libavcodec/dnxhdenc.c

   1 /*
   2  * VC3/DNxHD encoder
   3  * Copyright (c) 2007 Baptiste Coudurier <baptiste dot coudurier at smartjog dot com>
   4  *
   5  * VC-3 encoder funded by the British Broadcasting Corporation
   6  *
   7  * This file is part of FFmpeg.
   8  *
   9  * FFmpeg is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public
  11  * License as published by the Free Software Foundation; either
  12  * version 2.1 of the License, or (at your option) any later version.
  13  *
  14  * FFmpeg is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with FFmpeg; if not, write to the Free Software
  21  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  22  */
  23
  24 //#define DEBUG
  25 #define RC_VARIANCE 1 // use variance or ssd for fast rc
  26
  27 #include "avcodec.h"
  28 #include "dsputil.h"
  29 #include "mpegvideo.h"
  30 #include "dnxhdenc.h"
  31
  32 int dct_quantize_c(MpegEncContext *s, DCTELEM *block, int n, int qscale, int *overflow);
  33
  34 #define LAMBDA_FRAC_BITS 10
  35
  36 static av_always_inline void dnxhd_get_pixels_8x4(DCTELEM *restrict block, const uint8_t *pixels, int line_size)
  37 {
  38     int i;
  39     for (i = 0; i < 4; i++) {
  40         block[0] = pixels[0]; block[1] = pixels[1];
  41         block[2] = pixels[2]; block[3] = pixels[3];
  42         block[4] = pixels[4]; block[5] = pixels[5];
  43         block[6] = pixels[6]; block[7] = pixels[7];
  44         pixels += line_size;
  45         block += 8;
  46     }
  47     memcpy(block   , block- 8, sizeof(*block)*8);
  48     memcpy(block+ 8, block-16, sizeof(*block)*8);
  49     memcpy(block+16, block-24, sizeof(*block)*8);
  50     memcpy(block+24, block-32, sizeof(*block)*8);
  51 }
  52
  53 static int dnxhd_init_vlc(DNXHDEncContext *ctx)
  54 {
  55     int i, j, level, run;
  56     int max_level = 1<<(ctx->cid_table->bit_depth+2);
  57
  58     FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->vlc_codes, max_level*4*sizeof(*ctx->vlc_codes), fail);
  59     FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->vlc_bits , max_level*4*sizeof(*ctx->vlc_bits ), fail);
  60     FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_codes, 63*2                               , fail);
  61     FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_bits , 63                                 , fail);
  62
  63     ctx->vlc_codes += max_level*2;
  64     ctx->vlc_bits  += max_level*2;
  65     for (level = -max_level; level < max_level; level++) {
  66         for (run = 0; run < 2; run++) {
  67             int index = (level<<1)|run;
  68             int sign, offset = 0, alevel = level;
  69
  70             MASK_ABS(sign, alevel);
  71             if (alevel > 64) {
  72                 offset = (alevel-1)>>6;
  73                 alevel -= offset<<6;
  74             }
  75             for (j = 0; j < 257; j++) {
  76                 if (ctx->cid_table->ac_level[j] == alevel &&
  77                     (!offset || (ctx->cid_table->ac_index_flag[j] && offset)) &&
  78                     (!run    || (ctx->cid_table->ac_run_flag  [j] && run))) {
  79                     assert(!ctx->vlc_codes[index]);
  80                     if (alevel) {
  81                         ctx->vlc_codes[index] = (ctx->cid_table->ac_codes[j]<<1)|(sign&1);
  82                         ctx->vlc_bits [index] = ctx->cid_table->ac_bits[j]+1;
  83                     } else {
  84                         ctx->vlc_codes[index] = ctx->cid_table->ac_codes[j];
  85                         ctx->vlc_bits [index] = ctx->cid_table->ac_bits [j];
  86                     }
  87                     break;
  88                 }
  89             }
  90             assert(!alevel || j < 257);
  91             if (offset) {
  92                 ctx->vlc_codes[index] = (ctx->vlc_codes[index]<<ctx->cid_table->index_bits)|offset;
  93                 ctx->vlc_bits [index]+= ctx->cid_table->index_bits;
  94             }
  95         }
  96     }
  97     for (i = 0; i < 62; i++) {
  98         int run = ctx->cid_table->run[i];
  99         assert(run < 63);
 100         ctx->run_codes[run] = ctx->cid_table->run_codes[i];
 101         ctx->run_bits [run] = ctx->cid_table->run_bits[i];
 102     }
 103     return 0;
 104  fail:
 105     return -1;
 106 }
 107
 108 static int dnxhd_init_qmat(DNXHDEncContext *ctx, int lbias, int cbias)
 109 {
 110     // init first elem to 1 to avoid div by 0 in convert_matrix
 111     uint16_t weight_matrix[64] = {1,}; // convert_matrix needs uint16_t*
 112     int qscale, i;
 113
 114     FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l,   (ctx->m.avctx->qmax+1) * 64 *     sizeof(int)     , fail);
 115     FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c,   (ctx->m.avctx->qmax+1) * 64 *     sizeof(int)     , fail);
 116     FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l16, (ctx->m.avctx->qmax+1) * 64 * 2 * sizeof(uint16_t), fail);
 117     FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c16, (ctx->m.avctx->qmax+1) * 64 * 2 * sizeof(uint16_t), fail);
 118
 119     for (i = 1; i < 64; i++) {
 120         int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]];
 121         weight_matrix[j] = ctx->cid_table->luma_weight[i];
 122     }
 123     ff_convert_matrix(&ctx->m.dsp, ctx->qmatrix_l, ctx->qmatrix_l16, weight_matrix,
 124                       ctx->m.intra_quant_bias, 1, ctx->m.avctx->qmax, 1);
 125     for (i = 1; i < 64; i++) {
 126         int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]];
 127         weight_matrix[j] = ctx->cid_table->chroma_weight[i];
 128     }
 129     ff_convert_matrix(&ctx->m.dsp, ctx->qmatrix_c, ctx->qmatrix_c16, weight_matrix,
 130                       ctx->m.intra_quant_bias, 1, ctx->m.avctx->qmax, 1);
 131     for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
 132         for (i = 0; i < 64; i++) {
 133             ctx->qmatrix_l  [qscale]   [i] <<= 2; ctx->qmatrix_c  [qscale]   [i] <<= 2;
 134             ctx->qmatrix_l16[qscale][0][i] <<= 2; ctx->qmatrix_l16[qscale][1][i] <<= 2;
 135             ctx->qmatrix_c16[qscale][0][i] <<= 2; ctx->qmatrix_c16[qscale][1][i] <<= 2;
 136         }
 137     }
 138     return 0;
 139  fail:
 140     return -1;
 141 }
 142
 143 static int dnxhd_init_rc(DNXHDEncContext *ctx)
 144 {
 145     FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_rc, 8160*ctx->m.avctx->qmax*sizeof(RCEntry), fail);
 146     if (ctx->m.avctx->mb_decision != FF_MB_DECISION_RD)
 147         FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_cmp, ctx->m.mb_num*sizeof(RCCMPEntry), fail);
 148
 149     ctx->frame_bits = (ctx->cid_table->coding_unit_size - 640 - 4) * 8;
 150     ctx->qscale = 1;
 151     ctx->lambda = 2<<LAMBDA_FRAC_BITS; // qscale 2
 152     return 0;
 153  fail:
 154     return -1;
 155 }
 156
 157 static int dnxhd_encode_init(AVCodecContext *avctx)
 158 {
 159     DNXHDEncContext *ctx = avctx->priv_data;
 160     int i, index;
 161
 162     ctx->cid = ff_dnxhd_find_cid(avctx);
 163     if (!ctx->cid || avctx->pix_fmt != PIX_FMT_YUV422P) {
 164         av_log(avctx, AV_LOG_ERROR, "video parameters incompatible with DNxHD\n");
 165         return -1;
 166     }
 167     av_log(avctx, AV_LOG_DEBUG, "cid %d\n", ctx->cid);
 168
 169     index = ff_dnxhd_get_cid_table(ctx->cid);
 170     ctx->cid_table = &ff_dnxhd_cid_table[index];
 171
 172     ctx->m.avctx = avctx;
 173     ctx->m.mb_intra = 1;
 174     ctx->m.h263_aic = 1;
 175
 176     ctx->get_pixels_8x4_sym = dnxhd_get_pixels_8x4;
 177
 178     dsputil_init(&ctx->m.dsp, avctx);
 179     ff_dct_common_init(&ctx->m);
 180 #if HAVE_MMX
 181     ff_dnxhd_init_mmx(ctx);
 182 #endif
 183     if (!ctx->m.dct_quantize)
 184         ctx->m.dct_quantize = dct_quantize_c;
 185
 186     ctx->m.mb_height = (avctx->height + 15) / 16;
 187     ctx->m.mb_width  = (avctx->width  + 15) / 16;
 188
 189     if (avctx->flags & CODEC_FLAG_INTERLACED_DCT) {
 190         ctx->interlaced = 1;
 191         ctx->m.mb_height /= 2;
 192     }
 193
 194     ctx->m.mb_num = ctx->m.mb_height * ctx->m.mb_width;
 195
 196     if (avctx->intra_quant_bias != FF_DEFAULT_QUANT_BIAS)
 197         ctx->m.intra_quant_bias = avctx->intra_quant_bias;
 198     if (dnxhd_init_qmat(ctx, ctx->m.intra_quant_bias, 0) < 0) // XXX tune lbias/cbias
 199         return -1;
 200
 201     if (dnxhd_init_vlc(ctx) < 0)
 202         return -1;
 203     if (dnxhd_init_rc(ctx) < 0)
 204         return -1;
 205
 206     FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_size, ctx->m.mb_height*sizeof(uint32_t), fail);
 207     FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_offs, ctx->m.mb_height*sizeof(uint32_t), fail);
 208     FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_bits,    ctx->m.mb_num   *sizeof(uint16_t), fail);
 209     FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_qscale,  ctx->m.mb_num   *sizeof(uint8_t) , fail);
 210
 211     ctx->frame.key_frame = 1;
 212     ctx->frame.pict_type = FF_I_TYPE;
 213     ctx->m.avctx->coded_frame = &ctx->frame;
 214
 215     if (avctx->thread_count > MAX_THREADS) {
 216         av_log(avctx, AV_LOG_ERROR, "too many threads\n");
 217         return -1;
 218     }
 219
 220     ctx->thread[0] = ctx;
 221     for (i = 1; i < avctx->thread_count; i++) {
 222         ctx->thread[i] =  av_malloc(sizeof(DNXHDEncContext));
 223         memcpy(ctx->thread[i], ctx, sizeof(DNXHDEncContext));
 224     }
 225
 226     return 0;
 227  fail: //for FF_ALLOCZ_OR_GOTO
 228     return -1;
 229 }
 230
 231 static int dnxhd_write_header(AVCodecContext *avctx, uint8_t *buf)
 232 {
 233     DNXHDEncContext *ctx = avctx->priv_data;
 234     const uint8_t header_prefix[5] = { 0x00,0x00,0x02,0x80,0x01 };
 235
 236     memset(buf, 0, 640);
 237
 238     memcpy(buf, header_prefix, 5);
 239     buf[5] = ctx->interlaced ? ctx->cur_field+2 : 0x01;
 240     buf[6] = 0x80; // crc flag off
 241     buf[7] = 0xa0; // reserved
 242     AV_WB16(buf + 0x18, avctx->height); // ALPF
 243     AV_WB16(buf + 0x1a, avctx->width);  // SPL
 244     AV_WB16(buf + 0x1d, avctx->height); // NAL
 245
 246     buf[0x21] = 0x38; // FIXME 8 bit per comp
 247     buf[0x22] = 0x88 + (ctx->frame.interlaced_frame<<2);
 248     AV_WB32(buf + 0x28, ctx->cid); // CID
 249     buf[0x2c] = ctx->interlaced ? 0 : 0x80;
 250
 251     buf[0x5f] = 0x01; // UDL
 252
 253     buf[0x167] = 0x02; // reserved
 254     AV_WB16(buf + 0x16a, ctx->m.mb_height * 4 + 4); // MSIPS
 255     buf[0x16d] = ctx->m.mb_height; // Ns
 256     buf[0x16f] = 0x10; // reserved
 257
 258     ctx->msip = buf + 0x170;
 259     return 0;
 260 }
 261
 262 static av_always_inline void dnxhd_encode_dc(DNXHDEncContext *ctx, int diff)
 263 {
 264     int nbits;
 265     if (diff < 0) {
 266         nbits = av_log2_16bit(-2*diff);
 267         diff--;
 268     } else {
 269         nbits = av_log2_16bit(2*diff);
 270     }
 271     put_bits(&ctx->m.pb, ctx->cid_table->dc_bits[nbits] + nbits,
 272              (ctx->cid_table->dc_codes[nbits]<<nbits) + (diff & ((1 << nbits) - 1)));
 273 }
 274
 275 static av_always_inline void dnxhd_encode_block(DNXHDEncContext *ctx, DCTELEM *block, int last_index, int n)
 276 {
 277     int last_non_zero = 0;
 278     int slevel, i, j;
 279
 280     dnxhd_encode_dc(ctx, block[0] - ctx->m.last_dc[n]);
 281     ctx->m.last_dc[n] = block[0];
 282
 283     for (i = 1; i <= last_index; i++) {
 284         j = ctx->m.intra_scantable.permutated[i];
 285         slevel = block[j];
 286         if (slevel) {
 287             int run_level = i - last_non_zero - 1;
 288             int rlevel = (slevel<<1)|!!run_level;
 289             put_bits(&ctx->m.pb, ctx->vlc_bits[rlevel], ctx->vlc_codes[rlevel]);
 290             if (run_level)
 291                 put_bits(&ctx->m.pb, ctx->run_bits[run_level], ctx->run_codes[run_level]);
 292             last_non_zero = i;
 293         }
 294     }
 295     put_bits(&ctx->m.pb, ctx->vlc_bits[0], ctx->vlc_codes[0]); // EOB
 296 }
 297
 298 static av_always_inline void dnxhd_unquantize_c(DNXHDEncContext *ctx, DCTELEM *block, int n, int qscale, int last_index)
 299 {
 300     const uint8_t *weight_matrix;
 301     int level;
 302     int i;
 303
 304     weight_matrix = (n&2) ? ctx->cid_table->chroma_weight : ctx->cid_table->luma_weight;
 305
 306     for (i = 1; i <= last_index; i++) {
 307         int j = ctx->m.intra_scantable.permutated[i];
 308         level = block[j];
 309         if (level) {
 310             if (level < 0) {
 311                 level = (1-2*level) * qscale * weight_matrix[i];
 312                 if (weight_matrix[i] != 32)
 313                     level += 32;
 314                 level >>= 6;
 315                 level = -level;
 316             } else {
 317                 level = (2*level+1) * qscale * weight_matrix[i];
 318                 if (weight_matrix[i] != 32)
 319                     level += 32;
 320                 level >>= 6;
 321             }
 322             block[j] = level;
 323         }
 324     }
 325 }
 326
 327 static av_always_inline int dnxhd_ssd_block(DCTELEM *qblock, DCTELEM *block)
 328 {
 329     int score = 0;
 330     int i;
 331     for (i = 0; i < 64; i++)
 332         score += (block[i]-qblock[i])*(block[i]-qblock[i]);
 333     return score;
 334 }
 335
 336 static av_always_inline int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, DCTELEM *block, int last_index)
 337 {
 338     int last_non_zero = 0;
 339     int bits = 0;
 340     int i, j, level;
 341     for (i = 1; i <= last_index; i++) {
 342         j = ctx->m.intra_scantable.permutated[i];
 343         level = block[j];
 344         if (level) {
 345             int run_level = i - last_non_zero - 1;
 346             bits += ctx->vlc_bits[(level<<1)|!!run_level]+ctx->run_bits[run_level];
 347             last_non_zero = i;
 348         }
 349     }
 350     return bits;
 351 }
 352
 353 static av_always_inline void dnxhd_get_blocks(DNXHDEncContext *ctx, int mb_x, int mb_y)
 354 {
 355     const uint8_t *ptr_y = ctx->thread[0]->src[0] + ((mb_y << 4) * ctx->m.linesize)   + (mb_x << 4);
 356     const uint8_t *ptr_u = ctx->thread[0]->src[1] + ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << 3);
 357     const uint8_t *ptr_v = ctx->thread[0]->src[2] + ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << 3);
 358     DSPContext *dsp = &ctx->m.dsp;
 359
 360     dsp->get_pixels(ctx->blocks[0], ptr_y    , ctx->m.linesize);
 361     dsp->get_pixels(ctx->blocks[1], ptr_y + 8, ctx->m.linesize);
 362     dsp->get_pixels(ctx->blocks[2], ptr_u    , ctx->m.uvlinesize);
 363     dsp->get_pixels(ctx->blocks[3], ptr_v    , ctx->m.uvlinesize);
 364
 365     if (mb_y+1 == ctx->m.mb_height && ctx->m.avctx->height == 1080) {
 366         if (ctx->interlaced) {
 367             ctx->get_pixels_8x4_sym(ctx->blocks[4], ptr_y + ctx->dct_y_offset    , ctx->m.linesize);
 368             ctx->get_pixels_8x4_sym(ctx->blocks[5], ptr_y + ctx->dct_y_offset + 8, ctx->m.linesize);
 369             ctx->get_pixels_8x4_sym(ctx->blocks[6], ptr_u + ctx->dct_uv_offset   , ctx->m.uvlinesize);
 370             ctx->get_pixels_8x4_sym(ctx->blocks[7], ptr_v + ctx->dct_uv_offset   , ctx->m.uvlinesize);
 371         } else {
 372             dsp->clear_block(ctx->blocks[4]); dsp->clear_block(ctx->blocks[5]);
 373             dsp->clear_block(ctx->blocks[6]); dsp->clear_block(ctx->blocks[7]);
 374         }
 375     } else {
 376         dsp->get_pixels(ctx->blocks[4], ptr_y + ctx->dct_y_offset    , ctx->m.linesize);
 377         dsp->get_pixels(ctx->blocks[5], ptr_y + ctx->dct_y_offset + 8, ctx->m.linesize);
 378         dsp->get_pixels(ctx->blocks[6], ptr_u + ctx->dct_uv_offset   , ctx->m.uvlinesize);
 379         dsp->get_pixels(ctx->blocks[7], ptr_v + ctx->dct_uv_offset   , ctx->m.uvlinesize);
 380     }
 381 }
 382
 383 static av_always_inline int dnxhd_switch_matrix(DNXHDEncContext *ctx, int i)
 384 {
 385     if (i&2) {
 386         ctx->m.q_intra_matrix16 = ctx->qmatrix_c16;
 387         ctx->m.q_intra_matrix   = ctx->qmatrix_c;
 388         return 1 + (i&1);
 389     } else {
 390         ctx->m.q_intra_matrix16 = ctx->qmatrix_l16;
 391         ctx->m.q_intra_matrix   = ctx->qmatrix_l;
 392         return 0;
 393     }
 394 }
 395
 396 static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg, int jobnr, int threadnr)
 397 {
 398     DNXHDEncContext *ctx = avctx->priv_data;
 399     int mb_y = jobnr, mb_x;
 400     int qscale = ctx->qscale;
 401     ctx = ctx->thread[threadnr];
 402
 403     ctx->m.last_dc[0] =
 404     ctx->m.last_dc[1] =
 405     ctx->m.last_dc[2] = 1024;
 406
 407     for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
 408         unsigned mb = mb_y * ctx->m.mb_width + mb_x;
 409         int ssd     = 0;
 410         int ac_bits = 0;
 411         int dc_bits = 0;
 412         int i;
 413
 414         dnxhd_get_blocks(ctx, mb_x, mb_y);
 415
 416         for (i = 0; i < 8; i++) {
 417             DECLARE_ALIGNED_16(DCTELEM, block[64]);
 418             DCTELEM *src_block = ctx->blocks[i];
 419             int overflow, nbits, diff, last_index;
 420             int n = dnxhd_switch_matrix(ctx, i);
 421
 422             memcpy(block, src_block, sizeof(block));
 423             last_index = ctx->m.dct_quantize((MpegEncContext*)ctx, block, i, qscale, &overflow);
 424             ac_bits += dnxhd_calc_ac_bits(ctx, block, last_index);
 425
 426             diff = block[0] - ctx->m.last_dc[n];
 427             if (diff < 0) nbits = av_log2_16bit(-2*diff);
 428             else          nbits = av_log2_16bit( 2*diff);
 429             dc_bits += ctx->cid_table->dc_bits[nbits] + nbits;
 430
 431             ctx->m.last_dc[n] = block[0];
 432
 433             if (avctx->mb_decision == FF_MB_DECISION_RD || !RC_VARIANCE) {
 434                 dnxhd_unquantize_c(ctx, block, i, qscale, last_index);
 435                 ctx->m.dsp.idct(block);
 436                 ssd += dnxhd_ssd_block(block, src_block);
 437             }
 438         }
 439         ctx->mb_rc[qscale][mb].ssd = ssd;
 440         ctx->mb_rc[qscale][mb].bits = ac_bits+dc_bits+12+8*ctx->vlc_bits[0];
 441     }
 442     return 0;
 443 }
 444
 445 static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg, int jobnr, int threadnr)
 446 {
 447     DNXHDEncContext *ctx = avctx->priv_data;
 448     int mb_y = jobnr, mb_x;
 449     ctx = ctx->thread[threadnr];
 450     init_put_bits(&ctx->m.pb, (uint8_t *)arg + 640 + ctx->slice_offs[jobnr], ctx->slice_size[jobnr]);
 451
 452     ctx->m.last_dc[0] =
 453     ctx->m.last_dc[1] =
 454     ctx->m.last_dc[2] = 1024;
 455     for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
 456         unsigned mb = mb_y * ctx->m.mb_width + mb_x;
 457         int qscale = ctx->mb_qscale[mb];
 458         int i;
 459
 460         put_bits(&ctx->m.pb, 12, qscale<<1);
 461
 462         dnxhd_get_blocks(ctx, mb_x, mb_y);
 463
 464         for (i = 0; i < 8; i++) {
 465             DCTELEM *block = ctx->blocks[i];
 466             int last_index, overflow;
 467             int n = dnxhd_switch_matrix(ctx, i);
 468             last_index = ctx->m.dct_quantize((MpegEncContext*)ctx, block, i, qscale, &overflow);
 469             //START_TIMER;
 470             dnxhd_encode_block(ctx, block, last_index, n);
 471             //STOP_TIMER("encode_block");
 472         }
 473     }
 474     if (put_bits_count(&ctx->m.pb)&31)
 475         put_bits(&ctx->m.pb, 32-(put_bits_count(&ctx->m.pb)&31), 0);
 476     flush_put_bits(&ctx->m.pb);
 477     return 0;
 478 }
 479
 480 static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx)
 481 {
 482     int mb_y, mb_x;
 483     int offset = 0;
 484     for (mb_y = 0; mb_y < ctx->m.mb_height; mb_y++) {
 485         int thread_size;
 486         ctx->slice_offs[mb_y] = offset;
 487             ctx->slice_size[mb_y] = 0;
 488             for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
 489                 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
 490                 ctx->slice_size[mb_y] += ctx->mb_bits[mb];
 491             }
 492             ctx->slice_size[mb_y] = (ctx->slice_size[mb_y]+31)&~31;
 493             ctx->slice_size[mb_y] >>= 3;
 494             thread_size = ctx->slice_size[mb_y];
 495         offset += thread_size;
 496     }
 497 }
 498
 499 static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg, int jobnr, int threadnr)
 500 {
 501     DNXHDEncContext *ctx = avctx->priv_data;
 502     int mb_y = jobnr, mb_x;
 503     ctx = ctx->thread[threadnr];
 504     for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
 505         unsigned mb  = mb_y * ctx->m.mb_width + mb_x;
 506         uint8_t *pix = ctx->thread[0]->src[0] + ((mb_y<<4) * ctx->m.linesize) + (mb_x<<4);
 507         int sum      = ctx->m.dsp.pix_sum(pix, ctx->m.linesize);
 508         int varc     = (ctx->m.dsp.pix_norm1(pix, ctx->m.linesize) - (((unsigned)(sum*sum))>>8)+128)>>8;
 509         ctx->mb_cmp[mb].value = varc;
 510         ctx->mb_cmp[mb].mb = mb;
 511     }
 512     return 0;
 513 }
 514
 515 static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx)
 516 {
 517     int lambda, up_step, down_step;
 518     int last_lower = INT_MAX, last_higher = 0;
 519     int x, y, q;
 520
 521     for (q = 1; q < avctx->qmax; q++) {
 522         ctx->qscale = q;
 523         avctx->execute2(avctx, dnxhd_calc_bits_thread, NULL, NULL, ctx->m.mb_height);
 524     }
 525     up_step = down_step = 2<<LAMBDA_FRAC_BITS;
 526     lambda = ctx->lambda;
 527
 528     for (;;) {
 529         int bits = 0;
 530         int end = 0;
 531         if (lambda == last_higher) {
 532             lambda++;
 533             end = 1; // need to set final qscales/bits
 534         }
 535         for (y = 0; y < ctx->m.mb_height; y++) {
 536             for (x = 0; x < ctx->m.mb_width; x++) {
 537                 unsigned min = UINT_MAX;
 538                 int qscale = 1;
 539                 int mb = y*ctx->m.mb_width+x;
 540                 for (q = 1; q < avctx->qmax; q++) {
 541                     unsigned score = ctx->mb_rc[q][mb].bits*lambda+(ctx->mb_rc[q][mb].ssd<<LAMBDA_FRAC_BITS);
 542                     if (score < min) {
 543                         min = score;
 544                         qscale = q;
 545                     }
 546                 }
 547                 bits += ctx->mb_rc[qscale][mb].bits;
 548                 ctx->mb_qscale[mb] = qscale;
 549                 ctx->mb_bits[mb] = ctx->mb_rc[qscale][mb].bits;
 550             }
 551             bits = (bits+31)&~31; // padding
 552             if (bits > ctx->frame_bits)
 553                 break;
 554         }
 555         //dprintf(ctx->m.avctx, "lambda %d, up %u, down %u, bits %d, frame %d\n",
 556         //        lambda, last_higher, last_lower, bits, ctx->frame_bits);
 557         if (end) {
 558             if (bits > ctx->frame_bits)
 559                 return -1;
 560             break;
 561         }
 562         if (bits < ctx->frame_bits) {
 563             last_lower = FFMIN(lambda, last_lower);
 564             if (last_higher != 0)
 565                 lambda = (lambda+last_higher)>>1;
 566             else
 567                 lambda -= down_step;
 568             down_step *= 5; // XXX tune ?
 569             up_step = 1<<LAMBDA_FRAC_BITS;
 570             lambda = FFMAX(1, lambda);
 571             if (lambda == last_lower)
 572                 break;
 573         } else {
 574             last_higher = FFMAX(lambda, last_higher);
 575             if (last_lower != INT_MAX)
 576                 lambda = (lambda+last_lower)>>1;
 577             else
 578                 lambda += up_step;
 579             up_step *= 5;
 580             down_step = 1<<LAMBDA_FRAC_BITS;
 581         }
 582     }
 583     //dprintf(ctx->m.avctx, "out lambda %d\n", lambda);
 584     ctx->lambda = lambda;
 585     return 0;
 586 }
 587
 588 static int dnxhd_find_qscale(DNXHDEncContext *ctx)
 589 {
 590     int bits = 0;
 591     int up_step = 1;
 592     int down_step = 1;
 593     int last_higher = 0;
 594     int last_lower = INT_MAX;
 595     int qscale;
 596     int x, y;
 597
 598     qscale = ctx->qscale;
 599     for (;;) {
 600         bits = 0;
 601         ctx->qscale = qscale;
 602         // XXX avoid recalculating bits
 603         ctx->m.avctx->execute2(ctx->m.avctx, dnxhd_calc_bits_thread, NULL, NULL, ctx->m.mb_height);
 604         for (y = 0; y < ctx->m.mb_height; y++) {
 605             for (x = 0; x < ctx->m.mb_width; x++)
 606                 bits += ctx->mb_rc[qscale][y*ctx->m.mb_width+x].bits;
 607             bits = (bits+31)&~31; // padding
 608             if (bits > ctx->frame_bits)
 609                 break;
 610         }
 611         //dprintf(ctx->m.avctx, "%d, qscale %d, bits %d, frame %d, higher %d, lower %d\n",
 612         //        ctx->m.avctx->frame_number, qscale, bits, ctx->frame_bits, last_higher, last_lower);
 613         if (bits < ctx->frame_bits) {
 614             if (qscale == 1)
 615                 return 1;
 616             if (last_higher == qscale - 1) {
 617                 qscale = last_higher;
 618                 break;
 619             }
 620             last_lower = FFMIN(qscale, last_lower);
 621             if (last_higher != 0)
 622                 qscale = (qscale+last_higher)>>1;
 623             else
 624                 qscale -= down_step++;
 625             if (qscale < 1)
 626                 qscale = 1;
 627             up_step = 1;
 628         } else {
 629             if (last_lower == qscale + 1)
 630                 break;
 631             last_higher = FFMAX(qscale, last_higher);
 632             if (last_lower != INT_MAX)
 633                 qscale = (qscale+last_lower)>>1;
 634             else
 635                 qscale += up_step++;
 636             down_step = 1;
 637             if (qscale >= ctx->m.avctx->qmax)
 638                 return -1;
 639         }
 640     }
 641     //dprintf(ctx->m.avctx, "out qscale %d\n", qscale);
 642     ctx->qscale = qscale;
 643     return 0;
 644 }
 645
 646 #define BUCKET_BITS 8
 647 #define RADIX_PASSES 4
 648 #define NBUCKETS (1 << BUCKET_BITS)
 649
 650 static inline int get_bucket(int value, int shift)
 651 {
 652     value >>= shift;
 653     value &= NBUCKETS - 1;
 654     return NBUCKETS - 1 - value;
 655 }
 656
 657 static void radix_count(const RCCMPEntry *data, int size, int buckets[RADIX_PASSES][NBUCKETS])
 658 {
 659     int i, j;
 660     memset(buckets, 0, sizeof(buckets[0][0]) * RADIX_PASSES * NBUCKETS);
 661     for (i = 0; i < size; i++) {
 662         int v = data[i].value;
 663         for (j = 0; j < RADIX_PASSES; j++) {
 664             buckets[j][get_bucket(v, 0)]++;
 665             v >>= BUCKET_BITS;
 666         }
 667         assert(!v);
 668     }
 669     for (j = 0; j < RADIX_PASSES; j++) {
 670         int offset = size;
 671         for (i = NBUCKETS - 1; i >= 0; i--)
 672             buckets[j][i] = offset -= buckets[j][i];
 673         assert(!buckets[j][0]);
 674     }
 675 }
 676
 677 static void radix_sort_pass(RCCMPEntry *dst, const RCCMPEntry *data, int size, int buckets[NBUCKETS], int pass)
 678 {
 679     int shift = pass * BUCKET_BITS;
 680     int i;
 681     for (i = 0; i < size; i++) {
 682         int v = get_bucket(data[i].value, shift);
 683         int pos = buckets[v]++;
 684         dst[pos] = data[i];
 685     }
 686 }
 687
 688 static void radix_sort(RCCMPEntry *data, int size)
 689 {
 690     int buckets[RADIX_PASSES][NBUCKETS];
 691     RCCMPEntry *tmp = av_malloc(sizeof(*tmp) * size);
 692     radix_count(data, size, buckets);
 693     radix_sort_pass(tmp, data, size, buckets[0], 0);
 694     radix_sort_pass(data, tmp, size, buckets[1], 1);
 695     if (buckets[2][NBUCKETS - 1] || buckets[3][NBUCKETS - 1]) {
 696         radix_sort_pass(tmp, data, size, buckets[2], 2);
 697         radix_sort_pass(data, tmp, size, buckets[3], 3);
 698     }
 699     av_free(tmp);
 700 }
 701
 702 static int dnxhd_encode_fast(AVCodecContext *avctx, DNXHDEncContext *ctx)
 703 {
 704     int max_bits = 0;
 705     int ret, x, y;
 706     if ((ret = dnxhd_find_qscale(ctx)) < 0)
 707         return -1;
 708     for (y = 0; y < ctx->m.mb_height; y++) {
 709         for (x = 0; x < ctx->m.mb_width; x++) {
 710             int mb = y*ctx->m.mb_width+x;
 711             int delta_bits;
 712             ctx->mb_qscale[mb] = ctx->qscale;
 713             ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale][mb].bits;
 714             max_bits += ctx->mb_rc[ctx->qscale][mb].bits;
 715             if (!RC_VARIANCE) {
 716                 delta_bits = ctx->mb_rc[ctx->qscale][mb].bits-ctx->mb_rc[ctx->qscale+1][mb].bits;
 717                 ctx->mb_cmp[mb].mb = mb;
 718                 ctx->mb_cmp[mb].value = delta_bits ?
 719                     ((ctx->mb_rc[ctx->qscale][mb].ssd-ctx->mb_rc[ctx->qscale+1][mb].ssd)*100)/delta_bits
 720                     : INT_MIN; //avoid increasing qscale
 721             }
 722         }
 723         max_bits += 31; //worst padding
 724     }
 725     if (!ret) {
 726         if (RC_VARIANCE)
 727             avctx->execute2(avctx, dnxhd_mb_var_thread, NULL, NULL, ctx->m.mb_height);
 728         radix_sort(ctx->mb_cmp, ctx->m.mb_num);
 729         for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) {
 730             int mb = ctx->mb_cmp[x].mb;
 731             max_bits -= ctx->mb_rc[ctx->qscale][mb].bits - ctx->mb_rc[ctx->qscale+1][mb].bits;
 732             ctx->mb_qscale[mb] = ctx->qscale+1;
 733             ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale+1][mb].bits;
 734         }
 735     }
 736     return 0;
 737 }
 738
 739 static void dnxhd_load_picture(DNXHDEncContext *ctx, const AVFrame *frame)
 740 {
 741     int i;
 742
 743     for (i = 0; i < 3; i++) {
 744         ctx->frame.data[i]     = frame->data[i];
 745         ctx->frame.linesize[i] = frame->linesize[i];
 746     }
 747
 748     for (i = 0; i < ctx->m.avctx->thread_count; i++) {
 749         ctx->thread[i]->m.linesize    = ctx->frame.linesize[0]<<ctx->interlaced;
 750         ctx->thread[i]->m.uvlinesize  = ctx->frame.linesize[1]<<ctx->interlaced;
 751         ctx->thread[i]->dct_y_offset  = ctx->m.linesize  *8;
 752         ctx->thread[i]->dct_uv_offset = ctx->m.uvlinesize*8;
 753     }
 754
 755     ctx->frame.interlaced_frame = frame->interlaced_frame;
 756     ctx->cur_field = frame->interlaced_frame && !frame->top_field_first;
 757 }
 758
 759 static int dnxhd_encode_picture(AVCodecContext *avctx, unsigned char *buf, int buf_size, void *data)
 760 {
 761     DNXHDEncContext *ctx = avctx->priv_data;
 762     int first_field = 1;
 763     int offset, i, ret;
 764
 765     if (buf_size < ctx->cid_table->frame_size) {
 766         av_log(avctx, AV_LOG_ERROR, "output buffer is too small to compress picture\n");
 767         return -1;
 768     }
 769
 770     dnxhd_load_picture(ctx, data);
 771
 772  encode_coding_unit:
 773     for (i = 0; i < 3; i++) {
 774         ctx->src[i] = ctx->frame.data[i];
 775         if (ctx->interlaced && ctx->cur_field)
 776             ctx->src[i] += ctx->frame.linesize[i];
 777     }
 778
 779     dnxhd_write_header(avctx, buf);
 780
 781     if (avctx->mb_decision == FF_MB_DECISION_RD)
 782         ret = dnxhd_encode_rdo(avctx, ctx);
 783     else
 784         ret = dnxhd_encode_fast(avctx, ctx);
 785     if (ret < 0) {
 786         av_log(avctx, AV_LOG_ERROR, "picture could not fit ratecontrol constraints\n");
 787         return -1;
 788     }
 789
 790     dnxhd_setup_threads_slices(ctx);
 791
 792     offset = 0;
 793     for (i = 0; i < ctx->m.mb_height; i++) {
 794         AV_WB32(ctx->msip + i * 4, offset);
 795         offset += ctx->slice_size[i];
 796         assert(!(ctx->slice_size[i] & 3));
 797     }
 798
 799     avctx->execute2(avctx, dnxhd_encode_thread, buf, NULL, ctx->m.mb_height);
 800
 801     assert(640 + offset + 4 <= ctx->cid_table->coding_unit_size);
 802     memset(buf + 640 + offset, 0, ctx->cid_table->coding_unit_size - 4 - offset - 640);
 803
 804     AV_WB32(buf + ctx->cid_table->coding_unit_size - 4, 0x600DC0DE); // EOF
 805
 806     if (ctx->interlaced && first_field) {
 807         first_field     = 0;
 808         ctx->cur_field ^= 1;
 809         buf      += ctx->cid_table->coding_unit_size;
 810         buf_size -= ctx->cid_table->coding_unit_size;
 811         goto encode_coding_unit;
 812     }
 813
 814     ctx->frame.quality = ctx->qscale*FF_QP2LAMBDA;
 815
 816     return ctx->cid_table->frame_size;
 817 }
 818
 819 static int dnxhd_encode_end(AVCodecContext *avctx)
 820 {
 821     DNXHDEncContext *ctx = avctx->priv_data;
 822     int max_level = 1<<(ctx->cid_table->bit_depth+2);
 823     int i;
 824
 825     av_free(ctx->vlc_codes-max_level*2);
 826     av_free(ctx->vlc_bits -max_level*2);
 827     av_freep(&ctx->run_codes);
 828     av_freep(&ctx->run_bits);
 829
 830     av_freep(&ctx->mb_bits);
 831     av_freep(&ctx->mb_qscale);
 832     av_freep(&ctx->mb_rc);
 833     av_freep(&ctx->mb_cmp);
 834     av_freep(&ctx->slice_size);
 835     av_freep(&ctx->slice_offs);
 836
 837     av_freep(&ctx->qmatrix_c);
 838     av_freep(&ctx->qmatrix_l);
 839     av_freep(&ctx->qmatrix_c16);
 840     av_freep(&ctx->qmatrix_l16);
 841
 842     for (i = 1; i < avctx->thread_count; i++)
 843         av_freep(&ctx->thread[i]);
 844
 845     return 0;
 846 }
 847
 848 AVCodec dnxhd_encoder = {
 849     "dnxhd",
 850     CODEC_TYPE_VIDEO,
 851     CODEC_ID_DNXHD,
 852     sizeof(DNXHDEncContext),
 853     dnxhd_encode_init,
 854     dnxhd_encode_picture,
 855     dnxhd_encode_end,
 856     .pix_fmts = (const enum PixelFormat[]){PIX_FMT_YUV422P, PIX_FMT_NONE},
 857     .long_name = NULL_IF_CONFIG_SMALL("VC3/DNxHD"),
 858 };