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"
32 int dct_quantize_c(MpegEncContext *s, DCTELEM *block, int n, int qscale, int *overflow);
34 #define LAMBDA_FRAC_BITS 10
36 static int dnxhd_init_vlc(DNXHDEncContext *ctx)
39 int max_level = 1<<(ctx->cid_table->bit_depth+2);
41 CHECKED_ALLOCZ(ctx->vlc_codes, max_level*4*sizeof(*ctx->vlc_codes));
42 CHECKED_ALLOCZ(ctx->vlc_bits, max_level*4*sizeof(*ctx->vlc_bits));
43 CHECKED_ALLOCZ(ctx->run_codes, 63*2);
44 CHECKED_ALLOCZ(ctx->run_bits, 63);
46 ctx->vlc_codes += max_level*2;
47 ctx->vlc_bits += max_level*2;
48 for (level = -max_level; level < max_level; level++) {
49 for (run = 0; run < 2; run++) {
50 int index = (level<<1)|run;
51 int sign, offset = 0, alevel = level;
53 MASK_ABS(sign, alevel);
55 offset = (alevel-1)>>6;
58 for (j = 0; j < 257; j++) {
59 if (ctx->cid_table->ac_level[j] == alevel &&
60 (!offset || (ctx->cid_table->ac_index_flag[j] && offset)) &&
61 (!run || (ctx->cid_table->ac_run_flag [j] && run))) {
62 assert(!ctx->vlc_codes[index]);
64 ctx->vlc_codes[index] = (ctx->cid_table->ac_codes[j]<<1)|(sign&1);
65 ctx->vlc_bits [index] = ctx->cid_table->ac_bits[j]+1;
67 ctx->vlc_codes[index] = ctx->cid_table->ac_codes[j];
68 ctx->vlc_bits [index] = ctx->cid_table->ac_bits [j];
73 assert(!alevel || j < 257);
75 ctx->vlc_codes[index] = (ctx->vlc_codes[index]<<ctx->cid_table->index_bits)|offset;
76 ctx->vlc_bits [index]+= ctx->cid_table->index_bits;
80 for (i = 0; i < 62; i++) {
81 int run = ctx->cid_table->run[i];
83 ctx->run_codes[run] = ctx->cid_table->run_codes[i];
84 ctx->run_bits [run] = ctx->cid_table->run_bits[i];
91 static int dnxhd_init_qmat(DNXHDEncContext *ctx, int lbias, int cbias)
93 // init first elem to 1 to avoid div by 0 in convert_matrix
94 uint16_t weight_matrix[64] = {1,}; // convert_matrix needs uint16_t*
97 CHECKED_ALLOCZ(ctx->qmatrix_l, (ctx->m.avctx->qmax+1) * 64 * sizeof(int));
98 CHECKED_ALLOCZ(ctx->qmatrix_c, (ctx->m.avctx->qmax+1) * 64 * sizeof(int));
99 CHECKED_ALLOCZ(ctx->qmatrix_l16, (ctx->m.avctx->qmax+1) * 64 * 2 * sizeof(uint16_t));
100 CHECKED_ALLOCZ(ctx->qmatrix_c16, (ctx->m.avctx->qmax+1) * 64 * 2 * sizeof(uint16_t));
102 for (i = 1; i < 64; i++) {
103 int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]];
104 weight_matrix[j] = ctx->cid_table->luma_weight[i];
106 ff_convert_matrix(&ctx->m.dsp, ctx->qmatrix_l, ctx->qmatrix_l16, weight_matrix,
107 ctx->m.intra_quant_bias, 1, ctx->m.avctx->qmax, 1);
108 for (i = 1; i < 64; i++) {
109 int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]];
110 weight_matrix[j] = ctx->cid_table->chroma_weight[i];
112 ff_convert_matrix(&ctx->m.dsp, ctx->qmatrix_c, ctx->qmatrix_c16, weight_matrix,
113 ctx->m.intra_quant_bias, 1, ctx->m.avctx->qmax, 1);
114 for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
115 for (i = 0; i < 64; i++) {
116 ctx->qmatrix_l [qscale] [i] <<= 2; ctx->qmatrix_c [qscale] [i] <<= 2;
117 ctx->qmatrix_l16[qscale][0][i] <<= 2; ctx->qmatrix_l16[qscale][1][i] <<= 2;
118 ctx->qmatrix_c16[qscale][0][i] <<= 2; ctx->qmatrix_c16[qscale][1][i] <<= 2;
126 static int dnxhd_init_rc(DNXHDEncContext *ctx)
128 CHECKED_ALLOCZ(ctx->mb_rc, 8160*ctx->m.avctx->qmax*sizeof(RCEntry));
129 if (ctx->m.avctx->mb_decision != FF_MB_DECISION_RD)
130 CHECKED_ALLOCZ(ctx->mb_cmp, ctx->m.mb_num*sizeof(RCCMPEntry));
132 ctx->frame_bits = (ctx->cid_table->coding_unit_size - 640 - 4) * 8;
134 ctx->lambda = 2<<LAMBDA_FRAC_BITS; // qscale 2
140 static int dnxhd_encode_init(AVCodecContext *avctx)
142 DNXHDEncContext *ctx = avctx->priv_data;
145 ctx->cid = ff_dnxhd_find_cid(avctx);
146 if (!ctx->cid || avctx->pix_fmt != PIX_FMT_YUV422P) {
147 av_log(avctx, AV_LOG_ERROR, "video parameters incompatible with DNxHD\n");
150 av_log(avctx, AV_LOG_DEBUG, "cid %d\n", ctx->cid);
152 index = ff_dnxhd_get_cid_table(ctx->cid);
153 ctx->cid_table = &ff_dnxhd_cid_table[index];
155 ctx->m.avctx = avctx;
159 dsputil_init(&ctx->m.dsp, avctx);
160 ff_dct_common_init(&ctx->m);
161 if (!ctx->m.dct_quantize)
162 ctx->m.dct_quantize = dct_quantize_c;
164 ctx->m.mb_height = (avctx->height + 15) / 16;
165 ctx->m.mb_width = (avctx->width + 15) / 16;
167 if (avctx->flags & CODEC_FLAG_INTERLACED_DCT) {
169 ctx->m.mb_height /= 2;
172 ctx->m.mb_num = ctx->m.mb_height * ctx->m.mb_width;
174 if (avctx->intra_quant_bias != FF_DEFAULT_QUANT_BIAS)
175 ctx->m.intra_quant_bias = avctx->intra_quant_bias;
176 if (dnxhd_init_qmat(ctx, ctx->m.intra_quant_bias, 0) < 0) // XXX tune lbias/cbias
179 if (dnxhd_init_vlc(ctx) < 0)
181 if (dnxhd_init_rc(ctx) < 0)
184 CHECKED_ALLOCZ(ctx->slice_size, ctx->m.mb_height*sizeof(uint32_t));
185 CHECKED_ALLOCZ(ctx->mb_bits, ctx->m.mb_num *sizeof(uint16_t));
186 CHECKED_ALLOCZ(ctx->mb_qscale, ctx->m.mb_num *sizeof(uint8_t));
188 ctx->frame.key_frame = 1;
189 ctx->frame.pict_type = FF_I_TYPE;
190 ctx->m.avctx->coded_frame = &ctx->frame;
192 if (avctx->thread_count > MAX_THREADS || (avctx->thread_count > ctx->m.mb_height)) {
193 av_log(avctx, AV_LOG_ERROR, "too many threads\n");
197 ctx->thread[0] = ctx;
198 for (i = 1; i < avctx->thread_count; i++) {
199 ctx->thread[i] = av_malloc(sizeof(DNXHDEncContext));
200 memcpy(ctx->thread[i], ctx, sizeof(DNXHDEncContext));
203 for (i = 0; i < avctx->thread_count; i++) {
204 ctx->thread[i]->m.start_mb_y = (ctx->m.mb_height*(i ) + avctx->thread_count/2) / avctx->thread_count;
205 ctx->thread[i]->m.end_mb_y = (ctx->m.mb_height*(i+1) + avctx->thread_count/2) / avctx->thread_count;
209 fail: //for CHECKED_ALLOCZ
213 static int dnxhd_write_header(AVCodecContext *avctx, uint8_t *buf)
215 DNXHDEncContext *ctx = avctx->priv_data;
216 const uint8_t header_prefix[5] = { 0x00,0x00,0x02,0x80,0x01 };
218 memcpy(buf, header_prefix, 5);
219 buf[5] = ctx->interlaced ? ctx->cur_field+2 : 0x01;
220 buf[6] = 0x80; // crc flag off
221 buf[7] = 0xa0; // reserved
222 AV_WB16(buf + 0x18, avctx->height); // ALPF
223 AV_WB16(buf + 0x1a, avctx->width); // SPL
224 AV_WB16(buf + 0x1d, avctx->height); // NAL
226 buf[0x21] = 0x38; // FIXME 8 bit per comp
227 buf[0x22] = 0x88 + (ctx->frame.interlaced_frame<<2);
228 AV_WB32(buf + 0x28, ctx->cid); // CID
229 buf[0x2c] = ctx->interlaced ? 0 : 0x80;
231 buf[0x5f] = 0x01; // UDL
233 buf[0x167] = 0x02; // reserved
234 AV_WB16(buf + 0x16a, ctx->m.mb_height * 4 + 4); // MSIPS
235 buf[0x16d] = ctx->m.mb_height; // Ns
236 buf[0x16f] = 0x10; // reserved
238 ctx->msip = buf + 0x170;
242 static av_always_inline void dnxhd_encode_dc(DNXHDEncContext *ctx, int diff)
246 nbits = av_log2_16bit(-2*diff);
249 nbits = av_log2_16bit(2*diff);
251 put_bits(&ctx->m.pb, ctx->cid_table->dc_bits[nbits] + nbits,
252 (ctx->cid_table->dc_codes[nbits]<<nbits) + (diff & ((1 << nbits) - 1)));
255 static av_always_inline void dnxhd_encode_block(DNXHDEncContext *ctx, DCTELEM *block, int last_index, int n)
257 int last_non_zero = 0;
260 dnxhd_encode_dc(ctx, block[0] - ctx->m.last_dc[n]);
261 ctx->m.last_dc[n] = block[0];
263 for (i = 1; i <= last_index; i++) {
264 j = ctx->m.intra_scantable.permutated[i];
267 int run_level = i - last_non_zero - 1;
268 int rlevel = (slevel<<1)|!!run_level;
269 put_bits(&ctx->m.pb, ctx->vlc_bits[rlevel], ctx->vlc_codes[rlevel]);
271 put_bits(&ctx->m.pb, ctx->run_bits[run_level], ctx->run_codes[run_level]);
275 put_bits(&ctx->m.pb, ctx->vlc_bits[0], ctx->vlc_codes[0]); // EOB
278 static av_always_inline void dnxhd_unquantize_c(DNXHDEncContext *ctx, DCTELEM *block, int n, int qscale, int last_index)
280 const uint8_t *weight_matrix;
284 weight_matrix = (n&2) ? ctx->cid_table->chroma_weight : ctx->cid_table->luma_weight;
286 for (i = 1; i <= last_index; i++) {
287 int j = ctx->m.intra_scantable.permutated[i];
291 level = (1-2*level) * qscale * weight_matrix[i];
292 if (weight_matrix[i] != 32)
297 level = (2*level+1) * qscale * weight_matrix[i];
298 if (weight_matrix[i] != 32)
307 static av_always_inline int dnxhd_ssd_block(DCTELEM *qblock, DCTELEM *block)
311 for (i = 0; i < 64; i++)
312 score += (block[i]-qblock[i])*(block[i]-qblock[i]);
316 static av_always_inline int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, DCTELEM *block, int last_index)
318 int last_non_zero = 0;
321 for (i = 1; i <= last_index; i++) {
322 j = ctx->m.intra_scantable.permutated[i];
325 int run_level = i - last_non_zero - 1;
326 bits += ctx->vlc_bits[(level<<1)|!!run_level]+ctx->run_bits[run_level];
333 static av_always_inline void dnxhd_get_pixels_4x8(DCTELEM *restrict block, const uint8_t *pixels, int line_size)
336 for (i = 0; i < 4; i++) {
337 block[0] = pixels[0]; block[1] = pixels[1];
338 block[2] = pixels[2]; block[3] = pixels[3];
339 block[4] = pixels[4]; block[5] = pixels[5];
340 block[6] = pixels[6]; block[7] = pixels[7];
344 memcpy(block , block- 8, sizeof(*block)*8);
345 memcpy(block+ 8, block-16, sizeof(*block)*8);
346 memcpy(block+16, block-24, sizeof(*block)*8);
347 memcpy(block+24, block-32, sizeof(*block)*8);
350 static av_always_inline void dnxhd_get_blocks(DNXHDEncContext *ctx, int mb_x, int mb_y)
352 const uint8_t *ptr_y = ctx->thread[0]->src[0] + ((mb_y << 4) * ctx->m.linesize) + (mb_x << 4);
353 const uint8_t *ptr_u = ctx->thread[0]->src[1] + ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << 3);
354 const uint8_t *ptr_v = ctx->thread[0]->src[2] + ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << 3);
355 DSPContext *dsp = &ctx->m.dsp;
357 dsp->get_pixels(ctx->blocks[0], ptr_y , ctx->m.linesize);
358 dsp->get_pixels(ctx->blocks[1], ptr_y + 8, ctx->m.linesize);
359 dsp->get_pixels(ctx->blocks[2], ptr_u , ctx->m.uvlinesize);
360 dsp->get_pixels(ctx->blocks[3], ptr_v , ctx->m.uvlinesize);
362 if (mb_y+1 == ctx->m.mb_height && ctx->m.avctx->height == 1080) {
363 if (ctx->interlaced) {
364 dnxhd_get_pixels_4x8(ctx->blocks[4], ptr_y + ctx->dct_y_offset , ctx->m.linesize);
365 dnxhd_get_pixels_4x8(ctx->blocks[5], ptr_y + ctx->dct_y_offset + 8, ctx->m.linesize);
366 dnxhd_get_pixels_4x8(ctx->blocks[6], ptr_u + ctx->dct_uv_offset , ctx->m.uvlinesize);
367 dnxhd_get_pixels_4x8(ctx->blocks[7], ptr_v + ctx->dct_uv_offset , ctx->m.uvlinesize);
369 dsp->clear_block(ctx->blocks[4]); dsp->clear_block(ctx->blocks[5]);
370 dsp->clear_block(ctx->blocks[6]); dsp->clear_block(ctx->blocks[7]);
373 dsp->get_pixels(ctx->blocks[4], ptr_y + ctx->dct_y_offset , ctx->m.linesize);
374 dsp->get_pixels(ctx->blocks[5], ptr_y + ctx->dct_y_offset + 8, ctx->m.linesize);
375 dsp->get_pixels(ctx->blocks[6], ptr_u + ctx->dct_uv_offset , ctx->m.uvlinesize);
376 dsp->get_pixels(ctx->blocks[7], ptr_v + ctx->dct_uv_offset , ctx->m.uvlinesize);
380 static av_always_inline int dnxhd_switch_matrix(DNXHDEncContext *ctx, int i)
383 ctx->m.q_intra_matrix16 = ctx->qmatrix_c16;
384 ctx->m.q_intra_matrix = ctx->qmatrix_c;
387 ctx->m.q_intra_matrix16 = ctx->qmatrix_l16;
388 ctx->m.q_intra_matrix = ctx->qmatrix_l;
393 static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg)
395 DNXHDEncContext *ctx = *(void**)arg;
397 int qscale = ctx->thread[0]->qscale;
399 for (mb_y = ctx->m.start_mb_y; mb_y < ctx->m.end_mb_y; mb_y++) {
402 ctx->m.last_dc[2] = 1024;
404 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
405 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
411 dnxhd_get_blocks(ctx, mb_x, mb_y);
413 for (i = 0; i < 8; i++) {
414 DECLARE_ALIGNED_16(DCTELEM, block[64]);
415 DCTELEM *src_block = ctx->blocks[i];
416 int overflow, nbits, diff, last_index;
417 int n = dnxhd_switch_matrix(ctx, i);
419 memcpy(block, src_block, sizeof(block));
420 last_index = ctx->m.dct_quantize((MpegEncContext*)ctx, block, i, qscale, &overflow);
421 ac_bits += dnxhd_calc_ac_bits(ctx, block, last_index);
423 diff = block[0] - ctx->m.last_dc[n];
424 if (diff < 0) nbits = av_log2_16bit(-2*diff);
425 else nbits = av_log2_16bit( 2*diff);
426 dc_bits += ctx->cid_table->dc_bits[nbits] + nbits;
428 ctx->m.last_dc[n] = block[0];
430 if (avctx->mb_decision == FF_MB_DECISION_RD || !RC_VARIANCE) {
431 dnxhd_unquantize_c(ctx, block, i, qscale, last_index);
432 ctx->m.dsp.idct(block);
433 ssd += dnxhd_ssd_block(block, src_block);
436 ctx->mb_rc[qscale][mb].ssd = ssd;
437 ctx->mb_rc[qscale][mb].bits = ac_bits+dc_bits+12+8*ctx->vlc_bits[0];
443 static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg)
445 DNXHDEncContext *ctx = *(void**)arg;
448 for (mb_y = ctx->m.start_mb_y; mb_y < ctx->m.end_mb_y; mb_y++) {
451 ctx->m.last_dc[2] = 1024;
452 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
453 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
454 int qscale = ctx->mb_qscale[mb];
457 put_bits(&ctx->m.pb, 12, qscale<<1);
459 dnxhd_get_blocks(ctx, mb_x, mb_y);
461 for (i = 0; i < 8; i++) {
462 DCTELEM *block = ctx->blocks[i];
463 int last_index, overflow;
464 int n = dnxhd_switch_matrix(ctx, i);
465 last_index = ctx->m.dct_quantize((MpegEncContext*)ctx, block, i, qscale, &overflow);
467 dnxhd_encode_block(ctx, block, last_index, n);
468 //STOP_TIMER("encode_block");
471 if (put_bits_count(&ctx->m.pb)&31)
472 put_bits(&ctx->m.pb, 32-(put_bits_count(&ctx->m.pb)&31), 0);
474 flush_put_bits(&ctx->m.pb);
478 static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx, uint8_t *buf)
482 for (i = 0; i < ctx->m.avctx->thread_count; i++) {
484 for (mb_y = ctx->thread[i]->m.start_mb_y; mb_y < ctx->thread[i]->m.end_mb_y; mb_y++) {
485 ctx->slice_size[mb_y] = 0;
486 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
487 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
488 ctx->slice_size[mb_y] += ctx->mb_bits[mb];
490 ctx->slice_size[mb_y] = (ctx->slice_size[mb_y]+31)&~31;
491 ctx->slice_size[mb_y] >>= 3;
492 thread_size += ctx->slice_size[mb_y];
494 init_put_bits(&ctx->thread[i]->m.pb, buf + 640 + offset, thread_size);
495 offset += thread_size;
499 static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg)
501 DNXHDEncContext *ctx = *(void**)arg;
503 for (mb_y = ctx->m.start_mb_y; mb_y < ctx->m.end_mb_y; mb_y++) {
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;
516 static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx)
518 int lambda, up_step, down_step;
519 int last_lower = INT_MAX, last_higher = 0;
522 for (q = 1; q < avctx->qmax; q++) {
524 avctx->execute(avctx, dnxhd_calc_bits_thread, (void**)&ctx->thread[0], NULL, avctx->thread_count, sizeof(void*));
526 up_step = down_step = 2<<LAMBDA_FRAC_BITS;
527 lambda = ctx->lambda;
532 if (lambda == last_higher) {
534 end = 1; // need to set final qscales/bits
536 for (y = 0; y < ctx->m.mb_height; y++) {
537 for (x = 0; x < ctx->m.mb_width; x++) {
538 unsigned min = UINT_MAX;
540 int mb = y*ctx->m.mb_width+x;
541 for (q = 1; q < avctx->qmax; q++) {
542 unsigned score = ctx->mb_rc[q][mb].bits*lambda+(ctx->mb_rc[q][mb].ssd<<LAMBDA_FRAC_BITS);
548 bits += ctx->mb_rc[qscale][mb].bits;
549 ctx->mb_qscale[mb] = qscale;
550 ctx->mb_bits[mb] = ctx->mb_rc[qscale][mb].bits;
552 bits = (bits+31)&~31; // padding
553 if (bits > ctx->frame_bits)
556 //dprintf(ctx->m.avctx, "lambda %d, up %u, down %u, bits %d, frame %d\n",
557 // lambda, last_higher, last_lower, bits, ctx->frame_bits);
559 if (bits > ctx->frame_bits)
563 if (bits < ctx->frame_bits) {
564 last_lower = FFMIN(lambda, last_lower);
565 if (last_higher != 0)
566 lambda = (lambda+last_higher)>>1;
569 down_step *= 5; // XXX tune ?
570 up_step = 1<<LAMBDA_FRAC_BITS;
571 lambda = FFMAX(1, lambda);
572 if (lambda == last_lower)
575 last_higher = FFMAX(lambda, last_higher);
576 if (last_lower != INT_MAX)
577 lambda = (lambda+last_lower)>>1;
581 down_step = 1<<LAMBDA_FRAC_BITS;
584 //dprintf(ctx->m.avctx, "out lambda %d\n", lambda);
585 ctx->lambda = lambda;
589 static int dnxhd_find_qscale(DNXHDEncContext *ctx)
595 int last_lower = INT_MAX;
599 qscale = ctx->qscale;
602 ctx->qscale = qscale;
603 // XXX avoid recalculating bits
604 ctx->m.avctx->execute(ctx->m.avctx, dnxhd_calc_bits_thread, (void**)&ctx->thread[0], NULL, ctx->m.avctx->thread_count, sizeof(void*));
605 for (y = 0; y < ctx->m.mb_height; y++) {
606 for (x = 0; x < ctx->m.mb_width; x++)
607 bits += ctx->mb_rc[qscale][y*ctx->m.mb_width+x].bits;
608 bits = (bits+31)&~31; // padding
609 if (bits > ctx->frame_bits)
612 //dprintf(ctx->m.avctx, "%d, qscale %d, bits %d, frame %d, higher %d, lower %d\n",
613 // ctx->m.avctx->frame_number, qscale, bits, ctx->frame_bits, last_higher, last_lower);
614 if (bits < ctx->frame_bits) {
617 if (last_higher == qscale - 1) {
618 qscale = last_higher;
621 last_lower = FFMIN(qscale, last_lower);
622 if (last_higher != 0)
623 qscale = (qscale+last_higher)>>1;
625 qscale -= down_step++;
630 if (last_lower == qscale + 1)
632 last_higher = FFMAX(qscale, last_higher);
633 if (last_lower != INT_MAX)
634 qscale = (qscale+last_lower)>>1;
638 if (qscale >= ctx->m.avctx->qmax)
642 //dprintf(ctx->m.avctx, "out qscale %d\n", qscale);
643 ctx->qscale = qscale;
647 static int dnxhd_rc_cmp(const void *a, const void *b)
649 return ((const RCCMPEntry *)b)->value - ((const RCCMPEntry *)a)->value;
652 static int dnxhd_encode_fast(AVCodecContext *avctx, DNXHDEncContext *ctx)
656 if ((ret = dnxhd_find_qscale(ctx)) < 0)
658 for (y = 0; y < ctx->m.mb_height; y++) {
659 for (x = 0; x < ctx->m.mb_width; x++) {
660 int mb = y*ctx->m.mb_width+x;
662 ctx->mb_qscale[mb] = ctx->qscale;
663 ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale][mb].bits;
664 max_bits += ctx->mb_rc[ctx->qscale][mb].bits;
666 delta_bits = ctx->mb_rc[ctx->qscale][mb].bits-ctx->mb_rc[ctx->qscale+1][mb].bits;
667 ctx->mb_cmp[mb].mb = mb;
668 ctx->mb_cmp[mb].value = delta_bits ?
669 ((ctx->mb_rc[ctx->qscale][mb].ssd-ctx->mb_rc[ctx->qscale+1][mb].ssd)*100)/delta_bits
670 : INT_MIN; //avoid increasing qscale
673 max_bits += 31; //worst padding
677 avctx->execute(avctx, dnxhd_mb_var_thread, (void**)&ctx->thread[0], NULL, avctx->thread_count, sizeof(void*));
678 qsort(ctx->mb_cmp, ctx->m.mb_num, sizeof(RCEntry), dnxhd_rc_cmp);
679 for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) {
680 int mb = ctx->mb_cmp[x].mb;
681 max_bits -= ctx->mb_rc[ctx->qscale][mb].bits - ctx->mb_rc[ctx->qscale+1][mb].bits;
682 ctx->mb_qscale[mb] = ctx->qscale+1;
683 ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale+1][mb].bits;
689 static void dnxhd_load_picture(DNXHDEncContext *ctx, const AVFrame *frame)
693 for (i = 0; i < 3; i++) {
694 ctx->frame.data[i] = frame->data[i];
695 ctx->frame.linesize[i] = frame->linesize[i];
698 for (i = 0; i < ctx->m.avctx->thread_count; i++) {
699 ctx->thread[i]->m.linesize = ctx->frame.linesize[0]<<ctx->interlaced;
700 ctx->thread[i]->m.uvlinesize = ctx->frame.linesize[1]<<ctx->interlaced;
701 ctx->thread[i]->dct_y_offset = ctx->m.linesize *8;
702 ctx->thread[i]->dct_uv_offset = ctx->m.uvlinesize*8;
705 ctx->frame.interlaced_frame = frame->interlaced_frame;
706 ctx->cur_field = frame->interlaced_frame && !frame->top_field_first;
709 static int dnxhd_encode_picture(AVCodecContext *avctx, unsigned char *buf, int buf_size, const void *data)
711 DNXHDEncContext *ctx = avctx->priv_data;
715 if (buf_size < ctx->cid_table->frame_size) {
716 av_log(avctx, AV_LOG_ERROR, "output buffer is too small to compress picture\n");
720 dnxhd_load_picture(ctx, data);
723 for (i = 0; i < 3; i++) {
724 ctx->src[i] = ctx->frame.data[i];
725 if (ctx->interlaced && ctx->cur_field)
726 ctx->src[i] += ctx->frame.linesize[i];
729 dnxhd_write_header(avctx, buf);
731 if (avctx->mb_decision == FF_MB_DECISION_RD)
732 ret = dnxhd_encode_rdo(avctx, ctx);
734 ret = dnxhd_encode_fast(avctx, ctx);
736 av_log(avctx, AV_LOG_ERROR, "picture could not fit ratecontrol constraints\n");
740 dnxhd_setup_threads_slices(ctx, buf);
743 for (i = 0; i < ctx->m.mb_height; i++) {
744 AV_WB32(ctx->msip + i * 4, offset);
745 offset += ctx->slice_size[i];
746 assert(!(ctx->slice_size[i] & 3));
749 avctx->execute(avctx, dnxhd_encode_thread, (void**)&ctx->thread[0], NULL, avctx->thread_count, sizeof(void*));
751 AV_WB32(buf + ctx->cid_table->coding_unit_size - 4, 0x600DC0DE); // EOF
753 if (ctx->interlaced && first_field) {
756 buf += ctx->cid_table->coding_unit_size;
757 buf_size -= ctx->cid_table->coding_unit_size;
758 goto encode_coding_unit;
761 ctx->frame.quality = ctx->qscale*FF_QP2LAMBDA;
763 return ctx->cid_table->frame_size;
766 static int dnxhd_encode_end(AVCodecContext *avctx)
768 DNXHDEncContext *ctx = avctx->priv_data;
769 int max_level = 1<<(ctx->cid_table->bit_depth+2);
772 av_free(ctx->vlc_codes-max_level*2);
773 av_free(ctx->vlc_bits -max_level*2);
774 av_freep(&ctx->run_codes);
775 av_freep(&ctx->run_bits);
777 av_freep(&ctx->mb_bits);
778 av_freep(&ctx->mb_qscale);
779 av_freep(&ctx->mb_rc);
780 av_freep(&ctx->mb_cmp);
781 av_freep(&ctx->slice_size);
783 av_freep(&ctx->qmatrix_c);
784 av_freep(&ctx->qmatrix_l);
785 av_freep(&ctx->qmatrix_c16);
786 av_freep(&ctx->qmatrix_l16);
788 for (i = 1; i < avctx->thread_count; i++)
789 av_freep(&ctx->thread[i]);
794 AVCodec dnxhd_encoder = {
798 sizeof(DNXHDEncContext),
800 dnxhd_encode_picture,
802 .pix_fmts = (enum PixelFormat[]){PIX_FMT_YUV422P, PIX_FMT_NONE},
803 .long_name = NULL_IF_CONFIG_SMALL("VC3/DNxHD"),