 651            if (level < 0) {
 652                level = (1 - 2 * level) * qscale * weight_matrix[i];
 653                if (ctx->bit_depth == 10) {
 654                    if (weight_matrix[i] != 8)
 655                        level += 8;
 656                    level >>= 4;
 657                } else {
 658                    if (weight_matrix[i] != 32)
 659                        level += 32;
 660                    level >>= 6;
 661                }
 662                level = -level;
 663            } else {
 664                level = (2 * level + 1) * qscale * weight_matrix[i];
 665                if (ctx->bit_depth == 10) {
 666                    if (weight_matrix[i] != 8)
 667                        level += 8;
 668                    level >>= 4;
 669                } else {
 670                    if (weight_matrix[i] != 32)
 671                        level += 32;
 672                    level >>= 6;
 673                }
 674            }
 675            block[j] = level;
 676        }
 677    }
 678}
 679
 680static av_always_inline int dnxhd_ssd_block(int16_t *qblock, int16_t *block)
 681{
 682    int score = 0;
 683    int i;
 684    for (i = 0; i < 64; i++)
 685        score += (block[i] - qblock[i]) * (block[i] - qblock[i]);
 686    return score;
 687}
 688
 689static av_always_inline
 690int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, int16_t *block, int last_index)
 691{
 692    int last_non_zero = 0;
 693    int bits = 0;
 694    int i, j, level;
 695    for (i = 1; i <= last_index; i++) {
 696        j = ctx->m.intra_scantable.permutated[i];
 697        level = block[j];
 698        if (level) {
 699            int run_level = i - last_non_zero - 1;
 700            bits += ctx->vlc_bits[level * (1 << 1) |
 701                    !!run_level] + ctx->run_bits[run_level];
 702            last_non_zero = i;
 703        }
 704    }
 705    return bits;
 706}
 707
 708static av_always_inline
 709void dnxhd_get_blocks(DNXHDEncContext *ctx, int mb_x, int mb_y)
 710{
 711    const int bs = ctx->block_width_l2;
 712    const int bw = 1 << bs;
 713    int dct_y_offset = ctx->dct_y_offset;
 714    int dct_uv_offset = ctx->dct_uv_offset;
 715    int linesize = ctx->m.linesize;
 716    int uvlinesize = ctx->m.uvlinesize;
 717    const uint8_t *ptr_y = ctx->thread[0]->src[0] +
 718                           ((mb_y << 4) * ctx->m.linesize) + (mb_x << bs + 1);
 719    const uint8_t *ptr_u = ctx->thread[0]->src[1] +
 720                           ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs + ctx->is_444);
 721    const uint8_t *ptr_v = ctx->thread[0]->src[2] +
 722                           ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs + ctx->is_444);
 723    PixblockDSPContext *pdsp = &ctx->m.pdsp;
 724    VideoDSPContext *vdsp = &ctx->m.vdsp;
 725
 726    if (ctx->bit_depth != 10 && vdsp->emulated_edge_mc && ((mb_x << 4) + 16 > ctx->m.avctx->width ||
 727                                                           (mb_y << 4) + 16 > ctx->m.avctx->height)) {
 728        int y_w = ctx->m.avctx->width  - (mb_x << 4);
 729        int y_h = ctx->m.avctx->height - (mb_y << 4);
 730        int uv_w = (y_w + 1) / 2;
 731        int uv_h = y_h;
 732        linesize = 16;
 733        uvlinesize = 8;
 734
 735        vdsp->emulated_edge_mc(&ctx->edge_buf_y[0], ptr_y,
 736                               linesize, ctx->m.linesize,
 737                               linesize, 16,
 738                               0, 0, y_w, y_h);
 739        vdsp->emulated_edge_mc(&ctx->edge_buf_uv[0][0], ptr_u,
 740                               uvlinesize, ctx->m.uvlinesize,
 741                               uvlinesize, 16,
 742                               0, 0, uv_w, uv_h);
 743        vdsp->emulated_edge_mc(&ctx->edge_buf_uv[1][0], ptr_v,
 744                               uvlinesize, ctx->m.uvlinesize,
 745                               uvlinesize, 16,
 746                               0, 0, uv_w, uv_h);
 747
 748        dct_y_offset =  bw * linesize;
 749        dct_uv_offset = bw * uvlinesize;
 750        ptr_y = &ctx->edge_buf_y[0];
 751        ptr_u = &ctx->edge_buf_uv[0][0];
 752        ptr_v = &ctx->edge_buf_uv[1][0];
 753    } else if (ctx->bit_depth == 10 && vdsp->emulated_edge_mc && ((mb_x << 4) + 16 > ctx->m.avctx->width ||
 754                                                                  (mb_y << 4) + 16 > ctx->m.avctx->height)) {
 755        int y_w = ctx->m.avctx->width  - (mb_x << 4);
 756        int y_h = ctx->m.avctx->height - (mb_y << 4);
 757        int uv_w = ctx->is_444 ? y_w : (y_w + 1) / 2;
 758        int uv_h = y_h;
 759        linesize = 32;
 760        uvlinesize = 16 + 16 * ctx->is_444;
 761
 762        vdsp->emulated_edge_mc(&ctx->edge_buf_y[0], ptr_y,
 763                               linesize, ctx->m.linesize,
 764                               linesize / 2, 16,
 765                               0, 0, y_w, y_h);
 766        vdsp->emulated_edge_mc(&ctx->edge_buf_uv[0][0], ptr_u,
 767                               uvlinesize, ctx->m.uvlinesize,
 768                               uvlinesize / 2, 16,
 769                               0, 0, uv_w, uv_h);
 770        vdsp->emulated_edge_mc(&ctx->edge_buf_uv[1][0], ptr_v,
 771                               uvlinesize, ctx->m.uvlinesize,
 772                               uvlinesize / 2, 16,
 773                               0, 0, uv_w, uv_h);
 774
 775        dct_y_offset =  bw * linesize / 2;
 776        dct_uv_offset = bw * uvlinesize / 2;
 777        ptr_y = &ctx->edge_buf_y[0];
 778        ptr_u = &ctx->edge_buf_uv[0][0];
 779        ptr_v = &ctx->edge_buf_uv[1][0];
 780    }
 781
 782    if (!ctx->is_444) {
 783        pdsp->get_pixels(ctx->blocks[0], ptr_y,      linesize);
 784        pdsp->get_pixels(ctx->blocks[1], ptr_y + bw, linesize);
 785        pdsp->get_pixels(ctx->blocks[2], ptr_u,      uvlinesize);
 786        pdsp->get_pixels(ctx->blocks[3], ptr_v,      uvlinesize);
 787
 788        if (mb_y + 1 == ctx->m.mb_height && ctx->m.avctx->height == 1080) {
 789            if (ctx->interlaced) {
 790                ctx->get_pixels_8x4_sym(ctx->blocks[4],
 791                                        ptr_y + dct_y_offset,
 792                                        linesize);
 793                ctx->get_pixels_8x4_sym(ctx->blocks[5],
 794                                        ptr_y + dct_y_offset + bw,
 795                                        linesize);
 796                ctx->get_pixels_8x4_sym(ctx->blocks[6],
 797                                        ptr_u + dct_uv_offset,
 798                                        uvlinesize);
 799                ctx->get_pixels_8x4_sym(ctx->blocks[7],
 800                                        ptr_v + dct_uv_offset,
 801                                        uvlinesize);
 802            } else {
 803                ctx->bdsp.clear_block(ctx->blocks[4]);
 804                ctx->bdsp.clear_block(ctx->blocks[5]);
 805                ctx->bdsp.clear_block(ctx->blocks[6]);
 806                ctx->bdsp.clear_block(ctx->blocks[7]);
 807            }
 808        } else {
 809            pdsp->get_pixels(ctx->blocks[4],
 810                             ptr_y + dct_y_offset, linesize);
 811            pdsp->get_pixels(ctx->blocks[5],
 812                             ptr_y + dct_y_offset + bw, linesize);
 813            pdsp->get_pixels(ctx->blocks[6],
 814                             ptr_u + dct_uv_offset, uvlinesize);
 815            pdsp->get_pixels(ctx->blocks[7],
 816                             ptr_v + dct_uv_offset, uvlinesize);
 817        }
 818    } else {
 819        pdsp->get_pixels(ctx->blocks[0], ptr_y,      linesize);
 820        pdsp->get_pixels(ctx->blocks[1], ptr_y + bw, linesize);
 821        pdsp->get_pixels(ctx->blocks[6], ptr_y + dct_y_offset, linesize);
 822        pdsp->get_pixels(ctx->blocks[7], ptr_y + dct_y_offset + bw, linesize);
 823
 824        pdsp->get_pixels(ctx->blocks[2], ptr_u,      uvlinesize);
 825        pdsp->get_pixels(ctx->blocks[3], ptr_u + bw, uvlinesize);
 826        pdsp->get_pixels(ctx->blocks[8], ptr_u + dct_uv_offset, uvlinesize);
 827        pdsp->get_pixels(ctx->blocks[9], ptr_u + dct_uv_offset + bw, uvlinesize);
 828
 829        pdsp->get_pixels(ctx->blocks[4], ptr_v,      uvlinesize);
 830        pdsp->get_pixels(ctx->blocks[5], ptr_v + bw, uvlinesize);
 831        pdsp->get_pixels(ctx->blocks[10], ptr_v + dct_uv_offset, uvlinesize);
 832        pdsp->get_pixels(ctx->blocks[11], ptr_v + dct_uv_offset + bw, uvlinesize);
 833    }
 834}
 835
 836static av_always_inline
 837int dnxhd_switch_matrix(DNXHDEncContext *ctx, int i)
 838{
 839    int x;
 840
 841    if (ctx->is_444) {
 842        x = (i >> 1) % 3;
 843    } else {
 844        const static uint8_t component[8]={0,0,1,2,0,0,1,2};
 845        x = component[i];
 846    }
 847    return x;
 848}
 849
 850static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg,
 851                                  int jobnr, int threadnr)
 852{
 853    DNXHDEncContext *ctx = avctx->priv_data;
 854    int mb_y = jobnr, mb_x;
 855    int qscale = ctx->qscale;
 856    LOCAL_ALIGNED_16(int16_t, block, [64]);
 857    ctx = ctx->thread[threadnr];
 858
 859    ctx->m.last_dc[0] =
 860    ctx->m.last_dc[1] =
 861    ctx->m.last_dc[2] = 1 << (ctx->bit_depth + 2);
 862
 863    for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
 864        unsigned mb = mb_y * ctx->m.mb_width + mb_x;
 865        int ssd     = 0;
 866        int ac_bits = 0;
 867        int dc_bits = 0;
 868        int i;
 869
 870        dnxhd_get_blocks(ctx, mb_x, mb_y);
 871
 872        for (i = 0; i < 8 + 4 * ctx->is_444; i++) {
 873            int16_t *src_block = ctx->blocks[i];
 874            int overflow, nbits, diff, last_index;
 875            int n = dnxhd_switch_matrix(ctx, i);
 876
 877            memcpy(block, src_block, 64 * sizeof(*block));
 878            last_index = ctx->m.dct_quantize(&ctx->m, block,
 879                                             ctx->is_444 ? 4 * (n > 0): 4 & (2*i),
 880                                             qscale, &overflow);
 881            ac_bits   += dnxhd_calc_ac_bits(ctx, block, last_index);
 882
 883            diff = block[0] - ctx->m.last_dc[n];
 884            if (diff < 0)
 885                nbits = av_log2_16bit(-2 * diff);
 886            else
 887                nbits = av_log2_16bit(2 * diff);
 888
 889            av_assert1(nbits < ctx->bit_depth + 4);
 890            dc_bits += ctx->cid_table->dc_bits[nbits] + nbits;
 891
 892            ctx->m.last_dc[n] = block[0];
 893
 894            if (avctx->mb_decision == FF_MB_DECISION_RD || !RC_VARIANCE) {
 895                dnxhd_unquantize_c(ctx, block, i, qscale, last_index);
 896                ctx->m.idsp.idct(block);
 897                ssd += dnxhd_ssd_block(block, src_block);
 898            }
 899        }
 900        ctx->mb_rc[(qscale * ctx->m.mb_num) + mb].ssd  = ssd;
 901        ctx->mb_rc[(qscale * ctx->m.mb_num) + mb].bits = ac_bits + dc_bits + 12 +
 902                                     (1 + ctx->is_444) * 8 * ctx->vlc_bits[0];
 903    }
 904    return 0;
 905}
 906
 907static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg,
 908                               int jobnr, int threadnr)
 909{
 910    DNXHDEncContext *ctx = avctx->priv_data;
 911    int mb_y = jobnr, mb_x;
 912    ctx = ctx->thread[threadnr];
 913    init_put_bits(&ctx->m.pb, (uint8_t *)arg + ctx->data_offset + ctx->slice_offs[jobnr],
 914                  ctx->slice_size[jobnr]);
 915
 916    ctx->m.last_dc[0] =
 917    ctx->m.last_dc[1] =
 918    ctx->m.last_dc[2] = 1 << (ctx->bit_depth + 2);
 919    for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
 920        unsigned mb = mb_y * ctx->m.mb_width + mb_x;
 921        int qscale = ctx->mb_qscale[mb];
 922        int i;
 923
 924        put_bits(&ctx->m.pb, 11, qscale);
 925        put_bits(&ctx->m.pb, 1, avctx->pix_fmt == AV_PIX_FMT_YUV444P10);
 926
 927        dnxhd_get_blocks(ctx, mb_x, mb_y);
 928
 929        for (i = 0; i < 8 + 4 * ctx->is_444; i++) {
 930            int16_t *block = ctx->blocks[i];
 931            int overflow, n = dnxhd_switch_matrix(ctx, i);
 932            int last_index = ctx->m.dct_quantize(&ctx->m, block,
 933                                                 ctx->is_444 ? (((i >> 1) % 3) < 1 ? 0 : 4): 4 & (2*i),
 934                                                 qscale, &overflow);
 935
 936            dnxhd_encode_block(ctx, block, last_index, n);
 937        }
 938    }
 939    if (put_bits_count(&ctx->m.pb) & 31)
 940        put_bits(&ctx->m.pb, 32 - (put_bits_count(&ctx->m.pb) & 31), 0);
 941    flush_put_bits(&ctx->m.pb);
 942    return 0;
 943}
 944
 945static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx)
 946{
 947    int mb_y, mb_x;
 948    int offset = 0;
 949    for (mb_y = 0; mb_y < ctx->m.mb_height; mb_y++) {
 950        int thread_size;
 951        ctx->slice_offs[mb_y] = offset;
 952        ctx->slice_size[mb_y] = 0;
 953        for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
 954            unsigned mb = mb_y * ctx->m.mb_width + mb_x;
 955            ctx->slice_size[mb_y] += ctx->mb_bits[mb];
 956        }
 957        ctx->slice_size[mb_y]   = (ctx->slice_size[mb_y] + 31) & ~31;
 958        ctx->slice_size[mb_y] >>= 3;
 959        thread_size = ctx->slice_size[mb_y];
 960        offset += thread_size;
 961    }
 962}
 963
 964static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg,
 965                               int jobnr, int threadnr)
 966{
 967    DNXHDEncContext *ctx = avctx->priv_data;
 968    int mb_y = jobnr, mb_x, x, y;
 969    int partial_last_row = (mb_y == ctx->m.mb_height - 1) &&
 970                           ((avctx->height >> ctx->interlaced) & 0xF);
 971
 972    ctx = ctx->thread[threadnr];
 973    if (ctx->bit_depth == 8) {
 974        uint8_t *pix = ctx->thread[0]->src[0] + ((mb_y << 4) * ctx->m.linesize);
 975        for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x, pix += 16) {
 976            unsigned mb = mb_y * ctx->m.mb_width + mb_x;
 977            int sum;
 978            int varc;
 979
 980            if (!partial_last_row && mb_x * 16 <= avctx->width - 16 && (avctx->width % 16) == 0) {
 981                sum  = ctx->m.mpvencdsp.pix_sum(pix, ctx->m.linesize);
 982                varc = ctx->m.mpvencdsp.pix_norm1(pix, ctx->m.linesize);
 983            } else {
 984                int bw = FFMIN(avctx->width - 16 * mb_x, 16);
 985                int bh = FFMIN((avctx->height >> ctx->interlaced) - 16 * mb_y, 16);
 986                sum = varc = 0;
 987                for (y = 0; y < bh; y++) {
 988                    for (x = 0; x < bw; x++) {
 989                        uint8_t val = pix[x + y * ctx->m.linesize];
 990                        sum  += val;
 991                        varc += val * val;
 992                    }
 993                }
 994            }
 995            varc = (varc - (((unsigned) sum * sum) >> 8) + 128) >> 8;
 996
 997            ctx->mb_cmp[mb].value = varc;
 998            ctx->mb_cmp[mb].mb    = mb;
 999        }
1000    } else { // 10-bit
1001        const int linesize = ctx->m.linesize >> 1;
1002        for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x) {
1003            uint16_t *pix = (uint16_t *)ctx->thread[0]->src[0] +
1004                            ((mb_y << 4) * linesize) + (mb_x << 4);
1005            unsigned mb  = mb_y * ctx->m.mb_width + mb_x;
1006            int sum = 0;
1007            int sqsum = 0;
1008            int bw = FFMIN(avctx->width - 16 * mb_x, 16);
1009            int bh = FFMIN((avctx->height >> ctx->interlaced) - 16 * mb_y, 16);
1010            int mean, sqmean;
1011            int i, j;
1012            // Macroblocks are 16x16 pixels, unlike DCT blocks which are 8x8.
1013            for (i = 0; i < bh; ++i) {
1014                for (j = 0; j < bw; ++j) {
1015                    // Turn 16-bit pixels into 10-bit ones.
1016                    const int sample = (unsigned) pix[j] >> 6;
1017                    sum   += sample;
1018                    sqsum += sample * sample;
1019                    // 2^10 * 2^10 * 16 * 16 = 2^28, which is less than INT_MAX
1020                }
1021                pix += linesize;
1022            }
1023            mean = sum >> 8; // 16*16 == 2^8
1024            sqmean = sqsum >> 8;
1025            ctx->mb_cmp[mb].value = sqmean - mean * mean;
1026            ctx->mb_cmp[mb].mb    = mb;
1027        }
1028    }
1029    return 0;
1030}
1031
1032static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx)
1033{
1034    int lambda, up_step, down_step;
1035    int last_lower = INT_MAX, last_higher = 0;
1036    int x, y, q;
1037
1038    for (q = 1; q < avctx->qmax; q++) {
1039        ctx->qscale = q;
1040        avctx->execute2(avctx, dnxhd_calc_bits_thread,
1041                        NULL, NULL, ctx->m.mb_height);
1042    }
1043    up_step = down_step = 2 << LAMBDA_FRAC_BITS;
1044    lambda  = ctx->lambda;
1045
1046    for (;;) {
1047        int bits = 0;
1048        int end  = 0;
1049        if (lambda == last_higher) {
1050            lambda++;
1051            end = 1; // need to set final qscales/bits
1052        }
1053        for (y = 0; y < ctx->m.mb_height; y++) {
1054            for (x = 0; x < ctx->m.mb_width; x++) {
1055                unsigned min = UINT_MAX;
1056                int qscale = 1;
1057                int mb     = y * ctx->m.mb_width + x;
1058                int rc = 0;
1059                for (q = 1; q < avctx->qmax; q++) {
1060                    int i = (q*ctx->m.mb_num) + mb;
1061                    unsigned score = ctx->mb_rc[i].bits * lambda +
1062                                     ((unsigned) ctx->mb_rc[i].ssd << LAMBDA_FRAC_BITS);
1063                    if (score < min) {
1064                        min    = score;
1065                        qscale = q;
1066                        rc = i;
1067                    }
1068                }
1069                bits += ctx->mb_rc[rc].bits;
1070                ctx->mb_qscale[mb] = qscale;
1071                ctx->mb_bits[mb]   = ctx->mb_rc[rc].bits;
1072            }
1073            bits = (bits + 31) & ~31; // padding
1074            if (bits > ctx->frame_bits)
1075                break;
1076        }
1077        if (end) {
1078            if (bits > ctx->frame_bits)
1079                return AVERROR(EINVAL);
1080            break;
1081        }
1082        if (bits < ctx->frame_bits) {
1083            last_lower = FFMIN(lambda, last_lower);
1084            if (last_higher != 0)
1085                lambda = (lambda+last_higher)>>1;
1086            else
1087                lambda -= down_step;
1088            down_step = FFMIN((int64_t)down_step*5, INT_MAX);
1089            up_step = 1<<LAMBDA_FRAC_BITS;
1090            lambda = FFMAX(1, lambda);
1091            if (lambda == last_lower)
1092                break;
1093        } else {
1094            last_higher = FFMAX(lambda, last_higher);
1095            if (last_lower != INT_MAX)
1096                lambda = (lambda+last_lower)>>1;
1097            else if ((int64_t)lambda + up_step > INT_MAX)
1098                return AVERROR(EINVAL);
1099            else
1100                lambda += up_step;
1101            up_step = FFMIN((int64_t)up_step*5, INT_MAX);
1102            down_step = 1<<LAMBDA_FRAC_BITS;
1103        }
1104    }
1105    ctx->lambda = lambda;
1106    return 0;
1107}
1108
1109static int dnxhd_find_qscale(DNXHDEncContext *ctx)
1110{
1111    int bits = 0;
1112    int up_step = 1;
1113    int down_step = 1;
1114    int last_higher = 0;
1115    int last_lower = INT_MAX;
1116    int qscale;
1117    int x, y;
1118
1119    qscale = ctx->qscale;
1120    for (;;) {
1121        bits = 0;
1122        ctx->qscale = qscale;
1123        // XXX avoid recalculating bits
1124        ctx->m.avctx->execute2(ctx->m.avctx, dnxhd_calc_bits_thread,
1125                               NULL, NULL, ctx->m.mb_height);
1126        for (y = 0; y < ctx->m.mb_height; y++) {
1127            for (x = 0; x < ctx->m.mb_width; x++)
1128                bits += ctx->mb_rc[(qscale*ctx->m.mb_num) + (y*ctx->m.mb_width+x)].bits;
1129            bits = (bits+31)&~31; // padding
1130            if (bits > ctx->frame_bits)
1131                break;
1132        }
1133        if (bits < ctx->frame_bits) {
1134            if (qscale == 1)
1135                return 1;
1136            if (last_higher == qscale - 1) {
1137                qscale = last_higher;
1138                break;
1139            }
1140            last_lower = FFMIN(qscale, last_lower);
1141            if (last_higher != 0)
1142                qscale = (qscale + last_higher) >> 1;
1143            else
1144                qscale -= down_step++;
1145            if (qscale < 1)
1146                qscale = 1;
1147            up_step = 1;
1148        } else {
1149            if (last_lower == qscale + 1)
1150                break;
1151            last_higher = FFMAX(qscale, last_higher);
1152            if (last_lower != INT_MAX)
1153                qscale = (qscale + last_lower) >> 1;
1154            else
1155                qscale += up_step++;
1156            down_step = 1;
1157            if (qscale >= ctx->m.avctx->qmax)
1158                return AVERROR(EINVAL);
1159        }
1160    }
1161    ctx->qscale = qscale;
1162    return 0;
1163}
1164
1165#define BUCKET_BITS 8
1166#define RADIX_PASSES 4
1167#define NBUCKETS (1 << BUCKET_BITS)
1168
1169static inline int get_bucket(int value, int shift)
1170{
1171    value >>= shift;
1172    value  &= NBUCKETS - 1;
1173    return NBUCKETS - 1 - value;
1174}
1175
1176static void radix_count(const RCCMPEntry *data, int size,
1177                        int buckets[RADIX_PASSES][NBUCKETS])
1178{
1179    int i, j;
1180    memset(buckets, 0, sizeof(buckets[0][0]) * RADIX_PASSES * NBUCKETS);
1181    for (i = 0; i < size; i++) {
1182        int v = data[i].value;
1183        for (j = 0; j < RADIX_PASSES; j++) {
1184            buckets[j][get_bucket(v, 0)]++;
1185            v >>= BUCKET_BITS;
1186        }
1187        av_assert1(!v);
1188    }
1189    for (j = 0; j < RADIX_PASSES; j++) {
1190        int offset = size;
1191        for (i = NBUCKETS - 1; i >= 0; i--)
1192            buckets[j][i] = offset -= buckets[j][i];
1193        av_assert1(!buckets[j][0]);
1194    }
1195}
1196
1197static void radix_sort_pass(RCCMPEntry *dst, const RCCMPEntry *data,
1198                            int size, int buckets[NBUCKETS], int pass)
1199{
1200    int shift = pass * BUCKET_BITS;
1201    int i;
1202    for (i = 0; i < size; i++) {
1203        int v   = get_bucket(data[i].value, shift);
1204        int pos = buckets[v]++;
1205        dst[pos] = data[i];
1206    }
1207}
1208
1209static void radix_sort(RCCMPEntry *data, RCCMPEntry *tmp, int size)
1210{
1211    int buckets[RADIX_PASSES][NBUCKETS];
1212    radix_count(data, size, buckets);
1213    radix_sort_pass(tmp, data, size, buckets[0], 0);
1214    radix_sort_pass(data, tmp, size, buckets[1], 1);
1215    if (buckets[2][NBUCKETS - 1] || buckets[3][NBUCKETS - 1]) {
1216        radix_sort_pass(tmp, data, size, buckets[2], 2);
1217        radix_sort_pass(data, tmp, size, buckets[3], 3);
1218    }
1219}
1220
1221static int dnxhd_encode_fast(AVCodecContext *avctx, DNXHDEncContext *ctx)
1222{
1223    int max_bits = 0;
1224    int ret, x, y;
1225    if ((ret = dnxhd_find_qscale(ctx)) < 0)
1226        return ret;
1227    for (y = 0; y < ctx->m.mb_height; y++) {
1228        for (x = 0; x < ctx->m.mb_width; x++) {
1229            int mb = y * ctx->m.mb_width + x;
1230            int rc = (ctx->qscale * ctx->m.mb_num ) + mb;
1231            int delta_bits;
1232            ctx->mb_qscale[mb] = ctx->qscale;
1233            ctx->mb_bits[mb] = ctx->mb_rc[rc].bits;
1234            max_bits += ctx->mb_rc[rc].bits;
1235            if (!RC_VARIANCE) {
1236                delta_bits = ctx->mb_rc[rc].bits -
1237                             ctx->mb_rc[rc + ctx->m.mb_num].bits;
1238                ctx->mb_cmp[mb].mb = mb;
1239                ctx->mb_cmp[mb].value =
1240                    delta_bits ? ((ctx->mb_rc[rc].ssd -
1241                                   ctx->mb_rc[rc + ctx->m.mb_num].ssd) * 100) /
1242                                  delta_bits
1243                               : INT_MIN; // avoid increasing qscale
1244            }
1245        }
1246        max_bits += 31; // worst padding
1247    }
1248    if (!ret) {
1249        if (RC_VARIANCE)
1250            avctx->execute2(avctx, dnxhd_mb_var_thread,
1251                            NULL, NULL, ctx->m.mb_height);
1252        radix_sort(ctx->mb_cmp, ctx->mb_cmp_tmp, ctx->m.mb_num);
1253        for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) {
1254            int mb = ctx->mb_cmp[x].mb;
1255            int rc = (ctx->qscale * ctx->m.mb_num ) + mb;
1256            max_bits -= ctx->mb_rc[rc].bits -
1257                        ctx->mb_rc[rc + ctx->m.mb_num].bits;
1258            ctx->mb_qscale[mb] = ctx->qscale + 1;
1259            ctx->mb_bits[mb]   = ctx->mb_rc[rc + ctx->m.mb_num].bits;
1260        }
1261    }
1262    return 0;
1263}
1264
1265static void dnxhd_load_picture(DNXHDEncContext *ctx, const AVFrame *frame)
1266{
1267    int i;
1268
1269    for (i = 0; i < ctx->m.avctx->thread_count; i++) {
1270        ctx->thread[i]->m.linesize    = frame->linesize[0] << ctx->interlaced;
1271        ctx->thread[i]->m.uvlinesize  = frame->linesize[1] << ctx->interlaced;
1272        ctx->thread[i]->dct_y_offset  = ctx->m.linesize  *8;
1273        ctx->thread[i]->dct_uv_offset = ctx->m.uvlinesize*8;
1274    }
1275
1276#if FF_API_CODED_FRAME
1277FF_DISABLE_DEPRECATION_WARNINGS
1278    ctx->m.avctx->coded_frame->interlaced_frame = frame->interlaced_frame;
1279FF_ENABLE_DEPRECATION_WARNINGS
1280#endif
1281    ctx->cur_field = frame->interlaced_frame && !frame->top_field_first;
1282}
1283
1284static int dnxhd_encode_picture(AVCodecContext *avctx, AVPacket *pkt,
1285                                const AVFrame *frame, int *got_packet)
1286{
1287    DNXHDEncContext *ctx = avctx->priv_data;
1288    int first_field = 1;
1289    int offset, i, ret;
1290    uint8_t *buf;
1291
1292    if ((ret = ff_alloc_packet2(avctx, pkt, ctx->frame_size, 0)) < 0)
1293        return ret;
1294    buf = pkt->data;
1295
1296    dnxhd_load_picture(ctx, frame);
1297
1298encode_coding_unit:
1299    for (i = 0; i < 3; i++) {
1300        ctx->src[i] = frame->data[i];
1301        if (ctx->interlaced && ctx->cur_field)
1302            ctx->src[i] += frame->linesize[i];
1303    }
1304
1305    dnxhd_write_header(avctx, buf);
1306
1307    if (avctx->mb_decision == FF_MB_DECISION_RD)
1308        ret = dnxhd_encode_rdo(avctx, ctx);
1309    else
1310        ret = dnxhd_encode_fast(avctx, ctx);
1311    if (ret < 0) {
1312        av_log(avctx, AV_LOG_ERROR,
1313               "picture could not fit ratecontrol constraints, increase qmax\n");
1314        return ret;
1315    }
1316
1317    dnxhd_setup_threads_slices(ctx);
1318
1319    offset = 0;
1320    for (i = 0; i < ctx->m.mb_height; i++) {
1321        AV_WB32(ctx->msip + i * 4, offset);
1322        offset += ctx->slice_size[i];
1323        av_assert1(!(ctx->slice_size[i] & 3));
1324    }
1325
1326    avctx->execute2(avctx, dnxhd_encode_thread, buf, NULL, ctx->m.mb_height);
1327
1328    av_assert1(ctx->data_offset + offset + 4 <= ctx->coding_unit_size);
1329    memset(buf + ctx->data_offset + offset, 0,
1330           ctx->coding_unit_size - 4 - offset - ctx->data_offset);
1331
1332    AV_WB32(buf + ctx->coding_unit_size - 4, 0x600DC0DE); // EOF
1333
1334    if (ctx->interlaced && first_field) {
1335        first_field     = 0;
1336        ctx->cur_field ^= 1;
1337        buf            += ctx->coding_unit_size;
1338        goto encode_coding_unit;
1339    }
1340
1341#if FF_API_CODED_FRAME
1342FF_DISABLE_DEPRECATION_WARNINGS
1343    avctx->coded_frame->quality = ctx->qscale * FF_QP2LAMBDA;
1344FF_ENABLE_DEPRECATION_WARNINGS
1345#endif
1346
1347    ff_side_data_set_encoder_stats(pkt, ctx->qscale * FF_QP2LAMBDA, NULL, 0, AV_PICTURE_TYPE_I);
1348
1349    pkt->flags |= AV_PKT_FLAG_KEY;
1350    *got_packet = 1;
1351    return 0;
1352}
1353
1354static av_cold int dnxhd_encode_end(AVCodecContext *avctx)
1355{
1356    DNXHDEncContext *ctx = avctx->priv_data;
1357    int i;
1358
1359    av_freep(&ctx->orig_vlc_codes);
1360    av_freep(&ctx->orig_vlc_bits);
1361    av_freep(&ctx->run_codes);
1362    av_freep(&ctx->run_bits);
1363
1364    av_freep(&ctx->mb_bits);
1365    av_freep(&ctx->mb_qscale);
1366    av_freep(&ctx->mb_rc);
1367    av_freep(&ctx->mb_cmp);
1368    av_freep(&ctx->mb_cmp_tmp);
1369    av_freep(&ctx->slice_size);
1370    av_freep(&ctx->slice_offs);
1371
1372    av_freep(&ctx->qmatrix_c);
1373    av_freep(&ctx->qmatrix_l);
1374    av_freep(&ctx->qmatrix_c16);
1375    av_freep(&ctx->qmatrix_l16);
1376
1377    if 

Large files files are truncated, but you can click here to view the full file