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 Libav.
9 * Libav 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 * Libav 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 Libav; 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
27 #include "libavutil/opt.h"
30 #include "mpegvideo.h"
31 #include "mpegvideo_common.h"
34 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
36 static const AVOption options[]={
37 {"nitris_compat", "encode with Avid Nitris compatibility", offsetof(DNXHDEncContext, nitris_compat), FF_OPT_TYPE_INT, {.dbl = 0}, 0, 1, VE},
40 static const AVClass class = { "dnxhd", av_default_item_name, options, LIBAVUTIL_VERSION_INT };
42 #define LAMBDA_FRAC_BITS 10
44 static av_always_inline void dnxhd_get_pixels_8x4(DCTELEM *restrict block, const uint8_t *pixels, int line_size)
47 for (i = 0; i < 4; i++) {
48 block[0] = pixels[0]; block[1] = pixels[1];
49 block[2] = pixels[2]; block[3] = pixels[3];
50 block[4] = pixels[4]; block[5] = pixels[5];
51 block[6] = pixels[6]; block[7] = pixels[7];
55 memcpy(block , block- 8, sizeof(*block)*8);
56 memcpy(block+ 8, block-16, sizeof(*block)*8);
57 memcpy(block+16, block-24, sizeof(*block)*8);
58 memcpy(block+24, block-32, sizeof(*block)*8);
61 static int dnxhd_init_vlc(DNXHDEncContext *ctx)
64 int max_level = 1<<(ctx->cid_table->bit_depth+2);
66 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->vlc_codes, max_level*4*sizeof(*ctx->vlc_codes), fail);
67 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->vlc_bits , max_level*4*sizeof(*ctx->vlc_bits ), fail);
68 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_codes, 63*2 , fail);
69 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_bits , 63 , fail);
71 ctx->vlc_codes += max_level*2;
72 ctx->vlc_bits += max_level*2;
73 for (level = -max_level; level < max_level; level++) {
74 for (run = 0; run < 2; run++) {
75 int index = (level<<1)|run;
76 int sign, offset = 0, alevel = level;
78 MASK_ABS(sign, alevel);
80 offset = (alevel-1)>>6;
83 for (j = 0; j < 257; j++) {
84 if (ctx->cid_table->ac_level[j] == alevel &&
85 (!offset || (ctx->cid_table->ac_index_flag[j] && offset)) &&
86 (!run || (ctx->cid_table->ac_run_flag [j] && run))) {
87 assert(!ctx->vlc_codes[index]);
89 ctx->vlc_codes[index] = (ctx->cid_table->ac_codes[j]<<1)|(sign&1);
90 ctx->vlc_bits [index] = ctx->cid_table->ac_bits[j]+1;
92 ctx->vlc_codes[index] = ctx->cid_table->ac_codes[j];
93 ctx->vlc_bits [index] = ctx->cid_table->ac_bits [j];
98 assert(!alevel || j < 257);
100 ctx->vlc_codes[index] = (ctx->vlc_codes[index]<<ctx->cid_table->index_bits)|offset;
101 ctx->vlc_bits [index]+= ctx->cid_table->index_bits;
105 for (i = 0; i < 62; i++) {
106 int run = ctx->cid_table->run[i];
108 ctx->run_codes[run] = ctx->cid_table->run_codes[i];
109 ctx->run_bits [run] = ctx->cid_table->run_bits[i];
116 static int dnxhd_init_qmat(DNXHDEncContext *ctx, int lbias, int cbias)
118 // init first elem to 1 to avoid div by 0 in convert_matrix
119 uint16_t weight_matrix[64] = {1,}; // convert_matrix needs uint16_t*
122 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l, (ctx->m.avctx->qmax+1) * 64 * sizeof(int) , fail);
123 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c, (ctx->m.avctx->qmax+1) * 64 * sizeof(int) , fail);
124 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l16, (ctx->m.avctx->qmax+1) * 64 * 2 * sizeof(uint16_t), fail);
125 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c16, (ctx->m.avctx->qmax+1) * 64 * 2 * sizeof(uint16_t), fail);
127 for (i = 1; i < 64; i++) {
128 int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]];
129 weight_matrix[j] = ctx->cid_table->luma_weight[i];
131 ff_convert_matrix(&ctx->m.dsp, ctx->qmatrix_l, ctx->qmatrix_l16, weight_matrix,
132 ctx->m.intra_quant_bias, 1, ctx->m.avctx->qmax, 1);
133 for (i = 1; i < 64; i++) {
134 int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]];
135 weight_matrix[j] = ctx->cid_table->chroma_weight[i];
137 ff_convert_matrix(&ctx->m.dsp, ctx->qmatrix_c, ctx->qmatrix_c16, weight_matrix,
138 ctx->m.intra_quant_bias, 1, ctx->m.avctx->qmax, 1);
139 for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
140 for (i = 0; i < 64; i++) {
141 ctx->qmatrix_l [qscale] [i] <<= 2; ctx->qmatrix_c [qscale] [i] <<= 2;
142 ctx->qmatrix_l16[qscale][0][i] <<= 2; ctx->qmatrix_l16[qscale][1][i] <<= 2;
143 ctx->qmatrix_c16[qscale][0][i] <<= 2; ctx->qmatrix_c16[qscale][1][i] <<= 2;
151 static int dnxhd_init_rc(DNXHDEncContext *ctx)
153 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_rc, 8160*ctx->m.avctx->qmax*sizeof(RCEntry), fail);
154 if (ctx->m.avctx->mb_decision != FF_MB_DECISION_RD)
155 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_cmp, ctx->m.mb_num*sizeof(RCCMPEntry), fail);
157 ctx->frame_bits = (ctx->cid_table->coding_unit_size - 640 - 4 - ctx->min_padding) * 8;
159 ctx->lambda = 2<<LAMBDA_FRAC_BITS; // qscale 2
165 static int dnxhd_encode_init(AVCodecContext *avctx)
167 DNXHDEncContext *ctx = avctx->priv_data;
170 ctx->cid = ff_dnxhd_find_cid(avctx);
171 if (!ctx->cid || avctx->pix_fmt != PIX_FMT_YUV422P) {
172 av_log(avctx, AV_LOG_ERROR, "video parameters incompatible with DNxHD\n");
175 av_log(avctx, AV_LOG_DEBUG, "cid %d\n", ctx->cid);
177 index = ff_dnxhd_get_cid_table(ctx->cid);
178 ctx->cid_table = &ff_dnxhd_cid_table[index];
180 ctx->m.avctx = avctx;
184 ctx->get_pixels_8x4_sym = dnxhd_get_pixels_8x4;
186 dsputil_init(&ctx->m.dsp, avctx);
187 ff_dct_common_init(&ctx->m);
189 ff_dnxhd_init_mmx(ctx);
191 if (!ctx->m.dct_quantize)
192 ctx->m.dct_quantize = dct_quantize_c;
194 ctx->m.mb_height = (avctx->height + 15) / 16;
195 ctx->m.mb_width = (avctx->width + 15) / 16;
197 if (avctx->flags & CODEC_FLAG_INTERLACED_DCT) {
199 ctx->m.mb_height /= 2;
202 ctx->m.mb_num = ctx->m.mb_height * ctx->m.mb_width;
204 if (avctx->intra_quant_bias != FF_DEFAULT_QUANT_BIAS)
205 ctx->m.intra_quant_bias = avctx->intra_quant_bias;
206 if (dnxhd_init_qmat(ctx, ctx->m.intra_quant_bias, 0) < 0) // XXX tune lbias/cbias
209 // Avid Nitris hardware decoder requires a minimum amount of padding in the coding unit payload
210 if (ctx->nitris_compat)
211 ctx->min_padding = 1600;
213 if (dnxhd_init_vlc(ctx) < 0)
215 if (dnxhd_init_rc(ctx) < 0)
218 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_size, ctx->m.mb_height*sizeof(uint32_t), fail);
219 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_offs, ctx->m.mb_height*sizeof(uint32_t), fail);
220 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_bits, ctx->m.mb_num *sizeof(uint16_t), fail);
221 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_qscale, ctx->m.mb_num *sizeof(uint8_t) , fail);
223 ctx->frame.key_frame = 1;
224 ctx->frame.pict_type = AV_PICTURE_TYPE_I;
225 ctx->m.avctx->coded_frame = &ctx->frame;
227 if (avctx->thread_count > MAX_THREADS) {
228 av_log(avctx, AV_LOG_ERROR, "too many threads\n");
232 ctx->thread[0] = ctx;
233 for (i = 1; i < avctx->thread_count; i++) {
234 ctx->thread[i] = av_malloc(sizeof(DNXHDEncContext));
235 memcpy(ctx->thread[i], ctx, sizeof(DNXHDEncContext));
239 fail: //for FF_ALLOCZ_OR_GOTO
243 static int dnxhd_write_header(AVCodecContext *avctx, uint8_t *buf)
245 DNXHDEncContext *ctx = avctx->priv_data;
246 const uint8_t header_prefix[5] = { 0x00,0x00,0x02,0x80,0x01 };
250 memcpy(buf, header_prefix, 5);
251 buf[5] = ctx->interlaced ? ctx->cur_field+2 : 0x01;
252 buf[6] = 0x80; // crc flag off
253 buf[7] = 0xa0; // reserved
254 AV_WB16(buf + 0x18, avctx->height>>ctx->interlaced); // ALPF
255 AV_WB16(buf + 0x1a, avctx->width); // SPL
256 AV_WB16(buf + 0x1d, avctx->height>>ctx->interlaced); // NAL
258 buf[0x21] = 0x38; // FIXME 8 bit per comp
259 buf[0x22] = 0x88 + (ctx->interlaced<<2);
260 AV_WB32(buf + 0x28, ctx->cid); // CID
261 buf[0x2c] = ctx->interlaced ? 0 : 0x80;
263 buf[0x5f] = 0x01; // UDL
265 buf[0x167] = 0x02; // reserved
266 AV_WB16(buf + 0x16a, ctx->m.mb_height * 4 + 4); // MSIPS
267 buf[0x16d] = ctx->m.mb_height; // Ns
268 buf[0x16f] = 0x10; // reserved
270 ctx->msip = buf + 0x170;
274 static av_always_inline void dnxhd_encode_dc(DNXHDEncContext *ctx, int diff)
278 nbits = av_log2_16bit(-2*diff);
281 nbits = av_log2_16bit(2*diff);
283 put_bits(&ctx->m.pb, ctx->cid_table->dc_bits[nbits] + nbits,
284 (ctx->cid_table->dc_codes[nbits]<<nbits) + (diff & ((1 << nbits) - 1)));
287 static av_always_inline void dnxhd_encode_block(DNXHDEncContext *ctx, DCTELEM *block, int last_index, int n)
289 int last_non_zero = 0;
292 dnxhd_encode_dc(ctx, block[0] - ctx->m.last_dc[n]);
293 ctx->m.last_dc[n] = block[0];
295 for (i = 1; i <= last_index; i++) {
296 j = ctx->m.intra_scantable.permutated[i];
299 int run_level = i - last_non_zero - 1;
300 int rlevel = (slevel<<1)|!!run_level;
301 put_bits(&ctx->m.pb, ctx->vlc_bits[rlevel], ctx->vlc_codes[rlevel]);
303 put_bits(&ctx->m.pb, ctx->run_bits[run_level], ctx->run_codes[run_level]);
307 put_bits(&ctx->m.pb, ctx->vlc_bits[0], ctx->vlc_codes[0]); // EOB
310 static av_always_inline void dnxhd_unquantize_c(DNXHDEncContext *ctx, DCTELEM *block, int n, int qscale, int last_index)
312 const uint8_t *weight_matrix;
316 weight_matrix = (n&2) ? ctx->cid_table->chroma_weight : ctx->cid_table->luma_weight;
318 for (i = 1; i <= last_index; i++) {
319 int j = ctx->m.intra_scantable.permutated[i];
323 level = (1-2*level) * qscale * weight_matrix[i];
324 if (weight_matrix[i] != 32)
329 level = (2*level+1) * qscale * weight_matrix[i];
330 if (weight_matrix[i] != 32)
339 static av_always_inline int dnxhd_ssd_block(DCTELEM *qblock, DCTELEM *block)
343 for (i = 0; i < 64; i++)
344 score += (block[i]-qblock[i])*(block[i]-qblock[i]);
348 static av_always_inline int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, DCTELEM *block, int last_index)
350 int last_non_zero = 0;
353 for (i = 1; i <= last_index; i++) {
354 j = ctx->m.intra_scantable.permutated[i];
357 int run_level = i - last_non_zero - 1;
358 bits += ctx->vlc_bits[(level<<1)|!!run_level]+ctx->run_bits[run_level];
365 static av_always_inline void dnxhd_get_blocks(DNXHDEncContext *ctx, int mb_x, int mb_y)
367 const uint8_t *ptr_y = ctx->thread[0]->src[0] + ((mb_y << 4) * ctx->m.linesize) + (mb_x << 4);
368 const uint8_t *ptr_u = ctx->thread[0]->src[1] + ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << 3);
369 const uint8_t *ptr_v = ctx->thread[0]->src[2] + ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << 3);
370 DSPContext *dsp = &ctx->m.dsp;
372 dsp->get_pixels(ctx->blocks[0], ptr_y , ctx->m.linesize);
373 dsp->get_pixels(ctx->blocks[1], ptr_y + 8, ctx->m.linesize);
374 dsp->get_pixels(ctx->blocks[2], ptr_u , ctx->m.uvlinesize);
375 dsp->get_pixels(ctx->blocks[3], ptr_v , ctx->m.uvlinesize);
377 if (mb_y+1 == ctx->m.mb_height && ctx->m.avctx->height == 1080) {
378 if (ctx->interlaced) {
379 ctx->get_pixels_8x4_sym(ctx->blocks[4], ptr_y + ctx->dct_y_offset , ctx->m.linesize);
380 ctx->get_pixels_8x4_sym(ctx->blocks[5], ptr_y + ctx->dct_y_offset + 8, ctx->m.linesize);
381 ctx->get_pixels_8x4_sym(ctx->blocks[6], ptr_u + ctx->dct_uv_offset , ctx->m.uvlinesize);
382 ctx->get_pixels_8x4_sym(ctx->blocks[7], ptr_v + ctx->dct_uv_offset , ctx->m.uvlinesize);
384 dsp->clear_block(ctx->blocks[4]); dsp->clear_block(ctx->blocks[5]);
385 dsp->clear_block(ctx->blocks[6]); dsp->clear_block(ctx->blocks[7]);
388 dsp->get_pixels(ctx->blocks[4], ptr_y + ctx->dct_y_offset , ctx->m.linesize);
389 dsp->get_pixels(ctx->blocks[5], ptr_y + ctx->dct_y_offset + 8, ctx->m.linesize);
390 dsp->get_pixels(ctx->blocks[6], ptr_u + ctx->dct_uv_offset , ctx->m.uvlinesize);
391 dsp->get_pixels(ctx->blocks[7], ptr_v + ctx->dct_uv_offset , ctx->m.uvlinesize);
395 static av_always_inline int dnxhd_switch_matrix(DNXHDEncContext *ctx, int i)
398 ctx->m.q_intra_matrix16 = ctx->qmatrix_c16;
399 ctx->m.q_intra_matrix = ctx->qmatrix_c;
402 ctx->m.q_intra_matrix16 = ctx->qmatrix_l16;
403 ctx->m.q_intra_matrix = ctx->qmatrix_l;
408 static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg, int jobnr, int threadnr)
410 DNXHDEncContext *ctx = avctx->priv_data;
411 int mb_y = jobnr, mb_x;
412 int qscale = ctx->qscale;
413 LOCAL_ALIGNED_16(DCTELEM, block, [64]);
414 ctx = ctx->thread[threadnr];
418 ctx->m.last_dc[2] = 1024;
420 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
421 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
427 dnxhd_get_blocks(ctx, mb_x, mb_y);
429 for (i = 0; i < 8; i++) {
430 DCTELEM *src_block = ctx->blocks[i];
431 int overflow, nbits, diff, last_index;
432 int n = dnxhd_switch_matrix(ctx, i);
434 memcpy(block, src_block, 64*sizeof(*block));
435 last_index = ctx->m.dct_quantize(&ctx->m, block, i, qscale, &overflow);
436 ac_bits += dnxhd_calc_ac_bits(ctx, block, last_index);
438 diff = block[0] - ctx->m.last_dc[n];
439 if (diff < 0) nbits = av_log2_16bit(-2*diff);
440 else nbits = av_log2_16bit( 2*diff);
441 dc_bits += ctx->cid_table->dc_bits[nbits] + nbits;
443 ctx->m.last_dc[n] = block[0];
445 if (avctx->mb_decision == FF_MB_DECISION_RD || !RC_VARIANCE) {
446 dnxhd_unquantize_c(ctx, block, i, qscale, last_index);
447 ctx->m.dsp.idct(block);
448 ssd += dnxhd_ssd_block(block, src_block);
451 ctx->mb_rc[qscale][mb].ssd = ssd;
452 ctx->mb_rc[qscale][mb].bits = ac_bits+dc_bits+12+8*ctx->vlc_bits[0];
457 static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg, int jobnr, int threadnr)
459 DNXHDEncContext *ctx = avctx->priv_data;
460 int mb_y = jobnr, mb_x;
461 ctx = ctx->thread[threadnr];
462 init_put_bits(&ctx->m.pb, (uint8_t *)arg + 640 + ctx->slice_offs[jobnr], ctx->slice_size[jobnr]);
466 ctx->m.last_dc[2] = 1024;
467 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
468 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
469 int qscale = ctx->mb_qscale[mb];
472 put_bits(&ctx->m.pb, 12, qscale<<1);
474 dnxhd_get_blocks(ctx, mb_x, mb_y);
476 for (i = 0; i < 8; i++) {
477 DCTELEM *block = ctx->blocks[i];
478 int last_index, overflow;
479 int n = dnxhd_switch_matrix(ctx, i);
480 last_index = ctx->m.dct_quantize(&ctx->m, block, i, qscale, &overflow);
482 dnxhd_encode_block(ctx, block, last_index, n);
483 //STOP_TIMER("encode_block");
486 if (put_bits_count(&ctx->m.pb)&31)
487 put_bits(&ctx->m.pb, 32-(put_bits_count(&ctx->m.pb)&31), 0);
488 flush_put_bits(&ctx->m.pb);
492 static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx)
496 for (mb_y = 0; mb_y < ctx->m.mb_height; mb_y++) {
498 ctx->slice_offs[mb_y] = offset;
499 ctx->slice_size[mb_y] = 0;
500 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
501 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
502 ctx->slice_size[mb_y] += ctx->mb_bits[mb];
504 ctx->slice_size[mb_y] = (ctx->slice_size[mb_y]+31)&~31;
505 ctx->slice_size[mb_y] >>= 3;
506 thread_size = ctx->slice_size[mb_y];
507 offset += thread_size;
511 static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg, int jobnr, int threadnr)
513 DNXHDEncContext *ctx = avctx->priv_data;
514 int mb_y = jobnr, mb_x;
515 ctx = ctx->thread[threadnr];
516 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
517 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
518 uint8_t *pix = ctx->thread[0]->src[0] + ((mb_y<<4) * ctx->m.linesize) + (mb_x<<4);
519 int sum = ctx->m.dsp.pix_sum(pix, ctx->m.linesize);
520 int varc = (ctx->m.dsp.pix_norm1(pix, ctx->m.linesize) - (((unsigned)(sum*sum))>>8)+128)>>8;
521 ctx->mb_cmp[mb].value = varc;
522 ctx->mb_cmp[mb].mb = mb;
527 static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx)
529 int lambda, up_step, down_step;
530 int last_lower = INT_MAX, last_higher = 0;
533 for (q = 1; q < avctx->qmax; q++) {
535 avctx->execute2(avctx, dnxhd_calc_bits_thread, NULL, NULL, ctx->m.mb_height);
537 up_step = down_step = 2<<LAMBDA_FRAC_BITS;
538 lambda = ctx->lambda;
543 if (lambda == last_higher) {
545 end = 1; // need to set final qscales/bits
547 for (y = 0; y < ctx->m.mb_height; y++) {
548 for (x = 0; x < ctx->m.mb_width; x++) {
549 unsigned min = UINT_MAX;
551 int mb = y*ctx->m.mb_width+x;
552 for (q = 1; q < avctx->qmax; q++) {
553 unsigned score = ctx->mb_rc[q][mb].bits*lambda+(ctx->mb_rc[q][mb].ssd<<LAMBDA_FRAC_BITS);
559 bits += ctx->mb_rc[qscale][mb].bits;
560 ctx->mb_qscale[mb] = qscale;
561 ctx->mb_bits[mb] = ctx->mb_rc[qscale][mb].bits;
563 bits = (bits+31)&~31; // padding
564 if (bits > ctx->frame_bits)
567 //av_dlog(ctx->m.avctx, "lambda %d, up %u, down %u, bits %d, frame %d\n",
568 // lambda, last_higher, last_lower, bits, ctx->frame_bits);
570 if (bits > ctx->frame_bits)
574 if (bits < ctx->frame_bits) {
575 last_lower = FFMIN(lambda, last_lower);
576 if (last_higher != 0)
577 lambda = (lambda+last_higher)>>1;
580 down_step *= 5; // XXX tune ?
581 up_step = 1<<LAMBDA_FRAC_BITS;
582 lambda = FFMAX(1, lambda);
583 if (lambda == last_lower)
586 last_higher = FFMAX(lambda, last_higher);
587 if (last_lower != INT_MAX)
588 lambda = (lambda+last_lower)>>1;
589 else if ((int64_t)lambda + up_step > INT_MAX)
593 up_step = FFMIN((int64_t)up_step*5, INT_MAX);
594 down_step = 1<<LAMBDA_FRAC_BITS;
597 //av_dlog(ctx->m.avctx, "out lambda %d\n", lambda);
598 ctx->lambda = lambda;
602 static int dnxhd_find_qscale(DNXHDEncContext *ctx)
608 int last_lower = INT_MAX;
612 qscale = ctx->qscale;
615 ctx->qscale = qscale;
616 // XXX avoid recalculating bits
617 ctx->m.avctx->execute2(ctx->m.avctx, dnxhd_calc_bits_thread, NULL, NULL, ctx->m.mb_height);
618 for (y = 0; y < ctx->m.mb_height; y++) {
619 for (x = 0; x < ctx->m.mb_width; x++)
620 bits += ctx->mb_rc[qscale][y*ctx->m.mb_width+x].bits;
621 bits = (bits+31)&~31; // padding
622 if (bits > ctx->frame_bits)
625 //av_dlog(ctx->m.avctx, "%d, qscale %d, bits %d, frame %d, higher %d, lower %d\n",
626 // ctx->m.avctx->frame_number, qscale, bits, ctx->frame_bits, last_higher, last_lower);
627 if (bits < ctx->frame_bits) {
630 if (last_higher == qscale - 1) {
631 qscale = last_higher;
634 last_lower = FFMIN(qscale, last_lower);
635 if (last_higher != 0)
636 qscale = (qscale+last_higher)>>1;
638 qscale -= down_step++;
643 if (last_lower == qscale + 1)
645 last_higher = FFMAX(qscale, last_higher);
646 if (last_lower != INT_MAX)
647 qscale = (qscale+last_lower)>>1;
651 if (qscale >= ctx->m.avctx->qmax)
655 //av_dlog(ctx->m.avctx, "out qscale %d\n", qscale);
656 ctx->qscale = qscale;
660 #define BUCKET_BITS 8
661 #define RADIX_PASSES 4
662 #define NBUCKETS (1 << BUCKET_BITS)
664 static inline int get_bucket(int value, int shift)
667 value &= NBUCKETS - 1;
668 return NBUCKETS - 1 - value;
671 static void radix_count(const RCCMPEntry *data, int size, int buckets[RADIX_PASSES][NBUCKETS])
674 memset(buckets, 0, sizeof(buckets[0][0]) * RADIX_PASSES * NBUCKETS);
675 for (i = 0; i < size; i++) {
676 int v = data[i].value;
677 for (j = 0; j < RADIX_PASSES; j++) {
678 buckets[j][get_bucket(v, 0)]++;
683 for (j = 0; j < RADIX_PASSES; j++) {
685 for (i = NBUCKETS - 1; i >= 0; i--)
686 buckets[j][i] = offset -= buckets[j][i];
687 assert(!buckets[j][0]);
691 static void radix_sort_pass(RCCMPEntry *dst, const RCCMPEntry *data, int size, int buckets[NBUCKETS], int pass)
693 int shift = pass * BUCKET_BITS;
695 for (i = 0; i < size; i++) {
696 int v = get_bucket(data[i].value, shift);
697 int pos = buckets[v]++;
702 static void radix_sort(RCCMPEntry *data, int size)
704 int buckets[RADIX_PASSES][NBUCKETS];
705 RCCMPEntry *tmp = av_malloc(sizeof(*tmp) * size);
706 radix_count(data, size, buckets);
707 radix_sort_pass(tmp, data, size, buckets[0], 0);
708 radix_sort_pass(data, tmp, size, buckets[1], 1);
709 if (buckets[2][NBUCKETS - 1] || buckets[3][NBUCKETS - 1]) {
710 radix_sort_pass(tmp, data, size, buckets[2], 2);
711 radix_sort_pass(data, tmp, size, buckets[3], 3);
716 static int dnxhd_encode_fast(AVCodecContext *avctx, DNXHDEncContext *ctx)
720 if ((ret = dnxhd_find_qscale(ctx)) < 0)
722 for (y = 0; y < ctx->m.mb_height; y++) {
723 for (x = 0; x < ctx->m.mb_width; x++) {
724 int mb = y*ctx->m.mb_width+x;
726 ctx->mb_qscale[mb] = ctx->qscale;
727 ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale][mb].bits;
728 max_bits += ctx->mb_rc[ctx->qscale][mb].bits;
730 delta_bits = ctx->mb_rc[ctx->qscale][mb].bits-ctx->mb_rc[ctx->qscale+1][mb].bits;
731 ctx->mb_cmp[mb].mb = mb;
732 ctx->mb_cmp[mb].value = delta_bits ?
733 ((ctx->mb_rc[ctx->qscale][mb].ssd-ctx->mb_rc[ctx->qscale+1][mb].ssd)*100)/delta_bits
734 : INT_MIN; //avoid increasing qscale
737 max_bits += 31; //worst padding
741 avctx->execute2(avctx, dnxhd_mb_var_thread, NULL, NULL, ctx->m.mb_height);
742 radix_sort(ctx->mb_cmp, ctx->m.mb_num);
743 for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) {
744 int mb = ctx->mb_cmp[x].mb;
745 max_bits -= ctx->mb_rc[ctx->qscale][mb].bits - ctx->mb_rc[ctx->qscale+1][mb].bits;
746 ctx->mb_qscale[mb] = ctx->qscale+1;
747 ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale+1][mb].bits;
753 static void dnxhd_load_picture(DNXHDEncContext *ctx, const AVFrame *frame)
757 for (i = 0; i < 3; i++) {
758 ctx->frame.data[i] = frame->data[i];
759 ctx->frame.linesize[i] = frame->linesize[i];
762 for (i = 0; i < ctx->m.avctx->thread_count; i++) {
763 ctx->thread[i]->m.linesize = ctx->frame.linesize[0]<<ctx->interlaced;
764 ctx->thread[i]->m.uvlinesize = ctx->frame.linesize[1]<<ctx->interlaced;
765 ctx->thread[i]->dct_y_offset = ctx->m.linesize *8;
766 ctx->thread[i]->dct_uv_offset = ctx->m.uvlinesize*8;
769 ctx->frame.interlaced_frame = frame->interlaced_frame;
770 ctx->cur_field = frame->interlaced_frame && !frame->top_field_first;
773 static int dnxhd_encode_picture(AVCodecContext *avctx, unsigned char *buf, int buf_size, void *data)
775 DNXHDEncContext *ctx = avctx->priv_data;
779 if (buf_size < ctx->cid_table->frame_size) {
780 av_log(avctx, AV_LOG_ERROR, "output buffer is too small to compress picture\n");
784 dnxhd_load_picture(ctx, data);
787 for (i = 0; i < 3; i++) {
788 ctx->src[i] = ctx->frame.data[i];
789 if (ctx->interlaced && ctx->cur_field)
790 ctx->src[i] += ctx->frame.linesize[i];
793 dnxhd_write_header(avctx, buf);
795 if (avctx->mb_decision == FF_MB_DECISION_RD)
796 ret = dnxhd_encode_rdo(avctx, ctx);
798 ret = dnxhd_encode_fast(avctx, ctx);
800 av_log(avctx, AV_LOG_ERROR,
801 "picture could not fit ratecontrol constraints, increase qmax\n");
805 dnxhd_setup_threads_slices(ctx);
808 for (i = 0; i < ctx->m.mb_height; i++) {
809 AV_WB32(ctx->msip + i * 4, offset);
810 offset += ctx->slice_size[i];
811 assert(!(ctx->slice_size[i] & 3));
814 avctx->execute2(avctx, dnxhd_encode_thread, buf, NULL, ctx->m.mb_height);
816 assert(640 + offset + 4 <= ctx->cid_table->coding_unit_size);
817 memset(buf + 640 + offset, 0, ctx->cid_table->coding_unit_size - 4 - offset - 640);
819 AV_WB32(buf + ctx->cid_table->coding_unit_size - 4, 0x600DC0DE); // EOF
821 if (ctx->interlaced && first_field) {
824 buf += ctx->cid_table->coding_unit_size;
825 buf_size -= ctx->cid_table->coding_unit_size;
826 goto encode_coding_unit;
829 ctx->frame.quality = ctx->qscale*FF_QP2LAMBDA;
831 return ctx->cid_table->frame_size;
834 static int dnxhd_encode_end(AVCodecContext *avctx)
836 DNXHDEncContext *ctx = avctx->priv_data;
837 int max_level = 1<<(ctx->cid_table->bit_depth+2);
840 av_free(ctx->vlc_codes-max_level*2);
841 av_free(ctx->vlc_bits -max_level*2);
842 av_freep(&ctx->run_codes);
843 av_freep(&ctx->run_bits);
845 av_freep(&ctx->mb_bits);
846 av_freep(&ctx->mb_qscale);
847 av_freep(&ctx->mb_rc);
848 av_freep(&ctx->mb_cmp);
849 av_freep(&ctx->slice_size);
850 av_freep(&ctx->slice_offs);
852 av_freep(&ctx->qmatrix_c);
853 av_freep(&ctx->qmatrix_l);
854 av_freep(&ctx->qmatrix_c16);
855 av_freep(&ctx->qmatrix_l16);
857 for (i = 1; i < avctx->thread_count; i++)
858 av_freep(&ctx->thread[i]);
863 AVCodec ff_dnxhd_encoder = {
867 sizeof(DNXHDEncContext),
869 dnxhd_encode_picture,
871 .capabilities = CODEC_CAP_SLICE_THREADS,
872 .pix_fmts = (const enum PixelFormat[]){PIX_FMT_YUV422P, PIX_FMT_NONE},
873 .long_name = NULL_IF_CONFIG_SMALL("VC3/DNxHD"),
874 .priv_class = &class,