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.dsp.idct_permutation[ff_zigzag_direct[i]];
204 weight_matrix[j] = ctx->cid_table->luma_weight[i];
206 ff_convert_matrix(&ctx->m.dsp, 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.dsp.idct_permutation[ff_zigzag_direct[i]];
211 weight_matrix[j] = ctx->cid_table->chroma_weight[i];
213 ff_convert_matrix(&ctx->m.dsp, 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.dsp.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_dct_common_init(&ctx->m);
312 if (!ctx->m.dct_quantize)
313 ctx->m.dct_quantize = ff_dct_quantize_c;
315 if (ctx->cid_table->bit_depth == 10) {
316 ctx->m.dct_quantize = dnxhd_10bit_dct_quantize;
317 ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym;
318 ctx->block_width_l2 = 4;
320 ctx->get_pixels_8x4_sym = dnxhd_8bit_get_pixels_8x4_sym;
321 ctx->block_width_l2 = 3;
325 ff_dnxhdenc_init_x86(ctx);
327 ctx->m.mb_height = (avctx->height + 15) / 16;
328 ctx->m.mb_width = (avctx->width + 15) / 16;
330 if (avctx->flags & CODEC_FLAG_INTERLACED_DCT) {
332 ctx->m.mb_height /= 2;
335 ctx->m.mb_num = ctx->m.mb_height * ctx->m.mb_width;
337 if (avctx->intra_quant_bias != FF_DEFAULT_QUANT_BIAS)
338 ctx->m.intra_quant_bias = avctx->intra_quant_bias;
339 // XXX tune lbias/cbias
340 if ((ret = dnxhd_init_qmat(ctx, ctx->m.intra_quant_bias, 0)) < 0)
343 /* Avid Nitris hardware decoder requires a minimum amount of padding
344 * in the coding unit payload */
345 if (ctx->nitris_compat)
346 ctx->min_padding = 1600;
348 if ((ret = dnxhd_init_vlc(ctx)) < 0)
350 if ((ret = dnxhd_init_rc(ctx)) < 0)
353 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_size,
354 ctx->m.mb_height * sizeof(uint32_t), fail);
355 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_offs,
356 ctx->m.mb_height * sizeof(uint32_t), fail);
357 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_bits,
358 ctx->m.mb_num * sizeof(uint16_t), fail);
359 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_qscale,
360 ctx->m.mb_num * sizeof(uint8_t), fail);
362 avctx->coded_frame = av_frame_alloc();
363 if (!avctx->coded_frame)
364 return AVERROR(ENOMEM);
366 avctx->coded_frame->key_frame = 1;
367 avctx->coded_frame->pict_type = AV_PICTURE_TYPE_I;
369 if (avctx->thread_count > MAX_THREADS) {
370 av_log(avctx, AV_LOG_ERROR, "too many threads\n");
371 return AVERROR(EINVAL);
374 ctx->thread[0] = ctx;
375 for (i = 1; i < avctx->thread_count; i++) {
376 ctx->thread[i] = av_malloc(sizeof(DNXHDEncContext));
377 memcpy(ctx->thread[i], ctx, sizeof(DNXHDEncContext));
381 fail: // for FF_ALLOCZ_OR_GOTO
382 return AVERROR(ENOMEM);
385 static int dnxhd_write_header(AVCodecContext *avctx, uint8_t *buf)
387 DNXHDEncContext *ctx = avctx->priv_data;
388 const uint8_t header_prefix[5] = { 0x00, 0x00, 0x02, 0x80, 0x01 };
392 memcpy(buf, header_prefix, 5);
393 buf[5] = ctx->interlaced ? ctx->cur_field + 2 : 0x01;
394 buf[6] = 0x80; // crc flag off
395 buf[7] = 0xa0; // reserved
396 AV_WB16(buf + 0x18, avctx->height >> ctx->interlaced); // ALPF
397 AV_WB16(buf + 0x1a, avctx->width); // SPL
398 AV_WB16(buf + 0x1d, avctx->height >> ctx->interlaced); // NAL
400 buf[0x21] = ctx->cid_table->bit_depth == 10 ? 0x58 : 0x38;
401 buf[0x22] = 0x88 + (ctx->interlaced << 2);
402 AV_WB32(buf + 0x28, ctx->cid); // CID
403 buf[0x2c] = ctx->interlaced ? 0 : 0x80;
405 buf[0x5f] = 0x01; // UDL
407 buf[0x167] = 0x02; // reserved
408 AV_WB16(buf + 0x16a, ctx->m.mb_height * 4 + 4); // MSIPS
409 buf[0x16d] = ctx->m.mb_height; // Ns
410 buf[0x16f] = 0x10; // reserved
412 ctx->msip = buf + 0x170;
416 static av_always_inline void dnxhd_encode_dc(DNXHDEncContext *ctx, int diff)
420 nbits = av_log2_16bit(-2 * diff);
423 nbits = av_log2_16bit(2 * diff);
425 put_bits(&ctx->m.pb, ctx->cid_table->dc_bits[nbits] + nbits,
426 (ctx->cid_table->dc_codes[nbits] << nbits) +
427 (diff & ((1 << nbits) - 1)));
430 static av_always_inline
431 void dnxhd_encode_block(DNXHDEncContext *ctx, int16_t *block,
432 int last_index, int n)
434 int last_non_zero = 0;
437 dnxhd_encode_dc(ctx, block[0] - ctx->m.last_dc[n]);
438 ctx->m.last_dc[n] = block[0];
440 for (i = 1; i <= last_index; i++) {
441 j = ctx->m.intra_scantable.permutated[i];
444 int run_level = i - last_non_zero - 1;
445 int rlevel = (slevel << 1) | !!run_level;
446 put_bits(&ctx->m.pb, ctx->vlc_bits[rlevel], ctx->vlc_codes[rlevel]);
448 put_bits(&ctx->m.pb, ctx->run_bits[run_level],
449 ctx->run_codes[run_level]);
453 put_bits(&ctx->m.pb, ctx->vlc_bits[0], ctx->vlc_codes[0]); // EOB
456 static av_always_inline
457 void dnxhd_unquantize_c(DNXHDEncContext *ctx, int16_t *block, int n,
458 int qscale, int last_index)
460 const uint8_t *weight_matrix;
464 weight_matrix = (n & 2) ? ctx->cid_table->chroma_weight
465 : ctx->cid_table->luma_weight;
467 for (i = 1; i <= last_index; i++) {
468 int j = ctx->m.intra_scantable.permutated[i];
472 level = (1 - 2 * level) * qscale * weight_matrix[i];
473 if (ctx->cid_table->bit_depth == 10) {
474 if (weight_matrix[i] != 8)
478 if (weight_matrix[i] != 32)
484 level = (2 * level + 1) * qscale * weight_matrix[i];
485 if (ctx->cid_table->bit_depth == 10) {
486 if (weight_matrix[i] != 8)
490 if (weight_matrix[i] != 32)
500 static av_always_inline int dnxhd_ssd_block(int16_t *qblock, int16_t *block)
504 for (i = 0; i < 64; i++)
505 score += (block[i] - qblock[i]) * (block[i] - qblock[i]);
509 static av_always_inline
510 int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, int16_t *block, int last_index)
512 int last_non_zero = 0;
515 for (i = 1; i <= last_index; i++) {
516 j = ctx->m.intra_scantable.permutated[i];
519 int run_level = i - last_non_zero - 1;
520 bits += ctx->vlc_bits[(level << 1) |
521 !!run_level] + ctx->run_bits[run_level];
528 static av_always_inline
529 void dnxhd_get_blocks(DNXHDEncContext *ctx, int mb_x, int mb_y)
531 const int bs = ctx->block_width_l2;
532 const int bw = 1 << bs;
533 const uint8_t *ptr_y = ctx->thread[0]->src[0] +
534 ((mb_y << 4) * ctx->m.linesize) + (mb_x << bs + 1);
535 const uint8_t *ptr_u = ctx->thread[0]->src[1] +
536 ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs);
537 const uint8_t *ptr_v = ctx->thread[0]->src[2] +
538 ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs);
539 DSPContext *dsp = &ctx->m.dsp;
541 dsp->get_pixels(ctx->blocks[0], ptr_y, ctx->m.linesize);
542 dsp->get_pixels(ctx->blocks[1], ptr_y + bw, ctx->m.linesize);
543 dsp->get_pixels(ctx->blocks[2], ptr_u, ctx->m.uvlinesize);
544 dsp->get_pixels(ctx->blocks[3], ptr_v, ctx->m.uvlinesize);
546 if (mb_y + 1 == ctx->m.mb_height && ctx->m.avctx->height == 1080) {
547 if (ctx->interlaced) {
548 ctx->get_pixels_8x4_sym(ctx->blocks[4],
549 ptr_y + ctx->dct_y_offset,
551 ctx->get_pixels_8x4_sym(ctx->blocks[5],
552 ptr_y + ctx->dct_y_offset + bw,
554 ctx->get_pixels_8x4_sym(ctx->blocks[6],
555 ptr_u + ctx->dct_uv_offset,
557 ctx->get_pixels_8x4_sym(ctx->blocks[7],
558 ptr_v + ctx->dct_uv_offset,
561 ctx->bdsp.clear_block(ctx->blocks[4]);
562 ctx->bdsp.clear_block(ctx->blocks[5]);
563 ctx->bdsp.clear_block(ctx->blocks[6]);
564 ctx->bdsp.clear_block(ctx->blocks[7]);
567 dsp->get_pixels(ctx->blocks[4],
568 ptr_y + ctx->dct_y_offset, ctx->m.linesize);
569 dsp->get_pixels(ctx->blocks[5],
570 ptr_y + ctx->dct_y_offset + bw, ctx->m.linesize);
571 dsp->get_pixels(ctx->blocks[6],
572 ptr_u + ctx->dct_uv_offset, ctx->m.uvlinesize);
573 dsp->get_pixels(ctx->blocks[7],
574 ptr_v + ctx->dct_uv_offset, ctx->m.uvlinesize);
578 static av_always_inline
579 int dnxhd_switch_matrix(DNXHDEncContext *ctx, int i)
582 ctx->m.q_intra_matrix16 = ctx->qmatrix_c16;
583 ctx->m.q_intra_matrix = ctx->qmatrix_c;
586 ctx->m.q_intra_matrix16 = ctx->qmatrix_l16;
587 ctx->m.q_intra_matrix = ctx->qmatrix_l;
592 static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg,
593 int jobnr, int threadnr)
595 DNXHDEncContext *ctx = avctx->priv_data;
596 int mb_y = jobnr, mb_x;
597 int qscale = ctx->qscale;
598 LOCAL_ALIGNED_16(int16_t, block, [64]);
599 ctx = ctx->thread[threadnr];
603 ctx->m.last_dc[2] = 1 << (ctx->cid_table->bit_depth + 2);
605 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
606 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
612 dnxhd_get_blocks(ctx, mb_x, mb_y);
614 for (i = 0; i < 8; i++) {
615 int16_t *src_block = ctx->blocks[i];
616 int overflow, nbits, diff, last_index;
617 int n = dnxhd_switch_matrix(ctx, i);
619 memcpy(block, src_block, 64 * sizeof(*block));
620 last_index = ctx->m.dct_quantize(&ctx->m, block, i,
622 ac_bits += dnxhd_calc_ac_bits(ctx, block, last_index);
624 diff = block[0] - ctx->m.last_dc[n];
626 nbits = av_log2_16bit(-2 * diff);
628 nbits = av_log2_16bit(2 * diff);
630 assert(nbits < ctx->cid_table->bit_depth + 4);
631 dc_bits += ctx->cid_table->dc_bits[nbits] + nbits;
633 ctx->m.last_dc[n] = block[0];
635 if (avctx->mb_decision == FF_MB_DECISION_RD || !RC_VARIANCE) {
636 dnxhd_unquantize_c(ctx, block, i, qscale, last_index);
637 ctx->m.dsp.idct(block);
638 ssd += dnxhd_ssd_block(block, src_block);
641 ctx->mb_rc[qscale][mb].ssd = ssd;
642 ctx->mb_rc[qscale][mb].bits = ac_bits + dc_bits + 12 +
643 8 * ctx->vlc_bits[0];
648 static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg,
649 int jobnr, int threadnr)
651 DNXHDEncContext *ctx = avctx->priv_data;
652 int mb_y = jobnr, mb_x;
653 ctx = ctx->thread[threadnr];
654 init_put_bits(&ctx->m.pb, (uint8_t *)arg + 640 + ctx->slice_offs[jobnr],
655 ctx->slice_size[jobnr]);
659 ctx->m.last_dc[2] = 1 << (ctx->cid_table->bit_depth + 2);
660 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
661 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
662 int qscale = ctx->mb_qscale[mb];
665 put_bits(&ctx->m.pb, 12, qscale << 1);
667 dnxhd_get_blocks(ctx, mb_x, mb_y);
669 for (i = 0; i < 8; i++) {
670 int16_t *block = ctx->blocks[i];
671 int overflow, n = dnxhd_switch_matrix(ctx, i);
672 int last_index = ctx->m.dct_quantize(&ctx->m, block, i,
675 dnxhd_encode_block(ctx, block, last_index, n);
676 // STOP_TIMER("encode_block");
679 if (put_bits_count(&ctx->m.pb) & 31)
680 put_bits(&ctx->m.pb, 32 - (put_bits_count(&ctx->m.pb) & 31), 0);
681 flush_put_bits(&ctx->m.pb);
685 static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx)
689 for (mb_y = 0; mb_y < ctx->m.mb_height; mb_y++) {
691 ctx->slice_offs[mb_y] = offset;
692 ctx->slice_size[mb_y] = 0;
693 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
694 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
695 ctx->slice_size[mb_y] += ctx->mb_bits[mb];
697 ctx->slice_size[mb_y] = (ctx->slice_size[mb_y] + 31) & ~31;
698 ctx->slice_size[mb_y] >>= 3;
699 thread_size = ctx->slice_size[mb_y];
700 offset += thread_size;
704 static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg,
705 int jobnr, int threadnr)
707 DNXHDEncContext *ctx = avctx->priv_data;
708 int mb_y = jobnr, mb_x, x, y;
709 int partial_last_row = (mb_y == ctx->m.mb_height - 1) &&
710 ((avctx->height >> ctx->interlaced) & 0xF);
712 ctx = ctx->thread[threadnr];
713 if (ctx->cid_table->bit_depth == 8) {
714 uint8_t *pix = ctx->thread[0]->src[0] + ((mb_y << 4) * ctx->m.linesize);
715 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x, pix += 16) {
716 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
720 if (!partial_last_row && mb_x * 16 <= avctx->width - 16) {
721 sum = ctx->m.dsp.pix_sum(pix, ctx->m.linesize);
722 varc = ctx->m.dsp.pix_norm1(pix, ctx->m.linesize);
724 int bw = FFMIN(avctx->width - 16 * mb_x, 16);
725 int bh = FFMIN((avctx->height >> ctx->interlaced) - 16 * mb_y, 16);
727 for (y = 0; y < bh; y++) {
728 for (x = 0; x < bw; x++) {
729 uint8_t val = pix[x + y * ctx->m.linesize];
735 varc = (varc - (((unsigned) sum * sum) >> 8) + 128) >> 8;
737 ctx->mb_cmp[mb].value = varc;
738 ctx->mb_cmp[mb].mb = mb;
741 int const linesize = ctx->m.linesize >> 1;
742 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x) {
743 uint16_t *pix = (uint16_t *)ctx->thread[0]->src[0] +
744 ((mb_y << 4) * linesize) + (mb_x << 4);
745 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
750 // Macroblocks are 16x16 pixels, unlike DCT blocks which are 8x8.
751 for (i = 0; i < 16; ++i) {
752 for (j = 0; j < 16; ++j) {
753 // Turn 16-bit pixels into 10-bit ones.
754 int const sample = (unsigned) pix[j] >> 6;
756 sqsum += sample * sample;
757 // 2^10 * 2^10 * 16 * 16 = 2^28, which is less than INT_MAX
761 mean = sum >> 8; // 16*16 == 2^8
763 ctx->mb_cmp[mb].value = sqmean - mean * mean;
764 ctx->mb_cmp[mb].mb = mb;
770 static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx)
772 int lambda, up_step, down_step;
773 int last_lower = INT_MAX, last_higher = 0;
776 for (q = 1; q < avctx->qmax; q++) {
778 avctx->execute2(avctx, dnxhd_calc_bits_thread,
779 NULL, NULL, ctx->m.mb_height);
781 up_step = down_step = 2 << LAMBDA_FRAC_BITS;
782 lambda = ctx->lambda;
787 if (lambda == last_higher) {
789 end = 1; // need to set final qscales/bits
791 for (y = 0; y < ctx->m.mb_height; y++) {
792 for (x = 0; x < ctx->m.mb_width; x++) {
793 unsigned min = UINT_MAX;
795 int mb = y * ctx->m.mb_width + x;
796 for (q = 1; q < avctx->qmax; q++) {
797 unsigned score = ctx->mb_rc[q][mb].bits * lambda +
798 ((unsigned) ctx->mb_rc[q][mb].ssd << LAMBDA_FRAC_BITS);
804 bits += ctx->mb_rc[qscale][mb].bits;
805 ctx->mb_qscale[mb] = qscale;
806 ctx->mb_bits[mb] = ctx->mb_rc[qscale][mb].bits;
808 bits = (bits + 31) & ~31; // padding
809 if (bits > ctx->frame_bits)
812 // av_dlog(ctx->m.avctx,
813 // "lambda %d, up %u, down %u, bits %d, frame %d\n",
814 // lambda, last_higher, last_lower, bits, ctx->frame_bits);
816 if (bits > ctx->frame_bits)
817 return AVERROR(EINVAL);
820 if (bits < ctx->frame_bits) {
821 last_lower = FFMIN(lambda, last_lower);
822 if (last_higher != 0)
823 lambda = (lambda+last_higher)>>1;
826 down_step = FFMIN((int64_t)down_step*5, INT_MAX);
827 up_step = 1<<LAMBDA_FRAC_BITS;
828 lambda = FFMAX(1, lambda);
829 if (lambda == last_lower)
832 last_higher = FFMAX(lambda, last_higher);
833 if (last_lower != INT_MAX)
834 lambda = (lambda+last_lower)>>1;
835 else if ((int64_t)lambda + up_step > INT_MAX)
836 return AVERROR(EINVAL);
839 up_step = FFMIN((int64_t)up_step*5, INT_MAX);
840 down_step = 1<<LAMBDA_FRAC_BITS;
843 //av_dlog(ctx->m.avctx, "out lambda %d\n", lambda);
844 ctx->lambda = lambda;
848 static int dnxhd_find_qscale(DNXHDEncContext *ctx)
854 int last_lower = INT_MAX;
858 qscale = ctx->qscale;
861 ctx->qscale = qscale;
862 // XXX avoid recalculating bits
863 ctx->m.avctx->execute2(ctx->m.avctx, dnxhd_calc_bits_thread,
864 NULL, NULL, ctx->m.mb_height);
865 for (y = 0; y < ctx->m.mb_height; y++) {
866 for (x = 0; x < ctx->m.mb_width; x++)
867 bits += ctx->mb_rc[qscale][y*ctx->m.mb_width+x].bits;
868 bits = (bits+31)&~31; // padding
869 if (bits > ctx->frame_bits)
872 // av_dlog(ctx->m.avctx,
873 // "%d, qscale %d, bits %d, frame %d, higher %d, lower %d\n",
874 // ctx->m.avctx->frame_number, qscale, bits, ctx->frame_bits,
875 // last_higher, last_lower);
876 if (bits < ctx->frame_bits) {
879 if (last_higher == qscale - 1) {
880 qscale = last_higher;
883 last_lower = FFMIN(qscale, last_lower);
884 if (last_higher != 0)
885 qscale = (qscale + last_higher) >> 1;
887 qscale -= down_step++;
892 if (last_lower == qscale + 1)
894 last_higher = FFMAX(qscale, last_higher);
895 if (last_lower != INT_MAX)
896 qscale = (qscale + last_lower) >> 1;
900 if (qscale >= ctx->m.avctx->qmax)
901 return AVERROR(EINVAL);
904 //av_dlog(ctx->m.avctx, "out qscale %d\n", qscale);
905 ctx->qscale = qscale;
909 #define BUCKET_BITS 8
910 #define RADIX_PASSES 4
911 #define NBUCKETS (1 << BUCKET_BITS)
913 static inline int get_bucket(int value, int shift)
916 value &= NBUCKETS - 1;
917 return NBUCKETS - 1 - value;
920 static void radix_count(const RCCMPEntry *data, int size,
921 int buckets[RADIX_PASSES][NBUCKETS])
924 memset(buckets, 0, sizeof(buckets[0][0]) * RADIX_PASSES * NBUCKETS);
925 for (i = 0; i < size; i++) {
926 int v = data[i].value;
927 for (j = 0; j < RADIX_PASSES; j++) {
928 buckets[j][get_bucket(v, 0)]++;
933 for (j = 0; j < RADIX_PASSES; j++) {
935 for (i = NBUCKETS - 1; i >= 0; i--)
936 buckets[j][i] = offset -= buckets[j][i];
937 assert(!buckets[j][0]);
941 static void radix_sort_pass(RCCMPEntry *dst, const RCCMPEntry *data,
942 int size, int buckets[NBUCKETS], int pass)
944 int shift = pass * BUCKET_BITS;
946 for (i = 0; i < size; i++) {
947 int v = get_bucket(data[i].value, shift);
948 int pos = buckets[v]++;
953 static void radix_sort(RCCMPEntry *data, int size)
955 int buckets[RADIX_PASSES][NBUCKETS];
956 RCCMPEntry *tmp = av_malloc(sizeof(*tmp) * size);
957 radix_count(data, size, buckets);
958 radix_sort_pass(tmp, data, size, buckets[0], 0);
959 radix_sort_pass(data, tmp, size, buckets[1], 1);
960 if (buckets[2][NBUCKETS - 1] || buckets[3][NBUCKETS - 1]) {
961 radix_sort_pass(tmp, data, size, buckets[2], 2);
962 radix_sort_pass(data, tmp, size, buckets[3], 3);
967 static int dnxhd_encode_fast(AVCodecContext *avctx, DNXHDEncContext *ctx)
971 if ((ret = dnxhd_find_qscale(ctx)) < 0)
973 for (y = 0; y < ctx->m.mb_height; y++) {
974 for (x = 0; x < ctx->m.mb_width; x++) {
975 int mb = y * ctx->m.mb_width + x;
977 ctx->mb_qscale[mb] = ctx->qscale;
978 ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale][mb].bits;
979 max_bits += ctx->mb_rc[ctx->qscale][mb].bits;
981 delta_bits = ctx->mb_rc[ctx->qscale][mb].bits -
982 ctx->mb_rc[ctx->qscale + 1][mb].bits;
983 ctx->mb_cmp[mb].mb = mb;
984 ctx->mb_cmp[mb].value =
985 delta_bits ? ((ctx->mb_rc[ctx->qscale][mb].ssd -
986 ctx->mb_rc[ctx->qscale + 1][mb].ssd) * 100) /
988 : INT_MIN; // avoid increasing qscale
991 max_bits += 31; // worst padding
995 avctx->execute2(avctx, dnxhd_mb_var_thread,
996 NULL, NULL, ctx->m.mb_height);
997 radix_sort(ctx->mb_cmp, ctx->m.mb_num);
998 for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) {
999 int mb = ctx->mb_cmp[x].mb;
1000 max_bits -= ctx->mb_rc[ctx->qscale][mb].bits -
1001 ctx->mb_rc[ctx->qscale + 1][mb].bits;
1002 ctx->mb_qscale[mb] = ctx->qscale + 1;
1003 ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale + 1][mb].bits;
1009 static void dnxhd_load_picture(DNXHDEncContext *ctx, const AVFrame *frame)
1013 for (i = 0; i < ctx->m.avctx->thread_count; i++) {
1014 ctx->thread[i]->m.linesize = frame->linesize[0] << ctx->interlaced;
1015 ctx->thread[i]->m.uvlinesize = frame->linesize[1] << ctx->interlaced;
1016 ctx->thread[i]->dct_y_offset = ctx->m.linesize *8;
1017 ctx->thread[i]->dct_uv_offset = ctx->m.uvlinesize*8;
1020 ctx->m.avctx->coded_frame->interlaced_frame = frame->interlaced_frame;
1021 ctx->cur_field = frame->interlaced_frame && !frame->top_field_first;
1024 static int dnxhd_encode_picture(AVCodecContext *avctx, AVPacket *pkt,
1025 const AVFrame *frame, int *got_packet)
1027 DNXHDEncContext *ctx = avctx->priv_data;
1028 int first_field = 1;
1032 if ((ret = ff_alloc_packet(pkt, ctx->cid_table->frame_size)) < 0) {
1033 av_log(avctx, AV_LOG_ERROR,
1034 "output buffer is too small to compress picture\n");
1039 dnxhd_load_picture(ctx, frame);
1042 for (i = 0; i < 3; i++) {
1043 ctx->src[i] = frame->data[i];
1044 if (ctx->interlaced && ctx->cur_field)
1045 ctx->src[i] += frame->linesize[i];
1048 dnxhd_write_header(avctx, buf);
1050 if (avctx->mb_decision == FF_MB_DECISION_RD)
1051 ret = dnxhd_encode_rdo(avctx, ctx);
1053 ret = dnxhd_encode_fast(avctx, ctx);
1055 av_log(avctx, AV_LOG_ERROR,
1056 "picture could not fit ratecontrol constraints, increase qmax\n");
1060 dnxhd_setup_threads_slices(ctx);
1063 for (i = 0; i < ctx->m.mb_height; i++) {
1064 AV_WB32(ctx->msip + i * 4, offset);
1065 offset += ctx->slice_size[i];
1066 assert(!(ctx->slice_size[i] & 3));
1069 avctx->execute2(avctx, dnxhd_encode_thread, buf, NULL, ctx->m.mb_height);
1071 assert(640 + offset + 4 <= ctx->cid_table->coding_unit_size);
1072 memset(buf + 640 + offset, 0,
1073 ctx->cid_table->coding_unit_size - 4 - offset - 640);
1075 AV_WB32(buf + ctx->cid_table->coding_unit_size - 4, 0x600DC0DE); // EOF
1077 if (ctx->interlaced && first_field) {
1079 ctx->cur_field ^= 1;
1080 buf += ctx->cid_table->coding_unit_size;
1081 goto encode_coding_unit;
1084 avctx->coded_frame->quality = ctx->qscale * FF_QP2LAMBDA;
1086 pkt->flags |= AV_PKT_FLAG_KEY;
1091 static av_cold int dnxhd_encode_end(AVCodecContext *avctx)
1093 DNXHDEncContext *ctx = avctx->priv_data;
1094 int max_level = 1 << (ctx->cid_table->bit_depth + 2);
1097 av_free(ctx->vlc_codes - max_level * 2);
1098 av_free(ctx->vlc_bits - max_level * 2);
1099 av_freep(&ctx->run_codes);
1100 av_freep(&ctx->run_bits);
1102 av_freep(&ctx->mb_bits);
1103 av_freep(&ctx->mb_qscale);
1104 av_freep(&ctx->mb_rc);
1105 av_freep(&ctx->mb_cmp);
1106 av_freep(&ctx->slice_size);
1107 av_freep(&ctx->slice_offs);
1109 av_freep(&ctx->qmatrix_c);
1110 av_freep(&ctx->qmatrix_l);
1111 av_freep(&ctx->qmatrix_c16);
1112 av_freep(&ctx->qmatrix_l16);
1114 for (i = 1; i < avctx->thread_count; i++)
1115 av_freep(&ctx->thread[i]);
1117 av_frame_free(&avctx->coded_frame);
1122 AVCodec ff_dnxhd_encoder = {
1124 .long_name = NULL_IF_CONFIG_SMALL("VC3/DNxHD"),
1125 .type = AVMEDIA_TYPE_VIDEO,
1126 .id = AV_CODEC_ID_DNXHD,
1127 .priv_data_size = sizeof(DNXHDEncContext),
1128 .init = dnxhd_encode_init,
1129 .encode2 = dnxhd_encode_picture,
1130 .close = dnxhd_encode_end,
1131 .capabilities = CODEC_CAP_SLICE_THREADS,
1132 .pix_fmts = (const enum AVPixelFormat[]) {
1134 AV_PIX_FMT_YUV422P10,
1137 .priv_class = &class,