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)
185 if (!ctx->cid || avctx->width != 1920 || avctx->height != 1080 || avctx->pix_fmt != PIX_FMT_YUV422P) {
186 av_log(avctx, AV_LOG_ERROR, "video parameters incompatible with DNxHD\n");
190 index = ff_dnxhd_get_cid_table(ctx->cid);
191 ctx->cid_table = &ff_dnxhd_cid_table[index];
193 ctx->m.avctx = avctx;
197 dsputil_init(&ctx->m.dsp, avctx);
198 ff_dct_common_init(&ctx->m);
199 if (!ctx->m.dct_quantize)
200 ctx->m.dct_quantize = dct_quantize_c;
202 ctx->m.mb_height = (avctx->height + 15) / 16;
203 ctx->m.mb_width = (avctx->width + 15) / 16;
205 if (avctx->flags & CODEC_FLAG_INTERLACED_DCT) {
207 ctx->m.mb_height /= 2;
210 ctx->m.mb_num = ctx->m.mb_height * ctx->m.mb_width;
212 if (avctx->intra_quant_bias != FF_DEFAULT_QUANT_BIAS)
213 ctx->m.intra_quant_bias = avctx->intra_quant_bias;
214 if (dnxhd_init_qmat(ctx, ctx->m.intra_quant_bias, 0) < 0) // XXX tune lbias/cbias
217 if (dnxhd_init_vlc(ctx) < 0)
219 if (dnxhd_init_rc(ctx) < 0)
222 CHECKED_ALLOCZ(ctx->slice_size, ctx->m.mb_height*sizeof(uint32_t));
223 CHECKED_ALLOCZ(ctx->mb_bits, ctx->m.mb_num *sizeof(uint16_t));
224 CHECKED_ALLOCZ(ctx->mb_qscale, ctx->m.mb_num *sizeof(uint8_t));
226 ctx->frame.key_frame = 1;
227 ctx->frame.pict_type = FF_I_TYPE;
228 ctx->m.avctx->coded_frame = &ctx->frame;
230 if (avctx->thread_count > MAX_THREADS || (avctx->thread_count > ctx->m.mb_height)) {
231 av_log(avctx, AV_LOG_ERROR, "too many threads\n");
235 ctx->thread[0] = ctx;
236 for (i = 1; i < avctx->thread_count; i++) {
237 ctx->thread[i] = av_malloc(sizeof(DNXHDEncContext));
238 memcpy(ctx->thread[i], ctx, sizeof(DNXHDEncContext));
241 for (i = 0; i < avctx->thread_count; i++) {
242 ctx->thread[i]->m.start_mb_y = (ctx->m.mb_height*(i ) + avctx->thread_count/2) / avctx->thread_count;
243 ctx->thread[i]->m.end_mb_y = (ctx->m.mb_height*(i+1) + avctx->thread_count/2) / avctx->thread_count;
247 fail: //for CHECKED_ALLOCZ
251 static int dnxhd_write_header(AVCodecContext *avctx, uint8_t *buf)
253 DNXHDEncContext *ctx = avctx->priv_data;
254 const uint8_t header_prefix[5] = { 0x00,0x00,0x02,0x80,0x01 };
256 memcpy(buf, header_prefix, 5);
257 buf[5] = ctx->interlaced ? ctx->cur_field+2 : 0x01;
258 buf[6] = 0x80; // crc flag off
259 buf[7] = 0xa0; // reserved
260 AV_WB16(buf + 0x18, avctx->height); // ALPF
261 AV_WB16(buf + 0x1a, avctx->width); // SPL
262 AV_WB16(buf + 0x1d, avctx->height); // NAL
264 buf[0x21] = 0x38; // FIXME 8 bit per comp
265 buf[0x22] = 0x88 + (ctx->frame.interlaced_frame<<2);
266 AV_WB32(buf + 0x28, ctx->cid); // CID
267 buf[0x2c] = ctx->interlaced ? 0 : 0x80;
269 buf[0x5f] = 0x01; // UDL
271 buf[0x167] = 0x02; // reserved
272 AV_WB16(buf + 0x16a, ctx->m.mb_height * 4 + 4); // MSIPS
273 buf[0x16d] = ctx->m.mb_height; // Ns
274 buf[0x16f] = 0x10; // reserved
276 ctx->msip = buf + 0x170;
280 static av_always_inline void dnxhd_encode_dc(DNXHDEncContext *ctx, int diff)
284 nbits = av_log2_16bit(-2*diff);
287 nbits = av_log2_16bit(2*diff);
289 put_bits(&ctx->m.pb, ctx->cid_table->dc_bits[nbits] + nbits,
290 (ctx->cid_table->dc_codes[nbits]<<nbits) + (diff & ((1 << nbits) - 1)));
293 static av_always_inline void dnxhd_encode_block(DNXHDEncContext *ctx, DCTELEM *block, int last_index, int n)
295 int last_non_zero = 0;
299 dnxhd_encode_dc(ctx, block[0] - ctx->m.last_dc[n]);
300 ctx->m.last_dc[n] = block[0];
302 for (i = 1; i <= last_index; i++) {
303 j = ctx->m.intra_scantable.permutated[i];
306 int run_level = i - last_non_zero - 1;
308 MASK_ABS(sign, slevel);
310 offset = (slevel-1) >> 6;
311 slevel = 256 | (slevel & 63); // level 64 is treated as 0
315 put_bits(&ctx->m.pb, ctx->table_vlc_bits[slevel]+1, (ctx->table_vlc_codes[slevel]<<1)|(sign&1));
317 put_bits(&ctx->m.pb, 4, offset);
321 put_bits(&ctx->m.pb, ctx->table_run_bits[run_level], ctx->table_run_codes[run_level]);
325 put_bits(&ctx->m.pb, ctx->table_vlc_bits[0], ctx->table_vlc_codes[0]); // EOB
328 static av_always_inline void dnxhd_unquantize_c(DNXHDEncContext *ctx, DCTELEM *block, int n, int qscale, int last_index)
330 const uint8_t *weight_matrix;
334 weight_matrix = (n&2) ? ctx->cid_table->chroma_weight : ctx->cid_table->luma_weight;
336 for (i = 1; i <= last_index; i++) {
337 int j = ctx->m.intra_scantable.permutated[i];
341 level = (1-2*level) * qscale * weight_matrix[i];
342 if (weight_matrix[i] != 32)
347 level = (2*level+1) * qscale * weight_matrix[i];
348 if (weight_matrix[i] != 32)
357 static av_always_inline int dnxhd_ssd_block(DCTELEM *qblock, DCTELEM *block)
361 for (i = 0; i < 64; i++)
362 score += (block[i]-qblock[i])*(block[i]-qblock[i]);
366 static av_always_inline int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, DCTELEM *block, int last_index)
368 int last_non_zero = 0;
371 for (i = 1; i <= last_index; i++) {
372 j = ctx->m.intra_scantable.permutated[i];
375 int run_level = i - last_non_zero - 1;
376 level = FFABS(level);
378 level = 256 | (level & 63); // level 64 is treated as 0
381 level |= (!!run_level)<<7;
382 bits += ctx->table_vlc_bits[level]+1 + ctx->table_run_bits[run_level];
389 static av_always_inline void dnxhd_get_pixels_4x8(DCTELEM *restrict block, const uint8_t *pixels, int line_size)
392 for (i = 0; i < 4; i++) {
393 block[0] = pixels[0];
394 block[1] = pixels[1];
395 block[2] = pixels[2];
396 block[3] = pixels[3];
397 block[4] = pixels[4];
398 block[5] = pixels[5];
399 block[6] = pixels[6];
400 block[7] = pixels[7];
404 memcpy(block , block- 8, sizeof(*block)*8);
405 memcpy(block+ 8, block-16, sizeof(*block)*8);
406 memcpy(block+16, block-24, sizeof(*block)*8);
407 memcpy(block+24, block-32, sizeof(*block)*8);
410 static av_always_inline void dnxhd_get_blocks(DNXHDEncContext *ctx, int mb_x, int mb_y)
412 const uint8_t *ptr_y = ctx->thread[0]->src[0] + ((mb_y << 4) * ctx->m.linesize) + (mb_x << 4);
413 const uint8_t *ptr_u = ctx->thread[0]->src[1] + ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << 3);
414 const uint8_t *ptr_v = ctx->thread[0]->src[2] + ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << 3);
415 DSPContext *dsp = &ctx->m.dsp;
417 dsp->get_pixels(ctx->blocks[0], ptr_y , ctx->m.linesize);
418 dsp->get_pixels(ctx->blocks[1], ptr_y + 8, ctx->m.linesize);
419 dsp->get_pixels(ctx->blocks[2], ptr_u , ctx->m.uvlinesize);
420 dsp->get_pixels(ctx->blocks[3], ptr_v , ctx->m.uvlinesize);
422 if (mb_y+1 == ctx->m.mb_height) {
423 if (ctx->interlaced) {
424 dnxhd_get_pixels_4x8(ctx->blocks[4], ptr_y + ctx->dct_y_offset , ctx->m.linesize);
425 dnxhd_get_pixels_4x8(ctx->blocks[5], ptr_y + ctx->dct_y_offset + 8, ctx->m.linesize);
426 dnxhd_get_pixels_4x8(ctx->blocks[6], ptr_u + ctx->dct_uv_offset , ctx->m.uvlinesize);
427 dnxhd_get_pixels_4x8(ctx->blocks[7], ptr_v + ctx->dct_uv_offset , ctx->m.uvlinesize);
429 memset(ctx->blocks[4], 0, 4*64*sizeof(DCTELEM));
431 dsp->get_pixels(ctx->blocks[4], ptr_y + ctx->dct_y_offset , ctx->m.linesize);
432 dsp->get_pixels(ctx->blocks[5], ptr_y + ctx->dct_y_offset + 8, ctx->m.linesize);
433 dsp->get_pixels(ctx->blocks[6], ptr_u + ctx->dct_uv_offset , ctx->m.uvlinesize);
434 dsp->get_pixels(ctx->blocks[7], ptr_v + ctx->dct_uv_offset , ctx->m.uvlinesize);
438 static av_always_inline int dnxhd_switch_matrix(DNXHDEncContext *ctx, int i)
441 ctx->m.q_intra_matrix16 = ctx->qmatrix_c16;
442 ctx->m.q_intra_matrix = ctx->qmatrix_c;
445 ctx->m.q_intra_matrix16 = ctx->qmatrix_l16;
446 ctx->m.q_intra_matrix = ctx->qmatrix_l;
451 static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg)
453 DNXHDEncContext *ctx = arg;
455 int qscale = ctx->thread[0]->qscale;
457 for (mb_y = ctx->m.start_mb_y; mb_y < ctx->m.end_mb_y; mb_y++) {
460 ctx->m.last_dc[2] = 1024;
462 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
463 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
469 dnxhd_get_blocks(ctx, mb_x, mb_y);
471 for (i = 0; i < 8; i++) {
472 DECLARE_ALIGNED_16(DCTELEM, block[64]);
473 DCTELEM *src_block = ctx->blocks[i];
474 int overflow, nbits, diff, last_index;
475 int n = dnxhd_switch_matrix(ctx, i);
477 memcpy(block, src_block, sizeof(block));
478 last_index = ctx->m.dct_quantize((MpegEncContext*)ctx, block, i, qscale, &overflow);
479 ac_bits += dnxhd_calc_ac_bits(ctx, block, last_index);
481 diff = block[0] - ctx->m.last_dc[n];
482 if (diff < 0) nbits = av_log2_16bit(-2*diff);
483 else nbits = av_log2_16bit( 2*diff);
484 dc_bits += ctx->cid_table->dc_bits[nbits] + nbits;
486 ctx->m.last_dc[n] = block[0];
488 if (avctx->mb_decision == FF_MB_DECISION_RD || !RC_VARIANCE) {
489 dnxhd_unquantize_c(ctx, block, i, qscale, last_index);
490 ctx->m.dsp.idct(block);
491 ssd += dnxhd_ssd_block(block, src_block);
494 ctx->mb_rc[qscale][mb].ssd = ssd;
495 ctx->mb_rc[qscale][mb].bits = ac_bits+dc_bits+12+8*ctx->table_vlc_bits[0];
501 static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg)
503 DNXHDEncContext *ctx = arg;
506 for (mb_y = ctx->m.start_mb_y; mb_y < ctx->m.end_mb_y; mb_y++) {
509 ctx->m.last_dc[2] = 1024;
510 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
511 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
512 int qscale = ctx->mb_qscale[mb];
515 put_bits(&ctx->m.pb, 12, qscale<<1);
517 dnxhd_get_blocks(ctx, mb_x, mb_y);
519 for (i = 0; i < 8; i++) {
520 DCTELEM *block = ctx->blocks[i];
521 int last_index, overflow;
522 int n = dnxhd_switch_matrix(ctx, i);
523 last_index = ctx->m.dct_quantize((MpegEncContext*)ctx, block, i, qscale, &overflow);
524 dnxhd_encode_block(ctx, block, last_index, n);
527 if (put_bits_count(&ctx->m.pb)&31)
528 put_bits(&ctx->m.pb, 32-(put_bits_count(&ctx->m.pb)&31), 0);
530 flush_put_bits(&ctx->m.pb);
534 static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx, uint8_t *buf)
538 for (i = 0; i < ctx->m.avctx->thread_count; i++) {
540 for (mb_y = ctx->thread[i]->m.start_mb_y; mb_y < ctx->thread[i]->m.end_mb_y; mb_y++) {
541 ctx->slice_size[mb_y] = 0;
542 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
543 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
544 ctx->slice_size[mb_y] += ctx->mb_bits[mb];
546 ctx->slice_size[mb_y] = (ctx->slice_size[mb_y]+31)&~31;
547 ctx->slice_size[mb_y] >>= 3;
548 thread_size += ctx->slice_size[mb_y];
550 init_put_bits(&ctx->thread[i]->m.pb, buf + 640 + offset, thread_size);
551 offset += thread_size;
555 static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg)
557 DNXHDEncContext *ctx = arg;
559 for (mb_y = ctx->m.start_mb_y; mb_y < ctx->m.end_mb_y; mb_y++) {
560 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
561 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
562 uint8_t *pix = ctx->thread[0]->src[0] + ((mb_y<<4) * ctx->m.linesize) + (mb_x<<4);
563 int sum = ctx->m.dsp.pix_sum(pix, ctx->m.linesize);
564 int varc = (ctx->m.dsp.pix_norm1(pix, ctx->m.linesize) - (((unsigned)(sum*sum))>>8)+128)>>8;
565 ctx->mb_cmp[mb].value = varc;
566 ctx->mb_cmp[mb].mb = mb;
572 static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx)
574 unsigned lambda, up_lambda, down_lambda;
577 for (q = 1; q < avctx->qmax; q++) {
579 avctx->execute(avctx, dnxhd_calc_bits_thread, (void**)&ctx->thread[0], NULL, avctx->thread_count);
581 up_lambda = avctx->qmax<<LAMBDA_FRAC_BITS;
582 down_lambda = 1; // higher ?
583 lambda = ctx->lambda;
588 if (lambda == up_lambda) {
590 end = 1; // need to set final qscales/bits
592 if (lambda == down_lambda) {
596 for (y = 0; y < ctx->m.mb_height; y++) {
597 for (x = 0; x < ctx->m.mb_width; x++) {
598 unsigned min = UINT_MAX;
600 int mb = y*ctx->m.mb_width+x;
601 for (q = 1; q < avctx->qmax; q++) {
602 unsigned score = ctx->mb_rc[q][mb].bits*lambda+(ctx->mb_rc[q][mb].ssd<<LAMBDA_FRAC_BITS);
608 bits += ctx->mb_rc[qscale][mb].bits;
609 ctx->mb_qscale[mb] = qscale;
610 ctx->mb_bits[mb] = ctx->mb_rc[qscale][mb].bits;
612 bits = (bits+31)&~31; // padding
613 if (bits > ctx->frame_bits)
616 //dprintf(ctx->m.avctx, "lambda %d, up %d, down %d, bits %d, frame %d\n", lambda, up_lambda, down_lambda, bits, ctx->frame_bits);
618 if (bits > ctx->frame_bits)
622 if (bits < ctx->frame_bits) {
624 lambda = (down_lambda+lambda)>>1;
626 down_lambda = lambda;
627 lambda = (up_lambda+lambda)>>1;
630 ctx->lambda = lambda;
634 static int dnxhd_find_qscale(DNXHDEncContext *ctx)
640 int last_lower = INT_MAX;
644 qscale = ctx->qscale;
647 ctx->qscale = qscale;
648 // XXX avoid recalculating bits
649 ctx->m.avctx->execute(ctx->m.avctx, dnxhd_calc_bits_thread, (void**)&ctx->thread[0], NULL, ctx->m.avctx->thread_count);
650 for (y = 0; y < ctx->m.mb_height; y++) {
651 for (x = 0; x < ctx->m.mb_width; x++)
652 bits += ctx->mb_rc[qscale][y*ctx->m.mb_width+x].bits;
653 bits = (bits+31)&~31; // padding
654 if (bits > ctx->frame_bits)
657 //dprintf(ctx->m.avctx, "%d, qscale %d, bits %d, frame %d, higher %d, lower %d\n",
658 // ctx->m.avctx->frame_number, qscale, bits, ctx->frame_bits, last_higher, last_lower);
659 if (bits < ctx->frame_bits) {
662 if (last_higher == qscale - 1) {
663 qscale = last_higher;
666 last_lower = FFMIN(qscale, last_lower);
667 if (last_higher != 0)
668 qscale = (qscale+last_higher)>>1;
670 qscale -= down_step++;
675 if (last_lower == qscale + 1)
677 last_higher = FFMAX(qscale, last_higher);
678 if (last_lower != INT_MAX)
679 qscale = (qscale+last_lower)>>1;
683 if (qscale >= ctx->m.avctx->qmax)
687 //dprintf(ctx->m.avctx, "out qscale %d\n", qscale);
688 ctx->qscale = qscale;
692 static int dnxhd_rc_cmp(const void *a, const void *b)
694 return ((RCCMPEntry *)b)->value - ((RCCMPEntry *)a)->value;
697 static int dnxhd_encode_variance(AVCodecContext *avctx, DNXHDEncContext *ctx)
701 if (dnxhd_find_qscale(ctx) < 0)
703 for (y = 0; y < ctx->m.mb_height; y++) {
704 for (x = 0; x < ctx->m.mb_width; x++) {
705 int mb = y*ctx->m.mb_width+x;
707 ctx->mb_qscale[mb] = ctx->qscale;
708 ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale][mb].bits;
709 max_bits += ctx->mb_rc[ctx->qscale][mb].bits;
711 delta_bits = ctx->mb_rc[ctx->qscale][mb].bits-ctx->mb_rc[ctx->qscale+1][mb].bits;
712 ctx->mb_cmp[mb].mb = mb;
713 ctx->mb_cmp[mb].value = delta_bits ?
714 ((ctx->mb_rc[ctx->qscale][mb].ssd-ctx->mb_rc[ctx->qscale+1][mb].ssd)*100)/delta_bits
715 : INT_MIN; //avoid increasing qscale
718 max_bits += 31; //worst padding
720 if (max_bits > ctx->frame_bits) {
722 avctx->execute(avctx, dnxhd_mb_var_thread, (void**)&ctx->thread[0], NULL, avctx->thread_count);
723 qsort(ctx->mb_cmp, ctx->m.mb_num, sizeof(RCEntry), dnxhd_rc_cmp);
724 for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) {
725 int mb = ctx->mb_cmp[x].mb;
726 max_bits -= ctx->mb_rc[ctx->qscale][mb].bits - ctx->mb_rc[ctx->qscale+1][mb].bits;
727 ctx->mb_qscale[mb] = ctx->qscale+1;
728 ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale+1][mb].bits;
734 static void dnxhd_load_picture(DNXHDEncContext *ctx, AVFrame *frame)
738 for (i = 0; i < 3; i++) {
739 ctx->frame.data[i] = frame->data[i];
740 ctx->frame.linesize[i] = frame->linesize[i];
743 for (i = 0; i < ctx->m.avctx->thread_count; i++) {
744 ctx->thread[i]->m.linesize = ctx->frame.linesize[0]<<ctx->interlaced;
745 ctx->thread[i]->m.uvlinesize = ctx->frame.linesize[1]<<ctx->interlaced;
746 ctx->thread[i]->dct_y_offset = ctx->m.linesize *8;
747 ctx->thread[i]->dct_uv_offset = ctx->m.uvlinesize*8;
750 ctx->frame.interlaced_frame = frame->interlaced_frame;
751 ctx->cur_field = frame->interlaced_frame && !frame->top_field_first;
754 static int dnxhd_encode_picture(AVCodecContext *avctx, unsigned char *buf, int buf_size, void *data)
756 DNXHDEncContext *ctx = avctx->priv_data;
760 if (buf_size < ctx->cid_table->frame_size) {
761 av_log(avctx, AV_LOG_ERROR, "output buffer is too small to compress picture\n");
765 dnxhd_load_picture(ctx, data);
768 for (i = 0; i < 3; i++) {
769 ctx->src[i] = ctx->frame.data[i];
770 if (ctx->interlaced && ctx->cur_field)
771 ctx->src[i] += ctx->frame.linesize[i];
774 dnxhd_write_header(avctx, buf);
776 if (avctx->mb_decision == FF_MB_DECISION_RD)
777 ret = dnxhd_encode_rdo(avctx, ctx);
779 ret = dnxhd_encode_variance(avctx, ctx);
781 av_log(avctx, AV_LOG_ERROR, "picture could not fit ratecontrol constraints\n");
785 dnxhd_setup_threads_slices(ctx, buf);
788 for (i = 0; i < ctx->m.mb_height; i++) {
789 AV_WB32(ctx->msip + i * 4, offset);
790 offset += ctx->slice_size[i];
791 assert(!(ctx->slice_size[i] & 3));
794 avctx->execute(avctx, dnxhd_encode_thread, (void**)&ctx->thread[0], NULL, avctx->thread_count);
796 AV_WB32(buf + ctx->cid_table->coding_unit_size - 4, 0x600DC0DE); // EOF
798 if (ctx->interlaced && first_field) {
801 buf += ctx->cid_table->coding_unit_size;
802 buf_size -= ctx->cid_table->coding_unit_size;
803 goto encode_coding_unit;
806 return ctx->cid_table->frame_size;
809 static int dnxhd_encode_end(AVCodecContext *avctx)
811 DNXHDEncContext *ctx = avctx->priv_data;
814 av_freep(&ctx->table_vlc_codes);
815 av_freep(&ctx->table_vlc_bits);
816 av_freep(&ctx->table_run_codes);
817 av_freep(&ctx->table_run_bits);
819 av_freep(&ctx->mb_bits);
820 av_freep(&ctx->mb_qscale);
821 av_freep(&ctx->mb_rc);
822 av_freep(&ctx->mb_cmp);
823 av_freep(&ctx->slice_size);
825 av_freep(&ctx->qmatrix_c);
826 av_freep(&ctx->qmatrix_l);
827 av_freep(&ctx->qmatrix_c16);
828 av_freep(&ctx->qmatrix_l16);
830 for (i = 1; i < avctx->thread_count; i++)
831 av_freep(&ctx->thread[i]);
836 AVCodec dnxhd_encoder = {
840 sizeof(DNXHDEncContext),
842 dnxhd_encode_picture,
844 .pix_fmts = (enum PixelFormat[]){PIX_FMT_YUV422P, -1},