3 * Copyright (c) 2007 Baptiste Coudurier <baptiste dot coudurier at smartjog dot com>
5 * VC-3 encoder funded by the British Broadcasting Corporation
7 * This file is part of FFmpeg.
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.
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.
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
25 #define RC_VARIANCE 1 // use variance or ssd for fast rc
29 #include "mpegvideo.h"
30 #include "dnxhddata.h"
42 int dct_quantize_c(MpegEncContext *s, DCTELEM *block, int n, int qscale, int *overflow);
44 typedef struct DNXHDEncContext {
45 MpegEncContext m; ///< Used for quantization dsp functions
49 const CIDEntry *cid_table;
50 uint8_t *msip; ///< Macroblock Scan Indices Payload
53 struct DNXHDEncContext *thread[MAX_THREADS];
55 unsigned dct_y_offset;
56 unsigned dct_uv_offset;
60 DECLARE_ALIGNED_16(DCTELEM, blocks[8][64]);
62 int (*qmatrix_c) [64];
63 int (*qmatrix_l) [64];
64 uint16_t (*qmatrix_l16)[2][64];
65 uint16_t (*qmatrix_c16)[2][64];
70 uint16_t *table_vlc_codes;
71 uint8_t *table_vlc_bits;
72 uint16_t *table_run_codes;
73 uint8_t *table_run_bits;
86 RCEntry (*mb_rc)[8160];
89 #define LAMBDA_FRAC_BITS 10
91 static int dnxhd_init_vlc(DNXHDEncContext *ctx)
95 CHECKED_ALLOCZ(ctx->table_vlc_codes, 449*2);
96 CHECKED_ALLOCZ(ctx->table_vlc_bits, 449);
97 CHECKED_ALLOCZ(ctx->table_run_codes, 63*2);
98 CHECKED_ALLOCZ(ctx->table_run_bits, 63);
100 for (i = 0; i < 257; i++) {
101 int level = ctx->cid_table->ac_level[i] +
102 (ctx->cid_table->ac_run_flag[i] << 7) + (ctx->cid_table->ac_index_flag[i] << 8);
104 if (ctx->cid_table->ac_level[i] == 64 && ctx->cid_table->ac_index_flag[i])
105 level -= 64; // use 0+(1<<8) level
106 ctx->table_vlc_codes[level] = ctx->cid_table->ac_codes[i];
107 ctx->table_vlc_bits [level] = ctx->cid_table->ac_bits[i];
109 for (i = 0; i < 62; i++) {
110 int run = ctx->cid_table->run[i];
112 ctx->table_run_codes[run] = ctx->cid_table->run_codes[i];
113 ctx->table_run_bits [run] = ctx->cid_table->run_bits[i];
120 static int dnxhd_init_qmat(DNXHDEncContext *ctx, int lbias, int cbias)
122 // init first elem to 1 to avoid div by 0 in convert_matrix
123 uint16_t weight_matrix[64] = {1,}; // convert_matrix needs uint16_t*
126 CHECKED_ALLOCZ(ctx->qmatrix_l, (ctx->m.avctx->qmax+1) * 64 * sizeof(int));
127 CHECKED_ALLOCZ(ctx->qmatrix_c, (ctx->m.avctx->qmax+1) * 64 * sizeof(int));
128 CHECKED_ALLOCZ(ctx->qmatrix_l16, (ctx->m.avctx->qmax+1) * 64 * 2 * sizeof(uint16_t));
129 CHECKED_ALLOCZ(ctx->qmatrix_c16, (ctx->m.avctx->qmax+1) * 64 * 2 * sizeof(uint16_t));
131 for (i = 1; i < 64; i++) {
132 int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]];
133 weight_matrix[j] = ctx->cid_table->luma_weight[i];
135 ff_convert_matrix(&ctx->m.dsp, ctx->qmatrix_l, ctx->qmatrix_l16, weight_matrix,
136 ctx->m.intra_quant_bias, 1, ctx->m.avctx->qmax, 1);
137 for (i = 1; i < 64; i++) {
138 int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]];
139 weight_matrix[j] = ctx->cid_table->chroma_weight[i];
141 ff_convert_matrix(&ctx->m.dsp, ctx->qmatrix_c, ctx->qmatrix_c16, weight_matrix,
142 ctx->m.intra_quant_bias, 1, ctx->m.avctx->qmax, 1);
143 for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
144 for (i = 0; i < 64; i++) {
145 ctx->qmatrix_l [qscale] [i] <<= 2; ctx->qmatrix_c [qscale] [i] <<= 2;
146 ctx->qmatrix_l16[qscale][0][i] <<= 2; ctx->qmatrix_l16[qscale][1][i] <<= 2;
147 ctx->qmatrix_c16[qscale][0][i] <<= 2; ctx->qmatrix_c16[qscale][1][i] <<= 2;
155 static int dnxhd_init_rc(DNXHDEncContext *ctx)
157 CHECKED_ALLOCZ(ctx->mb_rc, 8160*ctx->m.avctx->qmax*sizeof(RCEntry));
158 if (ctx->m.avctx->mb_decision != FF_MB_DECISION_RD)
159 CHECKED_ALLOCZ(ctx->mb_cmp, ctx->m.mb_num*sizeof(RCCMPEntry));
161 ctx->frame_bits = (ctx->cid_table->coding_unit_size - 640 - 4) * 8;
163 ctx->lambda = 2<<LAMBDA_FRAC_BITS; // qscale 2
169 static int dnxhd_encode_init(AVCodecContext *avctx)
171 DNXHDEncContext *ctx = avctx->priv_data;
174 if (avctx->flags & CODEC_FLAG_INTERLACED_DCT) {
175 if (avctx->bit_rate == 120000000)
177 else if (avctx->bit_rate == 185000000)
180 if (avctx->bit_rate == 120000000)
182 else if (avctx->bit_rate == 185000000)
184 else if (avctx->bit_rate == 36000000)
187 if (!ctx->cid || avctx->width != 1920 || avctx->height != 1080 || avctx->pix_fmt != PIX_FMT_YUV422P) {
188 av_log(avctx, AV_LOG_ERROR, "video parameters incompatible with DNxHD\n");
192 index = ff_dnxhd_get_cid_table(ctx->cid);
193 ctx->cid_table = &ff_dnxhd_cid_table[index];
195 ctx->m.avctx = avctx;
199 dsputil_init(&ctx->m.dsp, avctx);
200 ff_dct_common_init(&ctx->m);
201 if (!ctx->m.dct_quantize)
202 ctx->m.dct_quantize = dct_quantize_c;
204 ctx->m.mb_height = (avctx->height + 15) / 16;
205 ctx->m.mb_width = (avctx->width + 15) / 16;
207 if (avctx->flags & CODEC_FLAG_INTERLACED_DCT) {
209 ctx->m.mb_height /= 2;
212 ctx->m.mb_num = ctx->m.mb_height * ctx->m.mb_width;
214 if (avctx->intra_quant_bias != FF_DEFAULT_QUANT_BIAS)
215 ctx->m.intra_quant_bias = avctx->intra_quant_bias;
216 if (dnxhd_init_qmat(ctx, ctx->m.intra_quant_bias, 0) < 0) // XXX tune lbias/cbias
219 if (dnxhd_init_vlc(ctx) < 0)
221 if (dnxhd_init_rc(ctx) < 0)
224 CHECKED_ALLOCZ(ctx->slice_size, ctx->m.mb_height*sizeof(uint32_t));
225 CHECKED_ALLOCZ(ctx->mb_bits, ctx->m.mb_num *sizeof(uint16_t));
226 CHECKED_ALLOCZ(ctx->mb_qscale, ctx->m.mb_num *sizeof(uint8_t));
228 ctx->frame.key_frame = 1;
229 ctx->frame.pict_type = FF_I_TYPE;
230 ctx->m.avctx->coded_frame = &ctx->frame;
232 if (avctx->thread_count > MAX_THREADS || (avctx->thread_count > ctx->m.mb_height)) {
233 av_log(avctx, AV_LOG_ERROR, "too many threads\n");
237 ctx->thread[0] = ctx;
238 for (i = 1; i < avctx->thread_count; i++) {
239 ctx->thread[i] = av_malloc(sizeof(DNXHDEncContext));
240 memcpy(ctx->thread[i], ctx, sizeof(DNXHDEncContext));
243 for (i = 0; i < avctx->thread_count; i++) {
244 ctx->thread[i]->m.start_mb_y = (ctx->m.mb_height*(i ) + avctx->thread_count/2) / avctx->thread_count;
245 ctx->thread[i]->m.end_mb_y = (ctx->m.mb_height*(i+1) + avctx->thread_count/2) / avctx->thread_count;
249 fail: //for CHECKED_ALLOCZ
253 static int dnxhd_write_header(AVCodecContext *avctx, uint8_t *buf)
255 DNXHDEncContext *ctx = avctx->priv_data;
256 const uint8_t header_prefix[5] = { 0x00,0x00,0x02,0x80,0x01 };
258 memcpy(buf, header_prefix, 5);
259 buf[5] = ctx->interlaced ? ctx->cur_field+2 : 0x01;
260 buf[6] = 0x80; // crc flag off
261 buf[7] = 0xa0; // reserved
262 AV_WB16(buf + 0x18, avctx->height); // ALPF
263 AV_WB16(buf + 0x1a, avctx->width); // SPL
264 AV_WB16(buf + 0x1d, avctx->height); // NAL
266 buf[0x21] = 0x38; // FIXME 8 bit per comp
267 buf[0x22] = 0x88 + (ctx->frame.interlaced_frame<<2);
268 AV_WB32(buf + 0x28, ctx->cid); // CID
269 buf[0x2c] = ctx->interlaced ? 0 : 0x80;
271 buf[0x5f] = 0x01; // UDL
273 buf[0x167] = 0x02; // reserved
274 AV_WB16(buf + 0x16a, ctx->m.mb_height * 4 + 4); // MSIPS
275 buf[0x16d] = ctx->m.mb_height; // Ns
276 buf[0x16f] = 0x10; // reserved
278 ctx->msip = buf + 0x170;
282 static av_always_inline void dnxhd_encode_dc(DNXHDEncContext *ctx, int diff)
286 nbits = av_log2_16bit(-2*diff);
289 nbits = av_log2_16bit(2*diff);
291 put_bits(&ctx->m.pb, ctx->cid_table->dc_bits[nbits] + nbits,
292 (ctx->cid_table->dc_codes[nbits]<<nbits) + (diff & ((1 << nbits) - 1)));
295 static av_always_inline void dnxhd_encode_block(DNXHDEncContext *ctx, DCTELEM *block, int last_index, int n)
297 int last_non_zero = 0;
301 dnxhd_encode_dc(ctx, block[0] - ctx->m.last_dc[n]);
302 ctx->m.last_dc[n] = block[0];
304 for (i = 1; i <= last_index; i++) {
305 j = ctx->m.intra_scantable.permutated[i];
308 int run_level = i - last_non_zero - 1;
310 MASK_ABS(sign, slevel);
312 offset = (slevel-1) >> 6;
313 slevel = 256 | (slevel & 63); // level 64 is treated as 0
317 put_bits(&ctx->m.pb, ctx->table_vlc_bits[slevel]+1, (ctx->table_vlc_codes[slevel]<<1)|(sign&1));
319 put_bits(&ctx->m.pb, 4, offset);
323 put_bits(&ctx->m.pb, ctx->table_run_bits[run_level], ctx->table_run_codes[run_level]);
327 put_bits(&ctx->m.pb, ctx->table_vlc_bits[0], ctx->table_vlc_codes[0]); // EOB
330 static av_always_inline void dnxhd_unquantize_c(DNXHDEncContext *ctx, DCTELEM *block, int n, int qscale, int last_index)
332 const uint8_t *weight_matrix;
336 weight_matrix = (n&2) ? ctx->cid_table->chroma_weight : ctx->cid_table->luma_weight;
338 for (i = 1; i <= last_index; i++) {
339 int j = ctx->m.intra_scantable.permutated[i];
343 level = (1-2*level) * qscale * weight_matrix[i];
344 if (weight_matrix[i] != 32)
349 level = (2*level+1) * qscale * weight_matrix[i];
350 if (weight_matrix[i] != 32)
359 static av_always_inline int dnxhd_ssd_block(DCTELEM *qblock, DCTELEM *block)
363 for (i = 0; i < 64; i++)
364 score += (block[i]-qblock[i])*(block[i]-qblock[i]);
368 static av_always_inline int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, DCTELEM *block, int last_index)
370 int last_non_zero = 0;
373 for (i = 1; i <= last_index; i++) {
374 j = ctx->m.intra_scantable.permutated[i];
377 int run_level = i - last_non_zero - 1;
378 level = FFABS(level);
380 level = 256 | (level & 63); // level 64 is treated as 0
383 level |= (!!run_level)<<7;
384 bits += ctx->table_vlc_bits[level]+1 + ctx->table_run_bits[run_level];
391 static av_always_inline void dnxhd_get_pixels_4x8(DCTELEM *restrict block, const uint8_t *pixels, int line_size)
394 for (i = 0; i < 4; i++) {
395 block[0] = pixels[0];
396 block[1] = pixels[1];
397 block[2] = pixels[2];
398 block[3] = pixels[3];
399 block[4] = pixels[4];
400 block[5] = pixels[5];
401 block[6] = pixels[6];
402 block[7] = pixels[7];
406 memcpy(block , block- 8, sizeof(*block)*8);
407 memcpy(block+ 8, block-16, sizeof(*block)*8);
408 memcpy(block+16, block-24, sizeof(*block)*8);
409 memcpy(block+24, block-32, sizeof(*block)*8);
412 static av_always_inline void dnxhd_get_blocks(DNXHDEncContext *ctx, int mb_x, int mb_y)
414 const uint8_t *ptr_y = ctx->thread[0]->src[0] + ((mb_y << 4) * ctx->m.linesize) + (mb_x << 4);
415 const uint8_t *ptr_u = ctx->thread[0]->src[1] + ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << 3);
416 const uint8_t *ptr_v = ctx->thread[0]->src[2] + ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << 3);
417 DSPContext *dsp = &ctx->m.dsp;
419 dsp->get_pixels(ctx->blocks[0], ptr_y , ctx->m.linesize);
420 dsp->get_pixels(ctx->blocks[1], ptr_y + 8, ctx->m.linesize);
421 dsp->get_pixels(ctx->blocks[2], ptr_u , ctx->m.uvlinesize);
422 dsp->get_pixels(ctx->blocks[3], ptr_v , ctx->m.uvlinesize);
424 if (mb_y+1 == ctx->m.mb_height) {
425 if (ctx->interlaced) {
426 dnxhd_get_pixels_4x8(ctx->blocks[4], ptr_y + ctx->dct_y_offset , ctx->m.linesize);
427 dnxhd_get_pixels_4x8(ctx->blocks[5], ptr_y + ctx->dct_y_offset + 8, ctx->m.linesize);
428 dnxhd_get_pixels_4x8(ctx->blocks[6], ptr_u + ctx->dct_uv_offset , ctx->m.uvlinesize);
429 dnxhd_get_pixels_4x8(ctx->blocks[7], ptr_v + ctx->dct_uv_offset , ctx->m.uvlinesize);
431 memset(ctx->blocks[4], 0, 4*64*sizeof(DCTELEM));
433 dsp->get_pixels(ctx->blocks[4], ptr_y + ctx->dct_y_offset , ctx->m.linesize);
434 dsp->get_pixels(ctx->blocks[5], ptr_y + ctx->dct_y_offset + 8, ctx->m.linesize);
435 dsp->get_pixels(ctx->blocks[6], ptr_u + ctx->dct_uv_offset , ctx->m.uvlinesize);
436 dsp->get_pixels(ctx->blocks[7], ptr_v + ctx->dct_uv_offset , ctx->m.uvlinesize);
440 static av_always_inline int dnxhd_switch_matrix(DNXHDEncContext *ctx, int i)
443 ctx->m.q_intra_matrix16 = ctx->qmatrix_c16;
444 ctx->m.q_intra_matrix = ctx->qmatrix_c;
447 ctx->m.q_intra_matrix16 = ctx->qmatrix_l16;
448 ctx->m.q_intra_matrix = ctx->qmatrix_l;
453 static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg)
455 DNXHDEncContext *ctx = arg;
457 int qscale = ctx->thread[0]->qscale;
459 for (mb_y = ctx->m.start_mb_y; mb_y < ctx->m.end_mb_y; mb_y++) {
462 ctx->m.last_dc[2] = 1024;
464 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
465 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
471 dnxhd_get_blocks(ctx, mb_x, mb_y);
473 for (i = 0; i < 8; i++) {
474 DECLARE_ALIGNED_16(DCTELEM, block[64]);
475 DCTELEM *src_block = ctx->blocks[i];
476 int overflow, nbits, diff, last_index;
477 int n = dnxhd_switch_matrix(ctx, i);
479 memcpy(block, src_block, sizeof(block));
480 last_index = ctx->m.dct_quantize((MpegEncContext*)ctx, block, i, qscale, &overflow);
481 ac_bits += dnxhd_calc_ac_bits(ctx, block, last_index);
483 diff = block[0] - ctx->m.last_dc[n];
484 if (diff < 0) nbits = av_log2_16bit(-2*diff);
485 else nbits = av_log2_16bit( 2*diff);
486 dc_bits += ctx->cid_table->dc_bits[nbits] + nbits;
488 ctx->m.last_dc[n] = block[0];
490 if (avctx->mb_decision == FF_MB_DECISION_RD || !RC_VARIANCE) {
491 dnxhd_unquantize_c(ctx, block, i, qscale, last_index);
492 ctx->m.dsp.idct(block);
493 ssd += dnxhd_ssd_block(block, src_block);
496 ctx->mb_rc[qscale][mb].ssd = ssd;
497 ctx->mb_rc[qscale][mb].bits = ac_bits+dc_bits+12+8*ctx->table_vlc_bits[0];
503 static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg)
505 DNXHDEncContext *ctx = arg;
508 for (mb_y = ctx->m.start_mb_y; mb_y < ctx->m.end_mb_y; mb_y++) {
511 ctx->m.last_dc[2] = 1024;
512 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
513 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
514 int qscale = ctx->mb_qscale[mb];
517 put_bits(&ctx->m.pb, 12, qscale<<1);
519 dnxhd_get_blocks(ctx, mb_x, mb_y);
521 for (i = 0; i < 8; i++) {
522 DCTELEM *block = ctx->blocks[i];
523 int last_index, overflow;
524 int n = dnxhd_switch_matrix(ctx, i);
525 last_index = ctx->m.dct_quantize((MpegEncContext*)ctx, block, i, qscale, &overflow);
526 dnxhd_encode_block(ctx, block, last_index, n);
529 if (put_bits_count(&ctx->m.pb)&31)
530 put_bits(&ctx->m.pb, 32-(put_bits_count(&ctx->m.pb)&31), 0);
532 flush_put_bits(&ctx->m.pb);
536 static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx, uint8_t *buf)
540 for (i = 0; i < ctx->m.avctx->thread_count; i++) {
542 for (mb_y = ctx->thread[i]->m.start_mb_y; mb_y < ctx->thread[i]->m.end_mb_y; mb_y++) {
543 ctx->slice_size[mb_y] = 0;
544 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
545 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
546 ctx->slice_size[mb_y] += ctx->mb_bits[mb];
548 ctx->slice_size[mb_y] = (ctx->slice_size[mb_y]+31)&~31;
549 ctx->slice_size[mb_y] >>= 3;
550 thread_size += ctx->slice_size[mb_y];
552 init_put_bits(&ctx->thread[i]->m.pb, buf + 640 + offset, thread_size);
553 offset += thread_size;
557 static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg)
559 DNXHDEncContext *ctx = arg;
561 for (mb_y = ctx->m.start_mb_y; mb_y < ctx->m.end_mb_y; mb_y++) {
562 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
563 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
564 uint8_t *pix = ctx->thread[0]->src[0] + ((mb_y<<4) * ctx->m.linesize) + (mb_x<<4);
565 int sum = ctx->m.dsp.pix_sum(pix, ctx->m.linesize);
566 int varc = (ctx->m.dsp.pix_norm1(pix, ctx->m.linesize) - (((unsigned)(sum*sum))>>8)+128)>>8;
567 ctx->mb_cmp[mb].value = varc;
568 ctx->mb_cmp[mb].mb = mb;
574 static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx)
576 int lambda, up_step, down_step;
577 int last_lower = INT_MAX, last_higher = 0;
580 for (q = 1; q < avctx->qmax; q++) {
582 avctx->execute(avctx, dnxhd_calc_bits_thread, (void**)&ctx->thread[0], NULL, avctx->thread_count);
584 up_step = down_step = 2<<LAMBDA_FRAC_BITS;
585 lambda = ctx->lambda;
590 if (lambda == last_higher) {
592 end = 1; // need to set final qscales/bits
594 for (y = 0; y < ctx->m.mb_height; y++) {
595 for (x = 0; x < ctx->m.mb_width; x++) {
596 unsigned min = UINT_MAX;
598 int mb = y*ctx->m.mb_width+x;
599 for (q = 1; q < avctx->qmax; q++) {
600 unsigned score = ctx->mb_rc[q][mb].bits*lambda+(ctx->mb_rc[q][mb].ssd<<LAMBDA_FRAC_BITS);
606 bits += ctx->mb_rc[qscale][mb].bits;
607 ctx->mb_qscale[mb] = qscale;
608 ctx->mb_bits[mb] = ctx->mb_rc[qscale][mb].bits;
610 bits = (bits+31)&~31; // padding
611 if (bits > ctx->frame_bits)
614 //dprintf(ctx->m.avctx, "lambda %d, up %u, down %u, bits %d, frame %d\n",
615 // lambda, last_higher, last_lower, bits, ctx->frame_bits);
617 if (bits > ctx->frame_bits)
621 if (bits < ctx->frame_bits) {
622 last_lower = FFMIN(lambda, last_lower);
623 if (last_higher != 0)
624 lambda = (lambda+last_higher)>>1;
627 down_step *= 5; // XXX tune ?
628 up_step = 1<<LAMBDA_FRAC_BITS;
629 lambda = FFMAX(1, lambda);
630 if (lambda == last_lower)
633 last_higher = FFMAX(lambda, last_higher);
634 if (last_lower != INT_MAX)
635 lambda = (lambda+last_lower)>>1;
639 down_step = 1<<LAMBDA_FRAC_BITS;
642 //dprintf(ctx->m.avctx, "out lambda %d\n", lambda);
643 ctx->lambda = lambda;
647 static int dnxhd_find_qscale(DNXHDEncContext *ctx)
653 int last_lower = INT_MAX;
657 qscale = ctx->qscale;
660 ctx->qscale = qscale;
661 // XXX avoid recalculating bits
662 ctx->m.avctx->execute(ctx->m.avctx, dnxhd_calc_bits_thread, (void**)&ctx->thread[0], NULL, ctx->m.avctx->thread_count);
663 for (y = 0; y < ctx->m.mb_height; y++) {
664 for (x = 0; x < ctx->m.mb_width; x++)
665 bits += ctx->mb_rc[qscale][y*ctx->m.mb_width+x].bits;
666 bits = (bits+31)&~31; // padding
667 if (bits > ctx->frame_bits)
670 //dprintf(ctx->m.avctx, "%d, qscale %d, bits %d, frame %d, higher %d, lower %d\n",
671 // ctx->m.avctx->frame_number, qscale, bits, ctx->frame_bits, last_higher, last_lower);
672 if (bits < ctx->frame_bits) {
675 if (last_higher == qscale - 1) {
676 qscale = last_higher;
679 last_lower = FFMIN(qscale, last_lower);
680 if (last_higher != 0)
681 qscale = (qscale+last_higher)>>1;
683 qscale -= down_step++;
688 if (last_lower == qscale + 1)
690 last_higher = FFMAX(qscale, last_higher);
691 if (last_lower != INT_MAX)
692 qscale = (qscale+last_lower)>>1;
696 if (qscale >= ctx->m.avctx->qmax)
700 //dprintf(ctx->m.avctx, "out qscale %d\n", qscale);
701 ctx->qscale = qscale;
705 static int dnxhd_rc_cmp(const void *a, const void *b)
707 return ((RCCMPEntry *)b)->value - ((RCCMPEntry *)a)->value;
710 static int dnxhd_encode_variance(AVCodecContext *avctx, DNXHDEncContext *ctx)
714 if (dnxhd_find_qscale(ctx) < 0)
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;
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;
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
731 max_bits += 31; //worst padding
733 if (max_bits > ctx->frame_bits) {
735 avctx->execute(avctx, dnxhd_mb_var_thread, (void**)&ctx->thread[0], NULL, avctx->thread_count);
736 qsort(ctx->mb_cmp, ctx->m.mb_num, sizeof(RCEntry), dnxhd_rc_cmp);
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;
747 static void dnxhd_load_picture(DNXHDEncContext *ctx, AVFrame *frame)
751 for (i = 0; i < 3; i++) {
752 ctx->frame.data[i] = frame->data[i];
753 ctx->frame.linesize[i] = frame->linesize[i];
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;
763 ctx->frame.interlaced_frame = frame->interlaced_frame;
764 ctx->cur_field = frame->interlaced_frame && !frame->top_field_first;
767 static int dnxhd_encode_picture(AVCodecContext *avctx, unsigned char *buf, int buf_size, void *data)
769 DNXHDEncContext *ctx = avctx->priv_data;
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");
778 dnxhd_load_picture(ctx, data);
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];
787 dnxhd_write_header(avctx, buf);
789 if (avctx->mb_decision == FF_MB_DECISION_RD)
790 ret = dnxhd_encode_rdo(avctx, ctx);
792 ret = dnxhd_encode_variance(avctx, ctx);
794 av_log(avctx, AV_LOG_ERROR, "picture could not fit ratecontrol constraints\n");
798 dnxhd_setup_threads_slices(ctx, buf);
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));
807 avctx->execute(avctx, dnxhd_encode_thread, (void**)&ctx->thread[0], NULL, avctx->thread_count);
809 AV_WB32(buf + ctx->cid_table->coding_unit_size - 4, 0x600DC0DE); // EOF
811 if (ctx->interlaced && first_field) {
814 buf += ctx->cid_table->coding_unit_size;
815 buf_size -= ctx->cid_table->coding_unit_size;
816 goto encode_coding_unit;
819 return ctx->cid_table->frame_size;
822 static int dnxhd_encode_end(AVCodecContext *avctx)
824 DNXHDEncContext *ctx = avctx->priv_data;
827 av_freep(&ctx->table_vlc_codes);
828 av_freep(&ctx->table_vlc_bits);
829 av_freep(&ctx->table_run_codes);
830 av_freep(&ctx->table_run_bits);
832 av_freep(&ctx->mb_bits);
833 av_freep(&ctx->mb_qscale);
834 av_freep(&ctx->mb_rc);
835 av_freep(&ctx->mb_cmp);
836 av_freep(&ctx->slice_size);
838 av_freep(&ctx->qmatrix_c);
839 av_freep(&ctx->qmatrix_l);
840 av_freep(&ctx->qmatrix_c16);
841 av_freep(&ctx->qmatrix_l16);
843 for (i = 1; i < avctx->thread_count; i++)
844 av_freep(&ctx->thread[i]);
849 AVCodec dnxhd_encoder = {
853 sizeof(DNXHDEncContext),
855 dnxhd_encode_picture,
857 .pix_fmts = (enum PixelFormat[]){PIX_FMT_YUV422P, -1},