3 * Copyright (c) 2007 Baptiste Coudurier <baptiste dot coudurier at smartjog dot com>
4 * Copyright (c) 2011 MirriAd Ltd
6 * VC-3 encoder funded by the British Broadcasting Corporation
7 * 10 bit support added by MirriAd Ltd, Joseph Artsimovich <joseph@mirriad.com>
9 * This file is part of FFmpeg.
11 * FFmpeg is free software; you can redistribute it and/or
12 * modify it under the terms of the GNU Lesser General Public
13 * License as published by the Free Software Foundation; either
14 * version 2.1 of the License, or (at your option) any later version.
16 * FFmpeg is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
19 * Lesser General Public License for more details.
21 * You should have received a copy of the GNU Lesser General Public
22 * License along with FFmpeg; if not, write to the Free Software
23 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
27 #define RC_VARIANCE 1 // use variance or ssd for fast rc
29 #include "libavutil/opt.h"
33 #include "mpegvideo.h"
37 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
38 #define DNX10BIT_QMAT_SHIFT 18 // The largest value that will not lead to overflow for 10bit samples.
40 static const AVOption options[]={
41 {"nitris_compat", "encode with Avid Nitris compatibility", offsetof(DNXHDEncContext, nitris_compat), AV_OPT_TYPE_INT, {.i64 = 0}, 0, 1, VE},
45 static const AVClass class = {
46 .class_name = "dnxhd",
47 .item_name = av_default_item_name,
49 .version = LIBAVUTIL_VERSION_INT,
52 #define LAMBDA_FRAC_BITS 10
54 static void dnxhd_8bit_get_pixels_8x4_sym(DCTELEM *av_restrict block, const uint8_t *pixels, int line_size)
57 for (i = 0; i < 4; i++) {
58 block[0] = pixels[0]; block[1] = pixels[1];
59 block[2] = pixels[2]; block[3] = pixels[3];
60 block[4] = pixels[4]; block[5] = pixels[5];
61 block[6] = pixels[6]; block[7] = pixels[7];
65 memcpy(block, block - 8, sizeof(*block) * 8);
66 memcpy(block + 8, block - 16, sizeof(*block) * 8);
67 memcpy(block + 16, block - 24, sizeof(*block) * 8);
68 memcpy(block + 24, block - 32, sizeof(*block) * 8);
71 static av_always_inline void dnxhd_10bit_get_pixels_8x4_sym(DCTELEM *av_restrict block, const uint8_t *pixels, int line_size)
74 const uint16_t* pixels16 = (const uint16_t*)pixels;
77 for (i = 0; i < 4; i++) {
78 block[0] = pixels16[0]; block[1] = pixels16[1];
79 block[2] = pixels16[2]; block[3] = pixels16[3];
80 block[4] = pixels16[4]; block[5] = pixels16[5];
81 block[6] = pixels16[6]; block[7] = pixels16[7];
82 pixels16 += line_size;
85 memcpy(block, block - 8, sizeof(*block) * 8);
86 memcpy(block + 8, block - 16, sizeof(*block) * 8);
87 memcpy(block + 16, block - 24, sizeof(*block) * 8);
88 memcpy(block + 24, block - 32, sizeof(*block) * 8);
91 static int dnxhd_10bit_dct_quantize(MpegEncContext *ctx, DCTELEM *block,
92 int n, int qscale, int *overflow)
94 const uint8_t *scantable= ctx->intra_scantable.scantable;
95 const int *qmat = n<4 ? ctx->q_intra_matrix[qscale] : ctx->q_chroma_intra_matrix[qscale];
96 int last_non_zero = 0;
101 // Divide by 4 with rounding, to compensate scaling of DCT coefficients
102 block[0] = (block[0] + 2) >> 2;
104 for (i = 1; i < 64; ++i) {
105 int j = scantable[i];
106 int sign = block[j] >> 31;
107 int level = (block[j] ^ sign) - sign;
108 level = level * qmat[j] >> DNX10BIT_QMAT_SHIFT;
109 block[j] = (level ^ sign) - sign;
114 return last_non_zero;
117 static int dnxhd_init_vlc(DNXHDEncContext *ctx)
119 int i, j, level, run;
120 int max_level = 1<<(ctx->cid_table->bit_depth+2);
122 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->vlc_codes, max_level*4*sizeof(*ctx->vlc_codes), fail);
123 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->vlc_bits, max_level*4*sizeof(*ctx->vlc_bits) , fail);
124 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_codes, 63*2, fail);
125 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_bits, 63, fail);
127 ctx->vlc_codes += max_level*2;
128 ctx->vlc_bits += max_level*2;
129 for (level = -max_level; level < max_level; level++) {
130 for (run = 0; run < 2; run++) {
131 int index = (level<<1)|run;
132 int sign, offset = 0, alevel = level;
134 MASK_ABS(sign, alevel);
136 offset = (alevel-1)>>6;
139 for (j = 0; j < 257; j++) {
140 if (ctx->cid_table->ac_level[j] >> 1 == alevel &&
141 (!offset || (ctx->cid_table->ac_flags[j] & 1) && offset) &&
142 (!run || (ctx->cid_table->ac_flags[j] & 2) && run)) {
143 av_assert1(!ctx->vlc_codes[index]);
145 ctx->vlc_codes[index] = (ctx->cid_table->ac_codes[j]<<1)|(sign&1);
146 ctx->vlc_bits [index] = ctx->cid_table->ac_bits[j]+1;
148 ctx->vlc_codes[index] = ctx->cid_table->ac_codes[j];
149 ctx->vlc_bits [index] = ctx->cid_table->ac_bits [j];
154 av_assert0(!alevel || j < 257);
156 ctx->vlc_codes[index] = (ctx->vlc_codes[index]<<ctx->cid_table->index_bits)|offset;
157 ctx->vlc_bits [index]+= ctx->cid_table->index_bits;
161 for (i = 0; i < 62; i++) {
162 int run = ctx->cid_table->run[i];
163 av_assert0(run < 63);
164 ctx->run_codes[run] = ctx->cid_table->run_codes[i];
165 ctx->run_bits [run] = ctx->cid_table->run_bits[i];
172 static int dnxhd_init_qmat(DNXHDEncContext *ctx, int lbias, int cbias)
174 // init first elem to 1 to avoid div by 0 in convert_matrix
175 uint16_t weight_matrix[64] = {1,}; // convert_matrix needs uint16_t*
177 const uint8_t *luma_weight_table = ctx->cid_table->luma_weight;
178 const uint8_t *chroma_weight_table = ctx->cid_table->chroma_weight;
180 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l, (ctx->m.avctx->qmax+1) * 64 * sizeof(int), fail);
181 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c, (ctx->m.avctx->qmax+1) * 64 * sizeof(int), fail);
182 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l16, (ctx->m.avctx->qmax+1) * 64 * 2 * sizeof(uint16_t), fail);
183 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c16, (ctx->m.avctx->qmax+1) * 64 * 2 * sizeof(uint16_t), fail);
185 if (ctx->cid_table->bit_depth == 8) {
186 for (i = 1; i < 64; i++) {
187 int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]];
188 weight_matrix[j] = ctx->cid_table->luma_weight[i];
190 ff_convert_matrix(&ctx->m.dsp, ctx->qmatrix_l, ctx->qmatrix_l16, weight_matrix,
191 ctx->m.intra_quant_bias, 1, ctx->m.avctx->qmax, 1);
192 for (i = 1; i < 64; i++) {
193 int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]];
194 weight_matrix[j] = ctx->cid_table->chroma_weight[i];
196 ff_convert_matrix(&ctx->m.dsp, ctx->qmatrix_c, ctx->qmatrix_c16, weight_matrix,
197 ctx->m.intra_quant_bias, 1, ctx->m.avctx->qmax, 1);
199 for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
200 for (i = 0; i < 64; i++) {
201 ctx->qmatrix_l [qscale] [i] <<= 2; ctx->qmatrix_c [qscale] [i] <<= 2;
202 ctx->qmatrix_l16[qscale][0][i] <<= 2; ctx->qmatrix_l16[qscale][1][i] <<= 2;
203 ctx->qmatrix_c16[qscale][0][i] <<= 2; ctx->qmatrix_c16[qscale][1][i] <<= 2;
208 for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
209 for (i = 1; i < 64; i++) {
210 int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]];
212 // The quantization formula from the VC-3 standard is:
213 // quantized = sign(block[i]) * floor(abs(block[i]/s) * p / (qscale * weight_table[i]))
214 // Where p is 32 for 8-bit samples and 8 for 10-bit ones.
215 // The s factor compensates scaling of DCT coefficients done by the DCT routines,
216 // and therefore is not present in standard. It's 8 for 8-bit samples and 4 for 10-bit ones.
217 // We want values of ctx->qtmatrix_l and ctx->qtmatrix_r to be:
218 // ((1 << DNX10BIT_QMAT_SHIFT) * (p / s)) / (qscale * weight_table[i])
219 // For 10-bit samples, p / s == 2
220 ctx->qmatrix_l[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) / (qscale * luma_weight_table[i]);
221 ctx->qmatrix_c[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) / (qscale * chroma_weight_table[i]);
226 ctx->m.q_chroma_intra_matrix16 = ctx->qmatrix_c16;
227 ctx->m.q_chroma_intra_matrix = ctx->qmatrix_c;
228 ctx->m.q_intra_matrix16 = ctx->qmatrix_l16;
229 ctx->m.q_intra_matrix = ctx->qmatrix_l;
236 static int dnxhd_init_rc(DNXHDEncContext *ctx)
238 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_rc, 8160*ctx->m.avctx->qmax*sizeof(RCEntry), fail);
239 if (ctx->m.avctx->mb_decision != FF_MB_DECISION_RD)
240 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_cmp, ctx->m.mb_num*sizeof(RCCMPEntry), fail);
242 ctx->frame_bits = (ctx->cid_table->coding_unit_size - 640 - 4 - ctx->min_padding) * 8;
244 ctx->lambda = 2<<LAMBDA_FRAC_BITS; // qscale 2
250 static int dnxhd_encode_init(AVCodecContext *avctx)
252 DNXHDEncContext *ctx = avctx->priv_data;
253 int i, index, bit_depth;
255 switch (avctx->pix_fmt) {
256 case PIX_FMT_YUV422P:
259 case PIX_FMT_YUV422P10:
263 av_log(avctx, AV_LOG_ERROR, "pixel format is incompatible with DNxHD\n");
267 ctx->cid = ff_dnxhd_find_cid(avctx, bit_depth);
269 av_log(avctx, AV_LOG_ERROR, "video parameters incompatible with DNxHD\n");
272 av_log(avctx, AV_LOG_DEBUG, "cid %d\n", ctx->cid);
274 index = ff_dnxhd_get_cid_table(ctx->cid);
275 ctx->cid_table = &ff_dnxhd_cid_table[index];
277 ctx->m.avctx = avctx;
281 avctx->bits_per_raw_sample = ctx->cid_table->bit_depth;
283 ff_dsputil_init(&ctx->m.dsp, avctx);
284 ff_dct_common_init(&ctx->m);
285 if (!ctx->m.dct_quantize)
286 ctx->m.dct_quantize = ff_dct_quantize_c;
288 if (ctx->cid_table->bit_depth == 10) {
289 ctx->m.dct_quantize = dnxhd_10bit_dct_quantize;
290 ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym;
291 ctx->block_width_l2 = 4;
293 ctx->get_pixels_8x4_sym = dnxhd_8bit_get_pixels_8x4_sym;
294 ctx->block_width_l2 = 3;
298 ff_dnxhdenc_init_x86(ctx);
300 ctx->m.mb_height = (avctx->height + 15) / 16;
301 ctx->m.mb_width = (avctx->width + 15) / 16;
303 if (avctx->flags & CODEC_FLAG_INTERLACED_DCT) {
305 ctx->m.mb_height /= 2;
308 ctx->m.mb_num = ctx->m.mb_height * ctx->m.mb_width;
310 if (avctx->intra_quant_bias != FF_DEFAULT_QUANT_BIAS)
311 ctx->m.intra_quant_bias = avctx->intra_quant_bias;
312 if (dnxhd_init_qmat(ctx, ctx->m.intra_quant_bias, 0) < 0) // XXX tune lbias/cbias
315 // Avid Nitris hardware decoder requires a minimum amount of padding in the coding unit payload
316 if (ctx->nitris_compat)
317 ctx->min_padding = 1600;
319 if (dnxhd_init_vlc(ctx) < 0)
321 if (dnxhd_init_rc(ctx) < 0)
324 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_size, ctx->m.mb_height*sizeof(uint32_t), fail);
325 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_offs, ctx->m.mb_height*sizeof(uint32_t), fail);
326 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_bits, ctx->m.mb_num *sizeof(uint16_t), fail);
327 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_qscale, ctx->m.mb_num *sizeof(uint8_t), fail);
329 ctx->frame.key_frame = 1;
330 ctx->frame.pict_type = AV_PICTURE_TYPE_I;
331 ctx->m.avctx->coded_frame = &ctx->frame;
333 if (avctx->thread_count > MAX_THREADS) {
334 av_log(avctx, AV_LOG_ERROR, "too many threads\n");
338 ctx->thread[0] = ctx;
339 for (i = 1; i < avctx->thread_count; i++) {
340 ctx->thread[i] = av_malloc(sizeof(DNXHDEncContext));
341 memcpy(ctx->thread[i], ctx, sizeof(DNXHDEncContext));
345 fail: //for FF_ALLOCZ_OR_GOTO
349 static int dnxhd_write_header(AVCodecContext *avctx, uint8_t *buf)
351 DNXHDEncContext *ctx = avctx->priv_data;
352 const uint8_t header_prefix[5] = { 0x00,0x00,0x02,0x80,0x01 };
356 memcpy(buf, header_prefix, 5);
357 buf[5] = ctx->interlaced ? ctx->cur_field+2 : 0x01;
358 buf[6] = 0x80; // crc flag off
359 buf[7] = 0xa0; // reserved
360 AV_WB16(buf + 0x18, avctx->height>>ctx->interlaced); // ALPF
361 AV_WB16(buf + 0x1a, avctx->width); // SPL
362 AV_WB16(buf + 0x1d, avctx->height>>ctx->interlaced); // NAL
364 buf[0x21] = ctx->cid_table->bit_depth == 10 ? 0x58 : 0x38;
365 buf[0x22] = 0x88 + (ctx->interlaced<<2);
366 AV_WB32(buf + 0x28, ctx->cid); // CID
367 buf[0x2c] = ctx->interlaced ? 0 : 0x80;
369 buf[0x5f] = 0x01; // UDL
371 buf[0x167] = 0x02; // reserved
372 AV_WB16(buf + 0x16a, ctx->m.mb_height * 4 + 4); // MSIPS
373 buf[0x16d] = ctx->m.mb_height; // Ns
374 buf[0x16f] = 0x10; // reserved
376 ctx->msip = buf + 0x170;
380 static av_always_inline void dnxhd_encode_dc(DNXHDEncContext *ctx, int diff)
384 nbits = av_log2_16bit(-2*diff);
387 nbits = av_log2_16bit(2*diff);
389 put_bits(&ctx->m.pb, ctx->cid_table->dc_bits[nbits] + nbits,
390 (ctx->cid_table->dc_codes[nbits]<<nbits) + (diff & ((1 << nbits) - 1)));
393 static av_always_inline void dnxhd_encode_block(DNXHDEncContext *ctx, DCTELEM *block, int last_index, int n)
395 int last_non_zero = 0;
398 dnxhd_encode_dc(ctx, block[0] - ctx->m.last_dc[n]);
399 ctx->m.last_dc[n] = block[0];
401 for (i = 1; i <= last_index; i++) {
402 j = ctx->m.intra_scantable.permutated[i];
405 int run_level = i - last_non_zero - 1;
406 int rlevel = (slevel<<1)|!!run_level;
407 put_bits(&ctx->m.pb, ctx->vlc_bits[rlevel], ctx->vlc_codes[rlevel]);
409 put_bits(&ctx->m.pb, ctx->run_bits[run_level], ctx->run_codes[run_level]);
413 put_bits(&ctx->m.pb, ctx->vlc_bits[0], ctx->vlc_codes[0]); // EOB
416 static av_always_inline void dnxhd_unquantize_c(DNXHDEncContext *ctx, DCTELEM *block, int n, int qscale, int last_index)
418 const uint8_t *weight_matrix;
422 weight_matrix = (n&2) ? ctx->cid_table->chroma_weight : ctx->cid_table->luma_weight;
424 for (i = 1; i <= last_index; i++) {
425 int j = ctx->m.intra_scantable.permutated[i];
429 level = (1-2*level) * qscale * weight_matrix[i];
430 if (ctx->cid_table->bit_depth == 10) {
431 if (weight_matrix[i] != 8)
435 if (weight_matrix[i] != 32)
441 level = (2*level+1) * qscale * weight_matrix[i];
442 if (ctx->cid_table->bit_depth == 10) {
443 if (weight_matrix[i] != 8)
447 if (weight_matrix[i] != 32)
457 static av_always_inline int dnxhd_ssd_block(DCTELEM *qblock, DCTELEM *block)
461 for (i = 0; i < 64; i++)
462 score += (block[i] - qblock[i]) * (block[i] - qblock[i]);
466 static av_always_inline int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, DCTELEM *block, int last_index)
468 int last_non_zero = 0;
471 for (i = 1; i <= last_index; i++) {
472 j = ctx->m.intra_scantable.permutated[i];
475 int run_level = i - last_non_zero - 1;
476 bits += ctx->vlc_bits[(level<<1)|!!run_level]+ctx->run_bits[run_level];
483 static av_always_inline void dnxhd_get_blocks(DNXHDEncContext *ctx, int mb_x, int mb_y)
485 const int bs = ctx->block_width_l2;
486 const int bw = 1 << bs;
487 const uint8_t *ptr_y = ctx->thread[0]->src[0] + ((mb_y << 4) * ctx->m.linesize) + (mb_x << bs+1);
488 const uint8_t *ptr_u = ctx->thread[0]->src[1] + ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs);
489 const uint8_t *ptr_v = ctx->thread[0]->src[2] + ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs);
490 DSPContext *dsp = &ctx->m.dsp;
492 dsp->get_pixels(ctx->blocks[0], ptr_y, ctx->m.linesize);
493 dsp->get_pixels(ctx->blocks[1], ptr_y + bw, ctx->m.linesize);
494 dsp->get_pixels(ctx->blocks[2], ptr_u, ctx->m.uvlinesize);
495 dsp->get_pixels(ctx->blocks[3], ptr_v, ctx->m.uvlinesize);
497 if (mb_y+1 == ctx->m.mb_height && ctx->m.avctx->height == 1080) {
498 if (ctx->interlaced) {
499 ctx->get_pixels_8x4_sym(ctx->blocks[4], ptr_y + ctx->dct_y_offset, ctx->m.linesize);
500 ctx->get_pixels_8x4_sym(ctx->blocks[5], ptr_y + ctx->dct_y_offset + bw, ctx->m.linesize);
501 ctx->get_pixels_8x4_sym(ctx->blocks[6], ptr_u + ctx->dct_uv_offset, ctx->m.uvlinesize);
502 ctx->get_pixels_8x4_sym(ctx->blocks[7], ptr_v + ctx->dct_uv_offset, ctx->m.uvlinesize);
504 dsp->clear_block(ctx->blocks[4]);
505 dsp->clear_block(ctx->blocks[5]);
506 dsp->clear_block(ctx->blocks[6]);
507 dsp->clear_block(ctx->blocks[7]);
510 dsp->get_pixels(ctx->blocks[4], ptr_y + ctx->dct_y_offset, ctx->m.linesize);
511 dsp->get_pixels(ctx->blocks[5], ptr_y + ctx->dct_y_offset + bw, ctx->m.linesize);
512 dsp->get_pixels(ctx->blocks[6], ptr_u + ctx->dct_uv_offset, ctx->m.uvlinesize);
513 dsp->get_pixels(ctx->blocks[7], ptr_v + ctx->dct_uv_offset, ctx->m.uvlinesize);
517 static av_always_inline int dnxhd_switch_matrix(DNXHDEncContext *ctx, int i)
519 const static uint8_t component[8]={0,0,1,2,0,0,1,2};
523 static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg, int jobnr, int threadnr)
525 DNXHDEncContext *ctx = avctx->priv_data;
526 int mb_y = jobnr, mb_x;
527 int qscale = ctx->qscale;
528 LOCAL_ALIGNED_16(DCTELEM, block, [64]);
529 ctx = ctx->thread[threadnr];
533 ctx->m.last_dc[2] = 1 << (ctx->cid_table->bit_depth + 2);
535 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
536 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
542 dnxhd_get_blocks(ctx, mb_x, mb_y);
544 for (i = 0; i < 8; i++) {
545 DCTELEM *src_block = ctx->blocks[i];
546 int overflow, nbits, diff, last_index;
547 int n = dnxhd_switch_matrix(ctx, i);
549 memcpy(block, src_block, 64*sizeof(*block));
550 last_index = ctx->m.dct_quantize(&ctx->m, block, 4&(2*i), qscale, &overflow);
551 ac_bits += dnxhd_calc_ac_bits(ctx, block, last_index);
553 diff = block[0] - ctx->m.last_dc[n];
554 if (diff < 0) nbits = av_log2_16bit(-2*diff);
555 else nbits = av_log2_16bit( 2*diff);
557 av_assert1(nbits < ctx->cid_table->bit_depth + 4);
558 dc_bits += ctx->cid_table->dc_bits[nbits] + nbits;
560 ctx->m.last_dc[n] = block[0];
562 if (avctx->mb_decision == FF_MB_DECISION_RD || !RC_VARIANCE) {
563 dnxhd_unquantize_c(ctx, block, i, qscale, last_index);
564 ctx->m.dsp.idct(block);
565 ssd += dnxhd_ssd_block(block, src_block);
568 ctx->mb_rc[qscale][mb].ssd = ssd;
569 ctx->mb_rc[qscale][mb].bits = ac_bits+dc_bits+12+8*ctx->vlc_bits[0];
574 static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg, int jobnr, int threadnr)
576 DNXHDEncContext *ctx = avctx->priv_data;
577 int mb_y = jobnr, mb_x;
578 ctx = ctx->thread[threadnr];
579 init_put_bits(&ctx->m.pb, (uint8_t *)arg + 640 + ctx->slice_offs[jobnr], ctx->slice_size[jobnr]);
583 ctx->m.last_dc[2] = 1 << (ctx->cid_table->bit_depth + 2);
584 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
585 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
586 int qscale = ctx->mb_qscale[mb];
589 put_bits(&ctx->m.pb, 12, qscale<<1);
591 dnxhd_get_blocks(ctx, mb_x, mb_y);
593 for (i = 0; i < 8; i++) {
594 DCTELEM *block = ctx->blocks[i];
595 int overflow, n = dnxhd_switch_matrix(ctx, i);
596 int last_index = ctx->m.dct_quantize(&ctx->m, block, 4&(2*i), qscale, &overflow);
598 dnxhd_encode_block(ctx, block, last_index, n);
599 //STOP_TIMER("encode_block");
602 if (put_bits_count(&ctx->m.pb)&31)
603 put_bits(&ctx->m.pb, 32-(put_bits_count(&ctx->m.pb)&31), 0);
604 flush_put_bits(&ctx->m.pb);
608 static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx)
612 for (mb_y = 0; mb_y < ctx->m.mb_height; mb_y++) {
614 ctx->slice_offs[mb_y] = offset;
615 ctx->slice_size[mb_y] = 0;
616 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
617 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
618 ctx->slice_size[mb_y] += ctx->mb_bits[mb];
620 ctx->slice_size[mb_y] = (ctx->slice_size[mb_y]+31)&~31;
621 ctx->slice_size[mb_y] >>= 3;
622 thread_size = ctx->slice_size[mb_y];
623 offset += thread_size;
627 static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg, int jobnr, int threadnr)
629 DNXHDEncContext *ctx = avctx->priv_data;
630 int mb_y = jobnr, mb_x;
631 ctx = ctx->thread[threadnr];
632 if (ctx->cid_table->bit_depth == 8) {
633 uint8_t *pix = ctx->thread[0]->src[0] + ((mb_y<<4) * ctx->m.linesize);
634 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x, pix += 16) {
635 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
636 int sum = ctx->m.dsp.pix_sum(pix, ctx->m.linesize);
637 int varc = (ctx->m.dsp.pix_norm1(pix, ctx->m.linesize) - (((unsigned)sum*sum)>>8)+128)>>8;
638 ctx->mb_cmp[mb].value = varc;
639 ctx->mb_cmp[mb].mb = mb;
642 int const linesize = ctx->m.linesize >> 1;
643 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x) {
644 uint16_t *pix = (uint16_t*)ctx->thread[0]->src[0] + ((mb_y << 4) * linesize) + (mb_x << 4);
645 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
650 // Macroblocks are 16x16 pixels, unlike DCT blocks which are 8x8.
651 for (i = 0; i < 16; ++i) {
652 for (j = 0; j < 16; ++j) {
653 // Turn 16-bit pixels into 10-bit ones.
654 int const sample = (unsigned)pix[j] >> 6;
656 sqsum += sample * sample;
657 // 2^10 * 2^10 * 16 * 16 = 2^28, which is less than INT_MAX
661 mean = sum >> 8; // 16*16 == 2^8
663 ctx->mb_cmp[mb].value = sqmean - mean * mean;
664 ctx->mb_cmp[mb].mb = mb;
670 static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx)
672 int lambda, up_step, down_step;
673 int last_lower = INT_MAX, last_higher = 0;
676 for (q = 1; q < avctx->qmax; q++) {
678 avctx->execute2(avctx, dnxhd_calc_bits_thread, NULL, NULL, ctx->m.mb_height);
680 up_step = down_step = 2<<LAMBDA_FRAC_BITS;
681 lambda = ctx->lambda;
686 if (lambda == last_higher) {
688 end = 1; // need to set final qscales/bits
690 for (y = 0; y < ctx->m.mb_height; y++) {
691 for (x = 0; x < ctx->m.mb_width; x++) {
692 unsigned min = UINT_MAX;
694 int mb = y*ctx->m.mb_width+x;
695 for (q = 1; q < avctx->qmax; q++) {
696 unsigned score = ctx->mb_rc[q][mb].bits*lambda+
697 ((unsigned)ctx->mb_rc[q][mb].ssd<<LAMBDA_FRAC_BITS);
703 bits += ctx->mb_rc[qscale][mb].bits;
704 ctx->mb_qscale[mb] = qscale;
705 ctx->mb_bits[mb] = ctx->mb_rc[qscale][mb].bits;
707 bits = (bits+31)&~31; // padding
708 if (bits > ctx->frame_bits)
711 //av_dlog(ctx->m.avctx, "lambda %d, up %u, down %u, bits %d, frame %d\n",
712 // lambda, last_higher, last_lower, bits, ctx->frame_bits);
714 if (bits > ctx->frame_bits)
718 if (bits < ctx->frame_bits) {
719 last_lower = FFMIN(lambda, last_lower);
720 if (last_higher != 0)
721 lambda = (lambda+last_higher)>>1;
724 down_step = FFMIN((int64_t)down_step*5, INT_MAX);
725 up_step = 1<<LAMBDA_FRAC_BITS;
726 lambda = FFMAX(1, lambda);
727 if (lambda == last_lower)
730 last_higher = FFMAX(lambda, last_higher);
731 if (last_lower != INT_MAX)
732 lambda = (lambda+last_lower)>>1;
733 else if ((int64_t)lambda + up_step > INT_MAX)
737 up_step = FFMIN((int64_t)up_step*5, INT_MAX);
738 down_step = 1<<LAMBDA_FRAC_BITS;
741 //av_dlog(ctx->m.avctx, "out lambda %d\n", lambda);
742 ctx->lambda = lambda;
746 static int dnxhd_find_qscale(DNXHDEncContext *ctx)
752 int last_lower = INT_MAX;
756 qscale = ctx->qscale;
759 ctx->qscale = qscale;
760 // XXX avoid recalculating bits
761 ctx->m.avctx->execute2(ctx->m.avctx, dnxhd_calc_bits_thread, NULL, NULL, ctx->m.mb_height);
762 for (y = 0; y < ctx->m.mb_height; y++) {
763 for (x = 0; x < ctx->m.mb_width; x++)
764 bits += ctx->mb_rc[qscale][y*ctx->m.mb_width+x].bits;
765 bits = (bits+31)&~31; // padding
766 if (bits > ctx->frame_bits)
769 //av_dlog(ctx->m.avctx, "%d, qscale %d, bits %d, frame %d, higher %d, lower %d\n",
770 // ctx->m.avctx->frame_number, qscale, bits, ctx->frame_bits, last_higher, last_lower);
771 if (bits < ctx->frame_bits) {
774 if (last_higher == qscale - 1) {
775 qscale = last_higher;
778 last_lower = FFMIN(qscale, last_lower);
779 if (last_higher != 0)
780 qscale = (qscale+last_higher)>>1;
782 qscale -= down_step++;
787 if (last_lower == qscale + 1)
789 last_higher = FFMAX(qscale, last_higher);
790 if (last_lower != INT_MAX)
791 qscale = (qscale+last_lower)>>1;
795 if (qscale >= ctx->m.avctx->qmax)
799 //av_dlog(ctx->m.avctx, "out qscale %d\n", qscale);
800 ctx->qscale = qscale;
804 #define BUCKET_BITS 8
805 #define RADIX_PASSES 4
806 #define NBUCKETS (1 << BUCKET_BITS)
808 static inline int get_bucket(int value, int shift)
811 value &= NBUCKETS - 1;
812 return NBUCKETS - 1 - value;
815 static void radix_count(const RCCMPEntry *data, int size, int buckets[RADIX_PASSES][NBUCKETS])
818 memset(buckets, 0, sizeof(buckets[0][0]) * RADIX_PASSES * NBUCKETS);
819 for (i = 0; i < size; i++) {
820 int v = data[i].value;
821 for (j = 0; j < RADIX_PASSES; j++) {
822 buckets[j][get_bucket(v, 0)]++;
827 for (j = 0; j < RADIX_PASSES; j++) {
829 for (i = NBUCKETS - 1; i >= 0; i--)
830 buckets[j][i] = offset -= buckets[j][i];
831 av_assert1(!buckets[j][0]);
835 static void radix_sort_pass(RCCMPEntry *dst, const RCCMPEntry *data, int size, int buckets[NBUCKETS], int pass)
837 int shift = pass * BUCKET_BITS;
839 for (i = 0; i < size; i++) {
840 int v = get_bucket(data[i].value, shift);
841 int pos = buckets[v]++;
846 static void radix_sort(RCCMPEntry *data, int size)
848 int buckets[RADIX_PASSES][NBUCKETS];
849 RCCMPEntry *tmp = av_malloc(sizeof(*tmp) * size);
850 radix_count(data, size, buckets);
851 radix_sort_pass(tmp, data, size, buckets[0], 0);
852 radix_sort_pass(data, tmp, size, buckets[1], 1);
853 if (buckets[2][NBUCKETS - 1] || buckets[3][NBUCKETS - 1]) {
854 radix_sort_pass(tmp, data, size, buckets[2], 2);
855 radix_sort_pass(data, tmp, size, buckets[3], 3);
860 static int dnxhd_encode_fast(AVCodecContext *avctx, DNXHDEncContext *ctx)
864 if ((ret = dnxhd_find_qscale(ctx)) < 0)
866 for (y = 0; y < ctx->m.mb_height; y++) {
867 for (x = 0; x < ctx->m.mb_width; x++) {
868 int mb = y*ctx->m.mb_width+x;
870 ctx->mb_qscale[mb] = ctx->qscale;
871 ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale][mb].bits;
872 max_bits += ctx->mb_rc[ctx->qscale][mb].bits;
874 delta_bits = ctx->mb_rc[ctx->qscale][mb].bits-ctx->mb_rc[ctx->qscale+1][mb].bits;
875 ctx->mb_cmp[mb].mb = mb;
876 ctx->mb_cmp[mb].value = delta_bits ?
877 ((ctx->mb_rc[ctx->qscale][mb].ssd-ctx->mb_rc[ctx->qscale+1][mb].ssd)*100)/delta_bits
878 : INT_MIN; //avoid increasing qscale
881 max_bits += 31; //worst padding
885 avctx->execute2(avctx, dnxhd_mb_var_thread, NULL, NULL, ctx->m.mb_height);
886 radix_sort(ctx->mb_cmp, ctx->m.mb_num);
887 for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) {
888 int mb = ctx->mb_cmp[x].mb;
889 max_bits -= ctx->mb_rc[ctx->qscale][mb].bits - ctx->mb_rc[ctx->qscale+1][mb].bits;
890 ctx->mb_qscale[mb] = ctx->qscale+1;
891 ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale+1][mb].bits;
897 static void dnxhd_load_picture(DNXHDEncContext *ctx, const AVFrame *frame)
901 for (i = 0; i < 3; i++) {
902 ctx->frame.data[i] = frame->data[i];
903 ctx->frame.linesize[i] = frame->linesize[i];
906 for (i = 0; i < ctx->m.avctx->thread_count; i++) {
907 ctx->thread[i]->m.linesize = ctx->frame.linesize[0]<<ctx->interlaced;
908 ctx->thread[i]->m.uvlinesize = ctx->frame.linesize[1]<<ctx->interlaced;
909 ctx->thread[i]->dct_y_offset = ctx->m.linesize *8;
910 ctx->thread[i]->dct_uv_offset = ctx->m.uvlinesize*8;
913 ctx->frame.interlaced_frame = frame->interlaced_frame;
914 ctx->cur_field = frame->interlaced_frame && !frame->top_field_first;
917 static int dnxhd_encode_picture(AVCodecContext *avctx, AVPacket *pkt,
918 const AVFrame *frame, int *got_packet)
920 DNXHDEncContext *ctx = avctx->priv_data;
925 if ((ret = ff_alloc_packet2(avctx, pkt, ctx->cid_table->frame_size)) < 0)
929 dnxhd_load_picture(ctx, frame);
932 for (i = 0; i < 3; i++) {
933 ctx->src[i] = ctx->frame.data[i];
934 if (ctx->interlaced && ctx->cur_field)
935 ctx->src[i] += ctx->frame.linesize[i];
938 dnxhd_write_header(avctx, buf);
940 if (avctx->mb_decision == FF_MB_DECISION_RD)
941 ret = dnxhd_encode_rdo(avctx, ctx);
943 ret = dnxhd_encode_fast(avctx, ctx);
945 av_log(avctx, AV_LOG_ERROR,
946 "picture could not fit ratecontrol constraints, increase qmax\n");
950 dnxhd_setup_threads_slices(ctx);
953 for (i = 0; i < ctx->m.mb_height; i++) {
954 AV_WB32(ctx->msip + i * 4, offset);
955 offset += ctx->slice_size[i];
956 av_assert1(!(ctx->slice_size[i] & 3));
959 avctx->execute2(avctx, dnxhd_encode_thread, buf, NULL, ctx->m.mb_height);
961 av_assert1(640 + offset + 4 <= ctx->cid_table->coding_unit_size);
962 memset(buf + 640 + offset, 0, ctx->cid_table->coding_unit_size - 4 - offset - 640);
964 AV_WB32(buf + ctx->cid_table->coding_unit_size - 4, 0x600DC0DE); // EOF
966 if (ctx->interlaced && first_field) {
969 buf += ctx->cid_table->coding_unit_size;
970 goto encode_coding_unit;
973 ctx->frame.quality = ctx->qscale*FF_QP2LAMBDA;
975 pkt->flags |= AV_PKT_FLAG_KEY;
980 static int dnxhd_encode_end(AVCodecContext *avctx)
982 DNXHDEncContext *ctx = avctx->priv_data;
983 int max_level = 1<<(ctx->cid_table->bit_depth+2);
986 av_free(ctx->vlc_codes-max_level*2);
987 av_free(ctx->vlc_bits -max_level*2);
988 av_freep(&ctx->run_codes);
989 av_freep(&ctx->run_bits);
991 av_freep(&ctx->mb_bits);
992 av_freep(&ctx->mb_qscale);
993 av_freep(&ctx->mb_rc);
994 av_freep(&ctx->mb_cmp);
995 av_freep(&ctx->slice_size);
996 av_freep(&ctx->slice_offs);
998 av_freep(&ctx->qmatrix_c);
999 av_freep(&ctx->qmatrix_l);
1000 av_freep(&ctx->qmatrix_c16);
1001 av_freep(&ctx->qmatrix_l16);
1003 for (i = 1; i < avctx->thread_count; i++)
1004 av_freep(&ctx->thread[i]);
1009 static const AVCodecDefault dnxhd_defaults[] = {
1010 { "qmax", "1024" }, /* Maximum quantization scale factor allowed for VC-3 */
1014 AVCodec ff_dnxhd_encoder = {
1016 .type = AVMEDIA_TYPE_VIDEO,
1017 .id = AV_CODEC_ID_DNXHD,
1018 .priv_data_size = sizeof(DNXHDEncContext),
1019 .init = dnxhd_encode_init,
1020 .encode2 = dnxhd_encode_picture,
1021 .close = dnxhd_encode_end,
1022 .capabilities = CODEC_CAP_SLICE_THREADS,
1023 .pix_fmts = (const enum PixelFormat[]){ PIX_FMT_YUV422P,
1026 .long_name = NULL_IF_CONFIG_SMALL("VC3/DNxHD"),
1027 .priv_class = &class,
1028 .defaults = dnxhd_defaults,