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 Libav.
11 * Libav 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 * Libav 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 Libav; if not, write to the Free Software
23 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
26 #include "libavutil/attributes.h"
27 #include "libavutil/internal.h"
28 #include "libavutil/opt.h"
29 #include "libavutil/timer.h"
35 #include "mpegvideo.h"
38 // The largest value that will not lead to overflow for 10bit samples.
39 #define DNX10BIT_QMAT_SHIFT 18
40 #define RC_VARIANCE 1 // use variance or ssd for fast rc
41 #define LAMBDA_FRAC_BITS 10
43 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
44 static const AVOption options[] = {
45 { "nitris_compat", "encode with Avid Nitris compatibility",
46 offsetof(DNXHDEncContext, nitris_compat), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 1, VE },
50 static const AVClass class = {
57 static void dnxhd_8bit_get_pixels_8x4_sym(int16_t *restrict block,
58 const uint8_t *pixels,
62 for (i = 0; i < 4; i++) {
74 memcpy(block, block - 8, sizeof(*block) * 8);
75 memcpy(block + 8, block - 16, sizeof(*block) * 8);
76 memcpy(block + 16, block - 24, sizeof(*block) * 8);
77 memcpy(block + 24, block - 32, sizeof(*block) * 8);
80 static av_always_inline
81 void dnxhd_10bit_get_pixels_8x4_sym(int16_t *restrict block,
82 const uint8_t *pixels,
89 for (i = 0; i < 4; i++) {
90 memcpy(block + i * 8, pixels + i * line_size, 8 * sizeof(*block));
91 memcpy(block - (i + 1) * 8, pixels + i * line_size, 8 * sizeof(*block));
95 static int dnxhd_10bit_dct_quantize(MpegEncContext *ctx, int16_t *block,
96 int n, int qscale, int *overflow)
98 const uint8_t *scantable= ctx->intra_scantable.scantable;
99 const int *qmat = ctx->q_intra_matrix[qscale];
100 int last_non_zero = 0;
103 ctx->dsp.fdct(block);
105 // Divide by 4 with rounding, to compensate scaling of DCT coefficients
106 block[0] = (block[0] + 2) >> 2;
108 for (i = 1; i < 64; ++i) {
109 int j = scantable[i];
110 int sign = block[j] >> 31;
111 int level = (block[j] ^ sign) - sign;
112 level = level * qmat[j] >> DNX10BIT_QMAT_SHIFT;
113 block[j] = (level ^ sign) - sign;
118 return last_non_zero;
121 static av_cold int dnxhd_init_vlc(DNXHDEncContext *ctx)
123 int i, j, level, run;
124 int max_level = 1 << (ctx->cid_table->bit_depth + 2);
126 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->vlc_codes,
127 max_level * 4 * sizeof(*ctx->vlc_codes), fail);
128 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->vlc_bits,
129 max_level * 4 * sizeof(*ctx->vlc_bits), fail);
130 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_codes,
132 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_bits,
135 ctx->vlc_codes += max_level * 2;
136 ctx->vlc_bits += max_level * 2;
137 for (level = -max_level; level < max_level; level++) {
138 for (run = 0; run < 2; run++) {
139 int index = (level << 1) | run;
140 int sign, offset = 0, alevel = level;
142 MASK_ABS(sign, alevel);
144 offset = (alevel - 1) >> 6;
145 alevel -= offset << 6;
147 for (j = 0; j < 257; j++) {
148 if (ctx->cid_table->ac_level[j] == alevel &&
149 (!offset || (ctx->cid_table->ac_index_flag[j] && offset)) &&
150 (!run || (ctx->cid_table->ac_run_flag [j] && run))) {
151 assert(!ctx->vlc_codes[index]);
153 ctx->vlc_codes[index] =
154 (ctx->cid_table->ac_codes[j] << 1) | (sign & 1);
155 ctx->vlc_bits[index] = ctx->cid_table->ac_bits[j] + 1;
157 ctx->vlc_codes[index] = ctx->cid_table->ac_codes[j];
158 ctx->vlc_bits[index] = ctx->cid_table->ac_bits[j];
163 assert(!alevel || j < 257);
165 ctx->vlc_codes[index] =
166 (ctx->vlc_codes[index] << ctx->cid_table->index_bits) | offset;
167 ctx->vlc_bits[index] += ctx->cid_table->index_bits;
171 for (i = 0; i < 62; i++) {
172 int run = ctx->cid_table->run[i];
174 ctx->run_codes[run] = ctx->cid_table->run_codes[i];
175 ctx->run_bits[run] = ctx->cid_table->run_bits[i];
179 return AVERROR(ENOMEM);
182 static av_cold int dnxhd_init_qmat(DNXHDEncContext *ctx, int lbias, int cbias)
184 // init first elem to 1 to avoid div by 0 in convert_matrix
185 uint16_t weight_matrix[64] = { 1, }; // convert_matrix needs uint16_t*
187 const uint8_t *luma_weight_table = ctx->cid_table->luma_weight;
188 const uint8_t *chroma_weight_table = ctx->cid_table->chroma_weight;
190 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l,
191 (ctx->m.avctx->qmax + 1) * 64 * sizeof(int), fail);
192 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c,
193 (ctx->m.avctx->qmax + 1) * 64 * sizeof(int), fail);
194 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l16,
195 (ctx->m.avctx->qmax + 1) * 64 * 2 * sizeof(uint16_t),
197 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c16,
198 (ctx->m.avctx->qmax + 1) * 64 * 2 * sizeof(uint16_t),
201 if (ctx->cid_table->bit_depth == 8) {
202 for (i = 1; i < 64; i++) {
203 int j = ctx->m.idsp.idct_permutation[ff_zigzag_direct[i]];
204 weight_matrix[j] = ctx->cid_table->luma_weight[i];
206 ff_convert_matrix(&ctx->m, ctx->qmatrix_l, ctx->qmatrix_l16,
207 weight_matrix, ctx->m.intra_quant_bias, 1,
208 ctx->m.avctx->qmax, 1);
209 for (i = 1; i < 64; i++) {
210 int j = ctx->m.idsp.idct_permutation[ff_zigzag_direct[i]];
211 weight_matrix[j] = ctx->cid_table->chroma_weight[i];
213 ff_convert_matrix(&ctx->m, ctx->qmatrix_c, ctx->qmatrix_c16,
214 weight_matrix, ctx->m.intra_quant_bias, 1,
215 ctx->m.avctx->qmax, 1);
217 for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
218 for (i = 0; i < 64; i++) {
219 ctx->qmatrix_l[qscale][i] <<= 2;
220 ctx->qmatrix_c[qscale][i] <<= 2;
221 ctx->qmatrix_l16[qscale][0][i] <<= 2;
222 ctx->qmatrix_l16[qscale][1][i] <<= 2;
223 ctx->qmatrix_c16[qscale][0][i] <<= 2;
224 ctx->qmatrix_c16[qscale][1][i] <<= 2;
229 for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
230 for (i = 1; i < 64; i++) {
231 int j = ctx->m.idsp.idct_permutation[ff_zigzag_direct[i]];
233 /* The quantization formula from the VC-3 standard is:
234 * quantized = sign(block[i]) * floor(abs(block[i]/s) * p /
235 * (qscale * weight_table[i]))
236 * Where p is 32 for 8-bit samples and 8 for 10-bit ones.
237 * The s factor compensates scaling of DCT coefficients done by
238 * the DCT routines, and therefore is not present in standard.
239 * It's 8 for 8-bit samples and 4 for 10-bit ones.
240 * We want values of ctx->qtmatrix_l and ctx->qtmatrix_r to be:
241 * ((1 << DNX10BIT_QMAT_SHIFT) * (p / s)) /
242 * (qscale * weight_table[i])
243 * For 10-bit samples, p / s == 2 */
244 ctx->qmatrix_l[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) /
245 (qscale * luma_weight_table[i]);
246 ctx->qmatrix_c[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) /
247 (qscale * chroma_weight_table[i]);
254 return AVERROR(ENOMEM);
257 static av_cold int dnxhd_init_rc(DNXHDEncContext *ctx)
259 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_rc,
260 8160 * ctx->m.avctx->qmax * sizeof(RCEntry), fail);
261 if (ctx->m.avctx->mb_decision != FF_MB_DECISION_RD)
262 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_cmp,
263 ctx->m.mb_num * sizeof(RCCMPEntry), fail);
265 ctx->frame_bits = (ctx->cid_table->coding_unit_size -
266 640 - 4 - ctx->min_padding) * 8;
268 ctx->lambda = 2 << LAMBDA_FRAC_BITS; // qscale 2
271 return AVERROR(ENOMEM);
274 static av_cold int dnxhd_encode_init(AVCodecContext *avctx)
276 DNXHDEncContext *ctx = avctx->priv_data;
277 int i, index, bit_depth, ret;
279 switch (avctx->pix_fmt) {
280 case AV_PIX_FMT_YUV422P:
283 case AV_PIX_FMT_YUV422P10:
287 av_log(avctx, AV_LOG_ERROR,
288 "pixel format is incompatible with DNxHD\n");
289 return AVERROR(EINVAL);
292 ctx->cid = ff_dnxhd_find_cid(avctx, bit_depth);
294 av_log(avctx, AV_LOG_ERROR,
295 "video parameters incompatible with DNxHD\n");
296 return AVERROR(EINVAL);
298 av_log(avctx, AV_LOG_DEBUG, "cid %d\n", ctx->cid);
300 index = ff_dnxhd_get_cid_table(ctx->cid);
301 ctx->cid_table = &ff_dnxhd_cid_table[index];
303 ctx->m.avctx = avctx;
307 avctx->bits_per_raw_sample = ctx->cid_table->bit_depth;
309 ff_blockdsp_init(&ctx->bdsp, avctx);
310 ff_dsputil_init(&ctx->m.dsp, avctx);
311 ff_idctdsp_init(&ctx->m.idsp, avctx);
312 ff_dct_common_init(&ctx->m);
313 if (!ctx->m.dct_quantize)
314 ctx->m.dct_quantize = ff_dct_quantize_c;
316 if (ctx->cid_table->bit_depth == 10) {
317 ctx->m.dct_quantize = dnxhd_10bit_dct_quantize;
318 ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym;
319 ctx->block_width_l2 = 4;
321 ctx->get_pixels_8x4_sym = dnxhd_8bit_get_pixels_8x4_sym;
322 ctx->block_width_l2 = 3;
326 ff_dnxhdenc_init_x86(ctx);
328 ctx->m.mb_height = (avctx->height + 15) / 16;
329 ctx->m.mb_width = (avctx->width + 15) / 16;
331 if (avctx->flags & CODEC_FLAG_INTERLACED_DCT) {
333 ctx->m.mb_height /= 2;
336 ctx->m.mb_num = ctx->m.mb_height * ctx->m.mb_width;
338 if (avctx->intra_quant_bias != FF_DEFAULT_QUANT_BIAS)
339 ctx->m.intra_quant_bias = avctx->intra_quant_bias;
340 // XXX tune lbias/cbias
341 if ((ret = dnxhd_init_qmat(ctx, ctx->m.intra_quant_bias, 0)) < 0)
344 /* Avid Nitris hardware decoder requires a minimum amount of padding
345 * in the coding unit payload */
346 if (ctx->nitris_compat)
347 ctx->min_padding = 1600;
349 if ((ret = dnxhd_init_vlc(ctx)) < 0)
351 if ((ret = dnxhd_init_rc(ctx)) < 0)
354 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_size,
355 ctx->m.mb_height * sizeof(uint32_t), fail);
356 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_offs,
357 ctx->m.mb_height * sizeof(uint32_t), fail);
358 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_bits,
359 ctx->m.mb_num * sizeof(uint16_t), fail);
360 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_qscale,
361 ctx->m.mb_num * sizeof(uint8_t), fail);
363 avctx->coded_frame = av_frame_alloc();
364 if (!avctx->coded_frame)
365 return AVERROR(ENOMEM);
367 avctx->coded_frame->key_frame = 1;
368 avctx->coded_frame->pict_type = AV_PICTURE_TYPE_I;
370 if (avctx->thread_count > MAX_THREADS) {
371 av_log(avctx, AV_LOG_ERROR, "too many threads\n");
372 return AVERROR(EINVAL);
375 ctx->thread[0] = ctx;
376 for (i = 1; i < avctx->thread_count; i++) {
377 ctx->thread[i] = av_malloc(sizeof(DNXHDEncContext));
378 memcpy(ctx->thread[i], ctx, sizeof(DNXHDEncContext));
382 fail: // for FF_ALLOCZ_OR_GOTO
383 return AVERROR(ENOMEM);
386 static int dnxhd_write_header(AVCodecContext *avctx, uint8_t *buf)
388 DNXHDEncContext *ctx = avctx->priv_data;
389 const uint8_t header_prefix[5] = { 0x00, 0x00, 0x02, 0x80, 0x01 };
393 memcpy(buf, header_prefix, 5);
394 buf[5] = ctx->interlaced ? ctx->cur_field + 2 : 0x01;
395 buf[6] = 0x80; // crc flag off
396 buf[7] = 0xa0; // reserved
397 AV_WB16(buf + 0x18, avctx->height >> ctx->interlaced); // ALPF
398 AV_WB16(buf + 0x1a, avctx->width); // SPL
399 AV_WB16(buf + 0x1d, avctx->height >> ctx->interlaced); // NAL
401 buf[0x21] = ctx->cid_table->bit_depth == 10 ? 0x58 : 0x38;
402 buf[0x22] = 0x88 + (ctx->interlaced << 2);
403 AV_WB32(buf + 0x28, ctx->cid); // CID
404 buf[0x2c] = ctx->interlaced ? 0 : 0x80;
406 buf[0x5f] = 0x01; // UDL
408 buf[0x167] = 0x02; // reserved
409 AV_WB16(buf + 0x16a, ctx->m.mb_height * 4 + 4); // MSIPS
410 buf[0x16d] = ctx->m.mb_height; // Ns
411 buf[0x16f] = 0x10; // reserved
413 ctx->msip = buf + 0x170;
417 static av_always_inline void dnxhd_encode_dc(DNXHDEncContext *ctx, int diff)
421 nbits = av_log2_16bit(-2 * diff);
424 nbits = av_log2_16bit(2 * diff);
426 put_bits(&ctx->m.pb, ctx->cid_table->dc_bits[nbits] + nbits,
427 (ctx->cid_table->dc_codes[nbits] << nbits) +
428 (diff & ((1 << nbits) - 1)));
431 static av_always_inline
432 void dnxhd_encode_block(DNXHDEncContext *ctx, int16_t *block,
433 int last_index, int n)
435 int last_non_zero = 0;
438 dnxhd_encode_dc(ctx, block[0] - ctx->m.last_dc[n]);
439 ctx->m.last_dc[n] = block[0];
441 for (i = 1; i <= last_index; i++) {
442 j = ctx->m.intra_scantable.permutated[i];
445 int run_level = i - last_non_zero - 1;
446 int rlevel = (slevel << 1) | !!run_level;
447 put_bits(&ctx->m.pb, ctx->vlc_bits[rlevel], ctx->vlc_codes[rlevel]);
449 put_bits(&ctx->m.pb, ctx->run_bits[run_level],
450 ctx->run_codes[run_level]);
454 put_bits(&ctx->m.pb, ctx->vlc_bits[0], ctx->vlc_codes[0]); // EOB
457 static av_always_inline
458 void dnxhd_unquantize_c(DNXHDEncContext *ctx, int16_t *block, int n,
459 int qscale, int last_index)
461 const uint8_t *weight_matrix;
465 weight_matrix = (n & 2) ? ctx->cid_table->chroma_weight
466 : ctx->cid_table->luma_weight;
468 for (i = 1; i <= last_index; i++) {
469 int j = ctx->m.intra_scantable.permutated[i];
473 level = (1 - 2 * level) * qscale * weight_matrix[i];
474 if (ctx->cid_table->bit_depth == 10) {
475 if (weight_matrix[i] != 8)
479 if (weight_matrix[i] != 32)
485 level = (2 * level + 1) * qscale * weight_matrix[i];
486 if (ctx->cid_table->bit_depth == 10) {
487 if (weight_matrix[i] != 8)
491 if (weight_matrix[i] != 32)
501 static av_always_inline int dnxhd_ssd_block(int16_t *qblock, int16_t *block)
505 for (i = 0; i < 64; i++)
506 score += (block[i] - qblock[i]) * (block[i] - qblock[i]);
510 static av_always_inline
511 int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, int16_t *block, int last_index)
513 int last_non_zero = 0;
516 for (i = 1; i <= last_index; i++) {
517 j = ctx->m.intra_scantable.permutated[i];
520 int run_level = i - last_non_zero - 1;
521 bits += ctx->vlc_bits[(level << 1) |
522 !!run_level] + ctx->run_bits[run_level];
529 static av_always_inline
530 void dnxhd_get_blocks(DNXHDEncContext *ctx, int mb_x, int mb_y)
532 const int bs = ctx->block_width_l2;
533 const int bw = 1 << bs;
534 const uint8_t *ptr_y = ctx->thread[0]->src[0] +
535 ((mb_y << 4) * ctx->m.linesize) + (mb_x << bs + 1);
536 const uint8_t *ptr_u = ctx->thread[0]->src[1] +
537 ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs);
538 const uint8_t *ptr_v = ctx->thread[0]->src[2] +
539 ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs);
540 DSPContext *dsp = &ctx->m.dsp;
542 dsp->get_pixels(ctx->blocks[0], ptr_y, ctx->m.linesize);
543 dsp->get_pixels(ctx->blocks[1], ptr_y + bw, ctx->m.linesize);
544 dsp->get_pixels(ctx->blocks[2], ptr_u, ctx->m.uvlinesize);
545 dsp->get_pixels(ctx->blocks[3], ptr_v, ctx->m.uvlinesize);
547 if (mb_y + 1 == ctx->m.mb_height && ctx->m.avctx->height == 1080) {
548 if (ctx->interlaced) {
549 ctx->get_pixels_8x4_sym(ctx->blocks[4],
550 ptr_y + ctx->dct_y_offset,
552 ctx->get_pixels_8x4_sym(ctx->blocks[5],
553 ptr_y + ctx->dct_y_offset + bw,
555 ctx->get_pixels_8x4_sym(ctx->blocks[6],
556 ptr_u + ctx->dct_uv_offset,
558 ctx->get_pixels_8x4_sym(ctx->blocks[7],
559 ptr_v + ctx->dct_uv_offset,
562 ctx->bdsp.clear_block(ctx->blocks[4]);
563 ctx->bdsp.clear_block(ctx->blocks[5]);
564 ctx->bdsp.clear_block(ctx->blocks[6]);
565 ctx->bdsp.clear_block(ctx->blocks[7]);
568 dsp->get_pixels(ctx->blocks[4],
569 ptr_y + ctx->dct_y_offset, ctx->m.linesize);
570 dsp->get_pixels(ctx->blocks[5],
571 ptr_y + ctx->dct_y_offset + bw, ctx->m.linesize);
572 dsp->get_pixels(ctx->blocks[6],
573 ptr_u + ctx->dct_uv_offset, ctx->m.uvlinesize);
574 dsp->get_pixels(ctx->blocks[7],
575 ptr_v + ctx->dct_uv_offset, ctx->m.uvlinesize);
579 static av_always_inline
580 int dnxhd_switch_matrix(DNXHDEncContext *ctx, int i)
583 ctx->m.q_intra_matrix16 = ctx->qmatrix_c16;
584 ctx->m.q_intra_matrix = ctx->qmatrix_c;
587 ctx->m.q_intra_matrix16 = ctx->qmatrix_l16;
588 ctx->m.q_intra_matrix = ctx->qmatrix_l;
593 static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg,
594 int jobnr, int threadnr)
596 DNXHDEncContext *ctx = avctx->priv_data;
597 int mb_y = jobnr, mb_x;
598 int qscale = ctx->qscale;
599 LOCAL_ALIGNED_16(int16_t, block, [64]);
600 ctx = ctx->thread[threadnr];
604 ctx->m.last_dc[2] = 1 << (ctx->cid_table->bit_depth + 2);
606 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
607 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
613 dnxhd_get_blocks(ctx, mb_x, mb_y);
615 for (i = 0; i < 8; i++) {
616 int16_t *src_block = ctx->blocks[i];
617 int overflow, nbits, diff, last_index;
618 int n = dnxhd_switch_matrix(ctx, i);
620 memcpy(block, src_block, 64 * sizeof(*block));
621 last_index = ctx->m.dct_quantize(&ctx->m, block, i,
623 ac_bits += dnxhd_calc_ac_bits(ctx, block, last_index);
625 diff = block[0] - ctx->m.last_dc[n];
627 nbits = av_log2_16bit(-2 * diff);
629 nbits = av_log2_16bit(2 * diff);
631 assert(nbits < ctx->cid_table->bit_depth + 4);
632 dc_bits += ctx->cid_table->dc_bits[nbits] + nbits;
634 ctx->m.last_dc[n] = block[0];
636 if (avctx->mb_decision == FF_MB_DECISION_RD || !RC_VARIANCE) {
637 dnxhd_unquantize_c(ctx, block, i, qscale, last_index);
638 ctx->m.idsp.idct(block);
639 ssd += dnxhd_ssd_block(block, src_block);
642 ctx->mb_rc[qscale][mb].ssd = ssd;
643 ctx->mb_rc[qscale][mb].bits = ac_bits + dc_bits + 12 +
644 8 * ctx->vlc_bits[0];
649 static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg,
650 int jobnr, int threadnr)
652 DNXHDEncContext *ctx = avctx->priv_data;
653 int mb_y = jobnr, mb_x;
654 ctx = ctx->thread[threadnr];
655 init_put_bits(&ctx->m.pb, (uint8_t *)arg + 640 + ctx->slice_offs[jobnr],
656 ctx->slice_size[jobnr]);
660 ctx->m.last_dc[2] = 1 << (ctx->cid_table->bit_depth + 2);
661 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
662 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
663 int qscale = ctx->mb_qscale[mb];
666 put_bits(&ctx->m.pb, 12, qscale << 1);
668 dnxhd_get_blocks(ctx, mb_x, mb_y);
670 for (i = 0; i < 8; i++) {
671 int16_t *block = ctx->blocks[i];
672 int overflow, n = dnxhd_switch_matrix(ctx, i);
673 int last_index = ctx->m.dct_quantize(&ctx->m, block, i,
676 dnxhd_encode_block(ctx, block, last_index, n);
677 // STOP_TIMER("encode_block");
680 if (put_bits_count(&ctx->m.pb) & 31)
681 put_bits(&ctx->m.pb, 32 - (put_bits_count(&ctx->m.pb) & 31), 0);
682 flush_put_bits(&ctx->m.pb);
686 static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx)
690 for (mb_y = 0; mb_y < ctx->m.mb_height; mb_y++) {
692 ctx->slice_offs[mb_y] = offset;
693 ctx->slice_size[mb_y] = 0;
694 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
695 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
696 ctx->slice_size[mb_y] += ctx->mb_bits[mb];
698 ctx->slice_size[mb_y] = (ctx->slice_size[mb_y] + 31) & ~31;
699 ctx->slice_size[mb_y] >>= 3;
700 thread_size = ctx->slice_size[mb_y];
701 offset += thread_size;
705 static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg,
706 int jobnr, int threadnr)
708 DNXHDEncContext *ctx = avctx->priv_data;
709 int mb_y = jobnr, mb_x, x, y;
710 int partial_last_row = (mb_y == ctx->m.mb_height - 1) &&
711 ((avctx->height >> ctx->interlaced) & 0xF);
713 ctx = ctx->thread[threadnr];
714 if (ctx->cid_table->bit_depth == 8) {
715 uint8_t *pix = ctx->thread[0]->src[0] + ((mb_y << 4) * ctx->m.linesize);
716 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x, pix += 16) {
717 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
721 if (!partial_last_row && mb_x * 16 <= avctx->width - 16) {
722 sum = ctx->m.dsp.pix_sum(pix, ctx->m.linesize);
723 varc = ctx->m.dsp.pix_norm1(pix, ctx->m.linesize);
725 int bw = FFMIN(avctx->width - 16 * mb_x, 16);
726 int bh = FFMIN((avctx->height >> ctx->interlaced) - 16 * mb_y, 16);
728 for (y = 0; y < bh; y++) {
729 for (x = 0; x < bw; x++) {
730 uint8_t val = pix[x + y * ctx->m.linesize];
736 varc = (varc - (((unsigned) sum * sum) >> 8) + 128) >> 8;
738 ctx->mb_cmp[mb].value = varc;
739 ctx->mb_cmp[mb].mb = mb;
742 int const linesize = ctx->m.linesize >> 1;
743 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x) {
744 uint16_t *pix = (uint16_t *)ctx->thread[0]->src[0] +
745 ((mb_y << 4) * linesize) + (mb_x << 4);
746 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
751 // Macroblocks are 16x16 pixels, unlike DCT blocks which are 8x8.
752 for (i = 0; i < 16; ++i) {
753 for (j = 0; j < 16; ++j) {
754 // Turn 16-bit pixels into 10-bit ones.
755 int const sample = (unsigned) pix[j] >> 6;
757 sqsum += sample * sample;
758 // 2^10 * 2^10 * 16 * 16 = 2^28, which is less than INT_MAX
762 mean = sum >> 8; // 16*16 == 2^8
764 ctx->mb_cmp[mb].value = sqmean - mean * mean;
765 ctx->mb_cmp[mb].mb = mb;
771 static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx)
773 int lambda, up_step, down_step;
774 int last_lower = INT_MAX, last_higher = 0;
777 for (q = 1; q < avctx->qmax; q++) {
779 avctx->execute2(avctx, dnxhd_calc_bits_thread,
780 NULL, NULL, ctx->m.mb_height);
782 up_step = down_step = 2 << LAMBDA_FRAC_BITS;
783 lambda = ctx->lambda;
788 if (lambda == last_higher) {
790 end = 1; // need to set final qscales/bits
792 for (y = 0; y < ctx->m.mb_height; y++) {
793 for (x = 0; x < ctx->m.mb_width; x++) {
794 unsigned min = UINT_MAX;
796 int mb = y * ctx->m.mb_width + x;
797 for (q = 1; q < avctx->qmax; q++) {
798 unsigned score = ctx->mb_rc[q][mb].bits * lambda +
799 ((unsigned) ctx->mb_rc[q][mb].ssd << LAMBDA_FRAC_BITS);
805 bits += ctx->mb_rc[qscale][mb].bits;
806 ctx->mb_qscale[mb] = qscale;
807 ctx->mb_bits[mb] = ctx->mb_rc[qscale][mb].bits;
809 bits = (bits + 31) & ~31; // padding
810 if (bits > ctx->frame_bits)
813 // av_dlog(ctx->m.avctx,
814 // "lambda %d, up %u, down %u, bits %d, frame %d\n",
815 // lambda, last_higher, last_lower, bits, ctx->frame_bits);
817 if (bits > ctx->frame_bits)
818 return AVERROR(EINVAL);
821 if (bits < ctx->frame_bits) {
822 last_lower = FFMIN(lambda, last_lower);
823 if (last_higher != 0)
824 lambda = (lambda+last_higher)>>1;
827 down_step = FFMIN((int64_t)down_step*5, INT_MAX);
828 up_step = 1<<LAMBDA_FRAC_BITS;
829 lambda = FFMAX(1, lambda);
830 if (lambda == last_lower)
833 last_higher = FFMAX(lambda, last_higher);
834 if (last_lower != INT_MAX)
835 lambda = (lambda+last_lower)>>1;
836 else if ((int64_t)lambda + up_step > INT_MAX)
837 return AVERROR(EINVAL);
840 up_step = FFMIN((int64_t)up_step*5, INT_MAX);
841 down_step = 1<<LAMBDA_FRAC_BITS;
844 //av_dlog(ctx->m.avctx, "out lambda %d\n", lambda);
845 ctx->lambda = lambda;
849 static int dnxhd_find_qscale(DNXHDEncContext *ctx)
855 int last_lower = INT_MAX;
859 qscale = ctx->qscale;
862 ctx->qscale = qscale;
863 // XXX avoid recalculating bits
864 ctx->m.avctx->execute2(ctx->m.avctx, dnxhd_calc_bits_thread,
865 NULL, NULL, ctx->m.mb_height);
866 for (y = 0; y < ctx->m.mb_height; y++) {
867 for (x = 0; x < ctx->m.mb_width; x++)
868 bits += ctx->mb_rc[qscale][y*ctx->m.mb_width+x].bits;
869 bits = (bits+31)&~31; // padding
870 if (bits > ctx->frame_bits)
873 // av_dlog(ctx->m.avctx,
874 // "%d, qscale %d, bits %d, frame %d, higher %d, lower %d\n",
875 // ctx->m.avctx->frame_number, qscale, bits, ctx->frame_bits,
876 // last_higher, last_lower);
877 if (bits < ctx->frame_bits) {
880 if (last_higher == qscale - 1) {
881 qscale = last_higher;
884 last_lower = FFMIN(qscale, last_lower);
885 if (last_higher != 0)
886 qscale = (qscale + last_higher) >> 1;
888 qscale -= down_step++;
893 if (last_lower == qscale + 1)
895 last_higher = FFMAX(qscale, last_higher);
896 if (last_lower != INT_MAX)
897 qscale = (qscale + last_lower) >> 1;
901 if (qscale >= ctx->m.avctx->qmax)
902 return AVERROR(EINVAL);
905 //av_dlog(ctx->m.avctx, "out qscale %d\n", qscale);
906 ctx->qscale = qscale;
910 #define BUCKET_BITS 8
911 #define RADIX_PASSES 4
912 #define NBUCKETS (1 << BUCKET_BITS)
914 static inline int get_bucket(int value, int shift)
917 value &= NBUCKETS - 1;
918 return NBUCKETS - 1 - value;
921 static void radix_count(const RCCMPEntry *data, int size,
922 int buckets[RADIX_PASSES][NBUCKETS])
925 memset(buckets, 0, sizeof(buckets[0][0]) * RADIX_PASSES * NBUCKETS);
926 for (i = 0; i < size; i++) {
927 int v = data[i].value;
928 for (j = 0; j < RADIX_PASSES; j++) {
929 buckets[j][get_bucket(v, 0)]++;
934 for (j = 0; j < RADIX_PASSES; j++) {
936 for (i = NBUCKETS - 1; i >= 0; i--)
937 buckets[j][i] = offset -= buckets[j][i];
938 assert(!buckets[j][0]);
942 static void radix_sort_pass(RCCMPEntry *dst, const RCCMPEntry *data,
943 int size, int buckets[NBUCKETS], int pass)
945 int shift = pass * BUCKET_BITS;
947 for (i = 0; i < size; i++) {
948 int v = get_bucket(data[i].value, shift);
949 int pos = buckets[v]++;
954 static void radix_sort(RCCMPEntry *data, int size)
956 int buckets[RADIX_PASSES][NBUCKETS];
957 RCCMPEntry *tmp = av_malloc(sizeof(*tmp) * size);
958 radix_count(data, size, buckets);
959 radix_sort_pass(tmp, data, size, buckets[0], 0);
960 radix_sort_pass(data, tmp, size, buckets[1], 1);
961 if (buckets[2][NBUCKETS - 1] || buckets[3][NBUCKETS - 1]) {
962 radix_sort_pass(tmp, data, size, buckets[2], 2);
963 radix_sort_pass(data, tmp, size, buckets[3], 3);
968 static int dnxhd_encode_fast(AVCodecContext *avctx, DNXHDEncContext *ctx)
972 if ((ret = dnxhd_find_qscale(ctx)) < 0)
974 for (y = 0; y < ctx->m.mb_height; y++) {
975 for (x = 0; x < ctx->m.mb_width; x++) {
976 int mb = y * ctx->m.mb_width + x;
978 ctx->mb_qscale[mb] = ctx->qscale;
979 ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale][mb].bits;
980 max_bits += ctx->mb_rc[ctx->qscale][mb].bits;
982 delta_bits = ctx->mb_rc[ctx->qscale][mb].bits -
983 ctx->mb_rc[ctx->qscale + 1][mb].bits;
984 ctx->mb_cmp[mb].mb = mb;
985 ctx->mb_cmp[mb].value =
986 delta_bits ? ((ctx->mb_rc[ctx->qscale][mb].ssd -
987 ctx->mb_rc[ctx->qscale + 1][mb].ssd) * 100) /
989 : INT_MIN; // avoid increasing qscale
992 max_bits += 31; // worst padding
996 avctx->execute2(avctx, dnxhd_mb_var_thread,
997 NULL, NULL, ctx->m.mb_height);
998 radix_sort(ctx->mb_cmp, ctx->m.mb_num);
999 for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) {
1000 int mb = ctx->mb_cmp[x].mb;
1001 max_bits -= ctx->mb_rc[ctx->qscale][mb].bits -
1002 ctx->mb_rc[ctx->qscale + 1][mb].bits;
1003 ctx->mb_qscale[mb] = ctx->qscale + 1;
1004 ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale + 1][mb].bits;
1010 static void dnxhd_load_picture(DNXHDEncContext *ctx, const AVFrame *frame)
1014 for (i = 0; i < ctx->m.avctx->thread_count; i++) {
1015 ctx->thread[i]->m.linesize = frame->linesize[0] << ctx->interlaced;
1016 ctx->thread[i]->m.uvlinesize = frame->linesize[1] << ctx->interlaced;
1017 ctx->thread[i]->dct_y_offset = ctx->m.linesize *8;
1018 ctx->thread[i]->dct_uv_offset = ctx->m.uvlinesize*8;
1021 ctx->m.avctx->coded_frame->interlaced_frame = frame->interlaced_frame;
1022 ctx->cur_field = frame->interlaced_frame && !frame->top_field_first;
1025 static int dnxhd_encode_picture(AVCodecContext *avctx, AVPacket *pkt,
1026 const AVFrame *frame, int *got_packet)
1028 DNXHDEncContext *ctx = avctx->priv_data;
1029 int first_field = 1;
1033 if ((ret = ff_alloc_packet(pkt, ctx->cid_table->frame_size)) < 0) {
1034 av_log(avctx, AV_LOG_ERROR,
1035 "output buffer is too small to compress picture\n");
1040 dnxhd_load_picture(ctx, frame);
1043 for (i = 0; i < 3; i++) {
1044 ctx->src[i] = frame->data[i];
1045 if (ctx->interlaced && ctx->cur_field)
1046 ctx->src[i] += frame->linesize[i];
1049 dnxhd_write_header(avctx, buf);
1051 if (avctx->mb_decision == FF_MB_DECISION_RD)
1052 ret = dnxhd_encode_rdo(avctx, ctx);
1054 ret = dnxhd_encode_fast(avctx, ctx);
1056 av_log(avctx, AV_LOG_ERROR,
1057 "picture could not fit ratecontrol constraints, increase qmax\n");
1061 dnxhd_setup_threads_slices(ctx);
1064 for (i = 0; i < ctx->m.mb_height; i++) {
1065 AV_WB32(ctx->msip + i * 4, offset);
1066 offset += ctx->slice_size[i];
1067 assert(!(ctx->slice_size[i] & 3));
1070 avctx->execute2(avctx, dnxhd_encode_thread, buf, NULL, ctx->m.mb_height);
1072 assert(640 + offset + 4 <= ctx->cid_table->coding_unit_size);
1073 memset(buf + 640 + offset, 0,
1074 ctx->cid_table->coding_unit_size - 4 - offset - 640);
1076 AV_WB32(buf + ctx->cid_table->coding_unit_size - 4, 0x600DC0DE); // EOF
1078 if (ctx->interlaced && first_field) {
1080 ctx->cur_field ^= 1;
1081 buf += ctx->cid_table->coding_unit_size;
1082 goto encode_coding_unit;
1085 avctx->coded_frame->quality = ctx->qscale * FF_QP2LAMBDA;
1087 pkt->flags |= AV_PKT_FLAG_KEY;
1092 static av_cold int dnxhd_encode_end(AVCodecContext *avctx)
1094 DNXHDEncContext *ctx = avctx->priv_data;
1095 int max_level = 1 << (ctx->cid_table->bit_depth + 2);
1098 av_free(ctx->vlc_codes - max_level * 2);
1099 av_free(ctx->vlc_bits - max_level * 2);
1100 av_freep(&ctx->run_codes);
1101 av_freep(&ctx->run_bits);
1103 av_freep(&ctx->mb_bits);
1104 av_freep(&ctx->mb_qscale);
1105 av_freep(&ctx->mb_rc);
1106 av_freep(&ctx->mb_cmp);
1107 av_freep(&ctx->slice_size);
1108 av_freep(&ctx->slice_offs);
1110 av_freep(&ctx->qmatrix_c);
1111 av_freep(&ctx->qmatrix_l);
1112 av_freep(&ctx->qmatrix_c16);
1113 av_freep(&ctx->qmatrix_l16);
1115 for (i = 1; i < avctx->thread_count; i++)
1116 av_freep(&ctx->thread[i]);
1118 av_frame_free(&avctx->coded_frame);
1123 AVCodec ff_dnxhd_encoder = {
1125 .long_name = NULL_IF_CONFIG_SMALL("VC3/DNxHD"),
1126 .type = AVMEDIA_TYPE_VIDEO,
1127 .id = AV_CODEC_ID_DNXHD,
1128 .priv_data_size = sizeof(DNXHDEncContext),
1129 .init = dnxhd_encode_init,
1130 .encode2 = dnxhd_encode_picture,
1131 .close = dnxhd_encode_end,
1132 .capabilities = CODEC_CAP_SLICE_THREADS,
1133 .pix_fmts = (const enum AVPixelFormat[]) {
1135 AV_PIX_FMT_YUV422P10,
1138 .priv_class = &class,