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->mb_bits,    ctx->m.mb_num   *sizeof(uint16_t), fail);
 208     FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_qscale,  ctx->m.mb_num   *sizeof(uint8_t) , fail);
 209
 210     ctx->frame.key_frame = 1;
 211     ctx->frame.pict_type = FF_I_TYPE;
 212     ctx->m.avctx->coded_frame = &ctx->frame;
 213
 214     if (avctx->thread_count > MAX_THREADS || (avctx->thread_count > ctx->m.mb_height)) {
 215         av_log(avctx, AV_LOG_ERROR, "too many threads\n");
 216         return -1;
 217     }
 218
 219     ctx->thread[0] = ctx;
 220     for (i = 1; i < avctx->thread_count; i++) {
 221         ctx->thread[i] =  av_malloc(sizeof(DNXHDEncContext));
 222         memcpy(ctx->thread[i], ctx, sizeof(DNXHDEncContext));
 223     }
 224
 225     for (i = 0; i < avctx->thread_count; i++) {
 226         ctx->thread[i]->m.start_mb_y = (ctx->m.mb_height*(i  ) + avctx->thread_count/2) / avctx->thread_count;
 227         ctx->thread[i]->m.end_mb_y   = (ctx->m.mb_height*(i+1) + avctx->thread_count/2) / avctx->thread_count;
 228     }
 229
 230     return 0;
 231  fail: //for FF_ALLOCZ_OR_GOTO
 232     return -1;
 233 }
 234
 235 static int dnxhd_write_header(AVCodecContext *avctx, uint8_t *buf)
 236 {
 237     DNXHDEncContext *ctx = avctx->priv_data;
 238     const uint8_t header_prefix[5] = { 0x00,0x00,0x02,0x80,0x01 };
 239
 240     memset(buf, 0, 640);
 241
 242     memcpy(buf, header_prefix, 5);
 243     buf[5] = ctx->interlaced ? ctx->cur_field+2 : 0x01;
 244     buf[6] = 0x80; // crc flag off
 245     buf[7] = 0xa0; // reserved
 246     AV_WB16(buf + 0x18, avctx->height); // ALPF
 247     AV_WB16(buf + 0x1a, avctx->width);  // SPL
 248     AV_WB16(buf + 0x1d, avctx->height); // NAL
 249
 250     buf[0x21] = 0x38; // FIXME 8 bit per comp
 251     buf[0x22] = 0x88 + (ctx->frame.interlaced_frame<<2);
 252     AV_WB32(buf + 0x28, ctx->cid); // CID
 253     buf[0x2c] = ctx->interlaced ? 0 : 0x80;
 254
 255     buf[0x5f] = 0x01; // UDL
 256
 257     buf[0x167] = 0x02; // reserved
 258     AV_WB16(buf + 0x16a, ctx->m.mb_height * 4 + 4); // MSIPS
 259     buf[0x16d] = ctx->m.mb_height; // Ns
 260     buf[0x16f] = 0x10; // reserved
 261
 262     ctx->msip = buf + 0x170;
 263     return 0;
 264 }
 265
 266 static av_always_inline void dnxhd_encode_dc(DNXHDEncContext *ctx, int diff)
 267 {
 268     int nbits;
 269     if (diff < 0) {
 270         nbits = av_log2_16bit(-2*diff);
 271         diff--;
 272     } else {
 273         nbits = av_log2_16bit(2*diff);
 274     }
 275     put_bits(&ctx->m.pb, ctx->cid_table->dc_bits[nbits] + nbits,
 276              (ctx->cid_table->dc_codes[nbits]<<nbits) + (diff & ((1 << nbits) - 1)));
 277 }
 278
 279 static av_always_inline void dnxhd_encode_block(DNXHDEncContext *ctx, DCTELEM *block, int last_index, int n)
 280 {
 281     int last_non_zero = 0;
 282     int slevel, i, j;
 283
 284     dnxhd_encode_dc(ctx, block[0] - ctx->m.last_dc[n]);
 285     ctx->m.last_dc[n] = block[0];
 286
 287     for (i = 1; i <= last_index; i++) {
 288         j = ctx->m.intra_scantable.permutated[i];
 289         slevel = block[j];
 290         if (slevel) {
 291             int run_level = i - last_non_zero - 1;
 292             int rlevel = (slevel<<1)|!!run_level;
 293             put_bits(&ctx->m.pb, ctx->vlc_bits[rlevel], ctx->vlc_codes[rlevel]);
 294             if (run_level)
 295                 put_bits(&ctx->m.pb, ctx->run_bits[run_level], ctx->run_codes[run_level]);
 296             last_non_zero = i;
 297         }
 298     }
 299     put_bits(&ctx->m.pb, ctx->vlc_bits[0], ctx->vlc_codes[0]); // EOB
 300 }
 301
 302 static av_always_inline void dnxhd_unquantize_c(DNXHDEncContext *ctx, DCTELEM *block, int n, int qscale, int last_index)
 303 {
 304     const uint8_t *weight_matrix;
 305     int level;
 306     int i;
 307
 308     weight_matrix = (n&2) ? ctx->cid_table->chroma_weight : ctx->cid_table->luma_weight;
 309
 310     for (i = 1; i <= last_index; i++) {
 311         int j = ctx->m.intra_scantable.permutated[i];
 312         level = block[j];
 313         if (level) {
 314             if (level < 0) {
 315                 level = (1-2*level) * qscale * weight_matrix[i];
 316                 if (weight_matrix[i] != 32)
 317                     level += 32;
 318                 level >>= 6;
 319                 level = -level;
 320             } else {
 321                 level = (2*level+1) * qscale * weight_matrix[i];
 322                 if (weight_matrix[i] != 32)
 323                     level += 32;
 324                 level >>= 6;
 325             }
 326             block[j] = level;
 327         }
 328     }
 329 }
 330
 331 static av_always_inline int dnxhd_ssd_block(DCTELEM *qblock, DCTELEM *block)
 332 {
 333     int score = 0;
 334     int i;
 335     for (i = 0; i < 64; i++)
 336         score += (block[i]-qblock[i])*(block[i]-qblock[i]);
 337     return score;
 338 }
 339
 340 static av_always_inline int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, DCTELEM *block, int last_index)
 341 {
 342     int last_non_zero = 0;
 343     int bits = 0;
 344     int i, j, level;
 345     for (i = 1; i <= last_index; i++) {
 346         j = ctx->m.intra_scantable.permutated[i];
 347         level = block[j];
 348         if (level) {
 349             int run_level = i - last_non_zero - 1;
 350             bits += ctx->vlc_bits[(level<<1)|!!run_level]+ctx->run_bits[run_level];
 351             last_non_zero = i;
 352         }
 353     }
 354     return bits;
 355 }
 356
 357 static av_always_inline void dnxhd_get_blocks(DNXHDEncContext *ctx, int mb_x, int mb_y)
 358 {
 359     const uint8_t *ptr_y = ctx->thread[0]->src[0] + ((mb_y << 4) * ctx->m.linesize)   + (mb_x << 4);
 360     const uint8_t *ptr_u = ctx->thread[0]->src[1] + ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << 3);
 361     const uint8_t *ptr_v = ctx->thread[0]->src[2] + ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << 3);
 362     DSPContext *dsp = &ctx->m.dsp;
 363
 364     dsp->get_pixels(ctx->blocks[0], ptr_y    , ctx->m.linesize);
 365     dsp->get_pixels(ctx->blocks[1], ptr_y + 8, ctx->m.linesize);
 366     dsp->get_pixels(ctx->blocks[2], ptr_u    , ctx->m.uvlinesize);
 367     dsp->get_pixels(ctx->blocks[3], ptr_v    , ctx->m.uvlinesize);
 368
 369     if (mb_y+1 == ctx->m.mb_height && ctx->m.avctx->height == 1080) {
 370         if (ctx->interlaced) {
 371             ctx->get_pixels_8x4_sym(ctx->blocks[4], ptr_y + ctx->dct_y_offset    , ctx->m.linesize);
 372             ctx->get_pixels_8x4_sym(ctx->blocks[5], ptr_y + ctx->dct_y_offset + 8, ctx->m.linesize);
 373             ctx->get_pixels_8x4_sym(ctx->blocks[6], ptr_u + ctx->dct_uv_offset   , ctx->m.uvlinesize);
 374             ctx->get_pixels_8x4_sym(ctx->blocks[7], ptr_v + ctx->dct_uv_offset   , ctx->m.uvlinesize);
 375         } else {
 376             dsp->clear_block(ctx->blocks[4]); dsp->clear_block(ctx->blocks[5]);
 377             dsp->clear_block(ctx->blocks[6]); dsp->clear_block(ctx->blocks[7]);
 378         }
 379     } else {
 380         dsp->get_pixels(ctx->blocks[4], ptr_y + ctx->dct_y_offset    , ctx->m.linesize);
 381         dsp->get_pixels(ctx->blocks[5], ptr_y + ctx->dct_y_offset + 8, ctx->m.linesize);
 382         dsp->get_pixels(ctx->blocks[6], ptr_u + ctx->dct_uv_offset   , ctx->m.uvlinesize);
 383         dsp->get_pixels(ctx->blocks[7], ptr_v + ctx->dct_uv_offset   , ctx->m.uvlinesize);
 384     }
 385 }
 386
 387 static av_always_inline int dnxhd_switch_matrix(DNXHDEncContext *ctx, int i)
 388 {
 389     if (i&2) {
 390         ctx->m.q_intra_matrix16 = ctx->qmatrix_c16;
 391         ctx->m.q_intra_matrix   = ctx->qmatrix_c;
 392         return 1 + (i&1);
 393     } else {
 394         ctx->m.q_intra_matrix16 = ctx->qmatrix_l16;
 395         ctx->m.q_intra_matrix   = ctx->qmatrix_l;
 396         return 0;
 397     }
 398 }
 399
 400 static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg)
 401 {
 402     DNXHDEncContext *ctx = *(void**)arg;
 403     int mb_y, mb_x;
 404     int qscale = ctx->thread[0]->qscale;
 405
 406     for (mb_y = ctx->m.start_mb_y; mb_y < ctx->m.end_mb_y; mb_y++) {
 407         ctx->m.last_dc[0] =
 408         ctx->m.last_dc[1] =
 409         ctx->m.last_dc[2] = 1024;
 410
 411         for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
 412             unsigned mb = mb_y * ctx->m.mb_width + mb_x;
 413             int ssd     = 0;
 414             int ac_bits = 0;
 415             int dc_bits = 0;
 416             int i;
 417
 418             dnxhd_get_blocks(ctx, mb_x, mb_y);
 419
 420             for (i = 0; i < 8; i++) {
 421                 DECLARE_ALIGNED_16(DCTELEM, block[64]);
 422                 DCTELEM *src_block = ctx->blocks[i];
 423                 int overflow, nbits, diff, last_index;
 424                 int n = dnxhd_switch_matrix(ctx, i);
 425
 426                 memcpy(block, src_block, sizeof(block));
 427                 last_index = ctx->m.dct_quantize((MpegEncContext*)ctx, block, i, qscale, &overflow);
 428                 ac_bits += dnxhd_calc_ac_bits(ctx, block, last_index);
 429
 430                 diff = block[0] - ctx->m.last_dc[n];
 431                 if (diff < 0) nbits = av_log2_16bit(-2*diff);
 432                 else          nbits = av_log2_16bit( 2*diff);
 433                 dc_bits += ctx->cid_table->dc_bits[nbits] + nbits;
 434
 435                 ctx->m.last_dc[n] = block[0];
 436
 437                 if (avctx->mb_decision == FF_MB_DECISION_RD || !RC_VARIANCE) {
 438                     dnxhd_unquantize_c(ctx, block, i, qscale, last_index);
 439                     ctx->m.dsp.idct(block);
 440                     ssd += dnxhd_ssd_block(block, src_block);
 441                 }
 442             }
 443             ctx->mb_rc[qscale][mb].ssd = ssd;
 444             ctx->mb_rc[qscale][mb].bits = ac_bits+dc_bits+12+8*ctx->vlc_bits[0];
 445         }
 446     }
 447     return 0;
 448 }
 449
 450 static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg)
 451 {
 452     DNXHDEncContext *ctx = *(void**)arg;
 453     int mb_y, mb_x;
 454
 455     for (mb_y = ctx->m.start_mb_y; mb_y < ctx->m.end_mb_y; mb_y++) {
 456         ctx->m.last_dc[0] =
 457         ctx->m.last_dc[1] =
 458         ctx->m.last_dc[2] = 1024;
 459         for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
 460             unsigned mb = mb_y * ctx->m.mb_width + mb_x;
 461             int qscale = ctx->mb_qscale[mb];
 462             int i;
 463
 464             put_bits(&ctx->m.pb, 12, qscale<<1);
 465
 466             dnxhd_get_blocks(ctx, mb_x, mb_y);
 467
 468             for (i = 0; i < 8; i++) {
 469                 DCTELEM *block = ctx->blocks[i];
 470                 int last_index, overflow;
 471                 int n = dnxhd_switch_matrix(ctx, i);
 472                 last_index = ctx->m.dct_quantize((MpegEncContext*)ctx, block, i, qscale, &overflow);
 473                 //START_TIMER;
 474                 dnxhd_encode_block(ctx, block, last_index, n);
 475                 //STOP_TIMER("encode_block");
 476             }
 477         }
 478         if (put_bits_count(&ctx->m.pb)&31)
 479             put_bits(&ctx->m.pb, 32-(put_bits_count(&ctx->m.pb)&31), 0);
 480     }
 481     flush_put_bits(&ctx->m.pb);
 482     return 0;
 483 }
 484
 485 static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx, uint8_t *buf)
 486 {
 487     int mb_y, mb_x;
 488     int i, offset = 0;
 489     for (i = 0; i < ctx->m.avctx->thread_count; i++) {
 490         int thread_size = 0;
 491         for (mb_y = ctx->thread[i]->m.start_mb_y; mb_y < ctx->thread[i]->m.end_mb_y; mb_y++) {
 492             ctx->slice_size[mb_y] = 0;
 493             for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
 494                 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
 495                 ctx->slice_size[mb_y] += ctx->mb_bits[mb];
 496             }
 497             ctx->slice_size[mb_y] = (ctx->slice_size[mb_y]+31)&~31;
 498             ctx->slice_size[mb_y] >>= 3;
 499             thread_size += ctx->slice_size[mb_y];
 500         }
 501         init_put_bits(&ctx->thread[i]->m.pb, buf + 640 + offset, thread_size);
 502         offset += thread_size;
 503     }
 504 }
 505
 506 static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg)
 507 {
 508     DNXHDEncContext *ctx = *(void**)arg;
 509     int mb_y, mb_x;
 510     for (mb_y = ctx->m.start_mb_y; mb_y < ctx->m.end_mb_y; mb_y++) {
 511         for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
 512             unsigned mb  = mb_y * ctx->m.mb_width + mb_x;
 513             uint8_t *pix = ctx->thread[0]->src[0] + ((mb_y<<4) * ctx->m.linesize) + (mb_x<<4);
 514             int sum      = ctx->m.dsp.pix_sum(pix, ctx->m.linesize);
 515             int varc     = (ctx->m.dsp.pix_norm1(pix, ctx->m.linesize) - (((unsigned)(sum*sum))>>8)+128)>>8;
 516             ctx->mb_cmp[mb].value = varc;
 517             ctx->mb_cmp[mb].mb = mb;
 518         }
 519     }
 520     return 0;
 521 }
 522
 523 static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx)
 524 {
 525     int lambda, up_step, down_step;
 526     int last_lower = INT_MAX, last_higher = 0;
 527     int x, y, q;
 528
 529     for (q = 1; q < avctx->qmax; q++) {
 530         ctx->qscale = q;
 531         avctx->execute(avctx, dnxhd_calc_bits_thread, (void**)&ctx->thread[0], NULL, avctx->thread_count, sizeof(void*));
 532     }
 533     up_step = down_step = 2<<LAMBDA_FRAC_BITS;
 534     lambda = ctx->lambda;
 535
 536     for (;;) {
 537         int bits = 0;
 538         int end = 0;
 539         if (lambda == last_higher) {
 540             lambda++;
 541             end = 1; // need to set final qscales/bits
 542         }
 543         for (y = 0; y < ctx->m.mb_height; y++) {
 544             for (x = 0; x < ctx->m.mb_width; x++) {
 545                 unsigned min = UINT_MAX;
 546                 int qscale = 1;
 547                 int mb = y*ctx->m.mb_width+x;
 548                 for (q = 1; q < avctx->qmax; q++) {
 549                     unsigned score = ctx->mb_rc[q][mb].bits*lambda+(ctx->mb_rc[q][mb].ssd<<LAMBDA_FRAC_BITS);
 550                     if (score < min) {
 551                         min = score;
 552                         qscale = q;
 553                     }
 554                 }
 555                 bits += ctx->mb_rc[qscale][mb].bits;
 556                 ctx->mb_qscale[mb] = qscale;
 557                 ctx->mb_bits[mb] = ctx->mb_rc[qscale][mb].bits;
 558             }
 559             bits = (bits+31)&~31; // padding
 560             if (bits > ctx->frame_bits)
 561                 break;
 562         }
 563         //dprintf(ctx->m.avctx, "lambda %d, up %u, down %u, bits %d, frame %d\n",
 564         //        lambda, last_higher, last_lower, bits, ctx->frame_bits);
 565         if (end) {
 566             if (bits > ctx->frame_bits)
 567                 return -1;
 568             break;
 569         }
 570         if (bits < ctx->frame_bits) {
 571             last_lower = FFMIN(lambda, last_lower);
 572             if (last_higher != 0)
 573                 lambda = (lambda+last_higher)>>1;
 574             else
 575                 lambda -= down_step;
 576             down_step *= 5; // XXX tune ?
 577             up_step = 1<<LAMBDA_FRAC_BITS;
 578             lambda = FFMAX(1, lambda);
 579             if (lambda == last_lower)
 580                 break;
 581         } else {
 582             last_higher = FFMAX(lambda, last_higher);
 583             if (last_lower != INT_MAX)
 584                 lambda = (lambda+last_lower)>>1;
 585             else
 586                 lambda += up_step;
 587             up_step *= 5;
 588             down_step = 1<<LAMBDA_FRAC_BITS;
 589         }
 590     }
 591     //dprintf(ctx->m.avctx, "out lambda %d\n", lambda);
 592     ctx->lambda = lambda;
 593     return 0;
 594 }
 595
 596 static int dnxhd_find_qscale(DNXHDEncContext *ctx)
 597 {
 598     int bits = 0;
 599     int up_step = 1;
 600     int down_step = 1;
 601     int last_higher = 0;
 602     int last_lower = INT_MAX;
 603     int qscale;
 604     int x, y;
 605
 606     qscale = ctx->qscale;
 607     for (;;) {
 608         bits = 0;
 609         ctx->qscale = qscale;
 610         // XXX avoid recalculating bits
 611         ctx->m.avctx->execute(ctx->m.avctx, dnxhd_calc_bits_thread, (void**)&ctx->thread[0], NULL, ctx->m.avctx->thread_count, sizeof(void*));
 612         for (y = 0; y < ctx->m.mb_height; y++) {
 613             for (x = 0; x < ctx->m.mb_width; x++)
 614                 bits += ctx->mb_rc[qscale][y*ctx->m.mb_width+x].bits;
 615             bits = (bits+31)&~31; // padding
 616             if (bits > ctx->frame_bits)
 617                 break;
 618         }
 619         //dprintf(ctx->m.avctx, "%d, qscale %d, bits %d, frame %d, higher %d, lower %d\n",
 620         //        ctx->m.avctx->frame_number, qscale, bits, ctx->frame_bits, last_higher, last_lower);
 621         if (bits < ctx->frame_bits) {
 622             if (qscale == 1)
 623                 return 1;
 624             if (last_higher == qscale - 1) {
 625                 qscale = last_higher;
 626                 break;
 627             }
 628             last_lower = FFMIN(qscale, last_lower);
 629             if (last_higher != 0)
 630                 qscale = (qscale+last_higher)>>1;
 631             else
 632                 qscale -= down_step++;
 633             if (qscale < 1)
 634                 qscale = 1;
 635             up_step = 1;
 636         } else {
 637             if (last_lower == qscale + 1)
 638                 break;
 639             last_higher = FFMAX(qscale, last_higher);
 640             if (last_lower != INT_MAX)
 641                 qscale = (qscale+last_lower)>>1;
 642             else
 643                 qscale += up_step++;
 644             down_step = 1;
 645             if (qscale >= ctx->m.avctx->qmax)
 646                 return -1;
 647         }
 648     }
 649     //dprintf(ctx->m.avctx, "out qscale %d\n", qscale);
 650     ctx->qscale = qscale;
 651     return 0;
 652 }
 653
 654 #define BUCKET_BITS 8
 655 #define RADIX_PASSES 4
 656 #define NBUCKETS (1 << BUCKET_BITS)
 657
 658 static inline int get_bucket(int value, int shift)
 659 {
 660     value >>= shift;
 661     value &= NBUCKETS - 1;
 662     return NBUCKETS - 1 - value;
 663 }
 664
 665 static void radix_count(const RCCMPEntry *data, int size, int buckets[RADIX_PASSES][NBUCKETS])
 666 {
 667     int i, j;
 668     memset(buckets, 0, sizeof(buckets[0][0]) * RADIX_PASSES * NBUCKETS);
 669     for (i = 0; i < size; i++) {
 670         int v = data[i].value;
 671         for (j = 0; j < RADIX_PASSES; j++) {
 672             buckets[j][get_bucket(v, 0)]++;
 673             v >>= BUCKET_BITS;
 674         }
 675         assert(!v);
 676     }
 677     for (j = 0; j < RADIX_PASSES; j++) {
 678         int offset = size;
 679         for (i = NBUCKETS - 1; i >= 0; i--)
 680             buckets[j][i] = offset -= buckets[j][i];
 681         assert(!buckets[j][0]);
 682     }
 683 }
 684
 685 static void radix_sort_pass(RCCMPEntry *dst, const RCCMPEntry *data, int size, int buckets[NBUCKETS], int pass)
 686 {
 687     int shift = pass * BUCKET_BITS;
 688     int i;
 689     for (i = 0; i < size; i++) {
 690         int v = get_bucket(data[i].value, shift);
 691         int pos = buckets[v]++;
 692         dst[pos] = data[i];
 693     }
 694 }
 695
 696 static void radix_sort(RCCMPEntry *data, int size)
 697 {
 698     int buckets[RADIX_PASSES][NBUCKETS];
 699     RCCMPEntry *tmp = av_malloc(sizeof(*tmp) * size);
 700     radix_count(data, size, buckets);
 701     radix_sort_pass(tmp, data, size, buckets[0], 0);
 702     radix_sort_pass(data, tmp, size, buckets[1], 1);
 703     if (buckets[2][NBUCKETS - 1] || buckets[3][NBUCKETS - 1]) {
 704         radix_sort_pass(tmp, data, size, buckets[2], 2);
 705         radix_sort_pass(data, tmp, size, buckets[3], 3);
 706     }
 707     av_free(tmp);
 708 }
 709
 710 static int dnxhd_encode_fast(AVCodecContext *avctx, DNXHDEncContext *ctx)
 711 {
 712     int max_bits = 0;
 713     int ret, x, y;
 714     if ((ret = dnxhd_find_qscale(ctx)) < 0)
 715         return -1;
 716     for (y = 0; y < ctx->m.mb_height; y++) {
 717         for (x = 0; x < ctx->m.mb_width; x++) {
 718             int mb = y*ctx->m.mb_width+x;
 719             int delta_bits;
 720             ctx->mb_qscale[mb] = ctx->qscale;
 721             ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale][mb].bits;
 722             max_bits += ctx->mb_rc[ctx->qscale][mb].bits;
 723             if (!RC_VARIANCE) {
 724                 delta_bits = ctx->mb_rc[ctx->qscale][mb].bits-ctx->mb_rc[ctx->qscale+1][mb].bits;
 725                 ctx->mb_cmp[mb].mb = mb;
 726                 ctx->mb_cmp[mb].value = delta_bits ?
 727                     ((ctx->mb_rc[ctx->qscale][mb].ssd-ctx->mb_rc[ctx->qscale+1][mb].ssd)*100)/delta_bits
 728                     : INT_MIN; //avoid increasing qscale
 729             }
 730         }
 731         max_bits += 31; //worst padding
 732     }
 733     if (!ret) {
 734         if (RC_VARIANCE)
 735             avctx->execute(avctx, dnxhd_mb_var_thread, (void**)&ctx->thread[0], NULL, avctx->thread_count, sizeof(void*));
 736         radix_sort(ctx->mb_cmp, ctx->m.mb_num);
 737         for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) {
 738             int mb = ctx->mb_cmp[x].mb;
 739             max_bits -= ctx->mb_rc[ctx->qscale][mb].bits - ctx->mb_rc[ctx->qscale+1][mb].bits;
 740             ctx->mb_qscale[mb] = ctx->qscale+1;
 741             ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale+1][mb].bits;
 742         }
 743     }
 744     return 0;
 745 }
 746
 747 static void dnxhd_load_picture(DNXHDEncContext *ctx, const AVFrame *frame)
 748 {
 749     int i;
 750
 751     for (i = 0; i < 3; i++) {
 752         ctx->frame.data[i]     = frame->data[i];
 753         ctx->frame.linesize[i] = frame->linesize[i];
 754     }
 755
 756     for (i = 0; i < ctx->m.avctx->thread_count; i++) {
 757         ctx->thread[i]->m.linesize    = ctx->frame.linesize[0]<<ctx->interlaced;
 758         ctx->thread[i]->m.uvlinesize  = ctx->frame.linesize[1]<<ctx->interlaced;
 759         ctx->thread[i]->dct_y_offset  = ctx->m.linesize  *8;
 760         ctx->thread[i]->dct_uv_offset = ctx->m.uvlinesize*8;
 761     }
 762
 763     ctx->frame.interlaced_frame = frame->interlaced_frame;
 764     ctx->cur_field = frame->interlaced_frame && !frame->top_field_first;
 765 }
 766
 767 static int dnxhd_encode_picture(AVCodecContext *avctx, unsigned char *buf, int buf_size, void *data)
 768 {
 769     DNXHDEncContext *ctx = avctx->priv_data;
 770     int first_field = 1;
 771     int offset, i, ret;
 772
 773     if (buf_size < ctx->cid_table->frame_size) {
 774         av_log(avctx, AV_LOG_ERROR, "output buffer is too small to compress picture\n");
 775         return -1;
 776     }
 777
 778     dnxhd_load_picture(ctx, data);
 779
 780  encode_coding_unit:
 781     for (i = 0; i < 3; i++) {
 782         ctx->src[i] = ctx->frame.data[i];
 783         if (ctx->interlaced && ctx->cur_field)
 784             ctx->src[i] += ctx->frame.linesize[i];
 785     }
 786
 787     dnxhd_write_header(avctx, buf);
 788
 789     if (avctx->mb_decision == FF_MB_DECISION_RD)
 790         ret = dnxhd_encode_rdo(avctx, ctx);
 791     else
 792         ret = dnxhd_encode_fast(avctx, ctx);
 793     if (ret < 0) {
 794         av_log(avctx, AV_LOG_ERROR, "picture could not fit ratecontrol constraints\n");
 795         return -1;
 796     }
 797
 798     dnxhd_setup_threads_slices(ctx, buf);
 799
 800     offset = 0;
 801     for (i = 0; i < ctx->m.mb_height; i++) {
 802         AV_WB32(ctx->msip + i * 4, offset);
 803         offset += ctx->slice_size[i];
 804         assert(!(ctx->slice_size[i] & 3));
 805     }
 806
 807     avctx->execute(avctx, dnxhd_encode_thread, (void**)&ctx->thread[0], NULL, avctx->thread_count, sizeof(void*));
 808
 809     assert(640 + offset + 4 <= ctx->cid_table->coding_unit_size);
 810     memset(buf + 640 + offset, 0, ctx->cid_table->coding_unit_size - 4 - offset - 640);
 811
 812     AV_WB32(buf + ctx->cid_table->coding_unit_size - 4, 0x600DC0DE); // EOF
 813
 814     if (ctx->interlaced && first_field) {
 815         first_field     = 0;
 816         ctx->cur_field ^= 1;
 817         buf      += ctx->cid_table->coding_unit_size;
 818         buf_size -= ctx->cid_table->coding_unit_size;
 819         goto encode_coding_unit;
 820     }
 821
 822     ctx->frame.quality = ctx->qscale*FF_QP2LAMBDA;
 823
 824     return ctx->cid_table->frame_size;
 825 }
 826
 827 static int dnxhd_encode_end(AVCodecContext *avctx)
 828 {
 829     DNXHDEncContext *ctx = avctx->priv_data;
 830     int max_level = 1<<(ctx->cid_table->bit_depth+2);
 831     int i;
 832
 833     av_free(ctx->vlc_codes-max_level*2);
 834     av_free(ctx->vlc_bits -max_level*2);
 835     av_freep(&ctx->run_codes);
 836     av_freep(&ctx->run_bits);
 837
 838     av_freep(&ctx->mb_bits);
 839     av_freep(&ctx->mb_qscale);
 840     av_freep(&ctx->mb_rc);
 841     av_freep(&ctx->mb_cmp);
 842     av_freep(&ctx->slice_size);
 843
 844     av_freep(&ctx->qmatrix_c);
 845     av_freep(&ctx->qmatrix_l);
 846     av_freep(&ctx->qmatrix_c16);
 847     av_freep(&ctx->qmatrix_l16);
 848
 849     for (i = 1; i < avctx->thread_count; i++)
 850         av_freep(&ctx->thread[i]);
 851
 852     return 0;
 853 }
 854
 855 AVCodec dnxhd_encoder = {
 856     "dnxhd",
 857     CODEC_TYPE_VIDEO,
 858     CODEC_ID_DNXHD,
 859     sizeof(DNXHDEncContext),
 860     dnxhd_encode_init,
 861     dnxhd_encode_picture,
 862     dnxhd_encode_end,
 863     .pix_fmts = (const enum PixelFormat[]){PIX_FMT_YUV422P, PIX_FMT_NONE},
 864     .long_name = NULL_IF_CONFIG_SMALL("VC3/DNxHD"),
 865 };