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"
36 #include "pixblockdsp.h"
39 // The largest value that will not lead to overflow for 10bit samples.
40 #define DNX10BIT_QMAT_SHIFT 18
41 #define RC_VARIANCE 1 // use variance or ssd for fast rc
42 #define LAMBDA_FRAC_BITS 10
44 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
45 static const AVOption options[] = {
46 { "nitris_compat", "encode with Avid Nitris compatibility",
47 offsetof(DNXHDEncContext, nitris_compat), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 1, VE },
51 static const AVClass class = {
58 static void dnxhd_8bit_get_pixels_8x4_sym(int16_t *restrict block,
59 const uint8_t *pixels,
63 for (i = 0; i < 4; i++) {
75 memcpy(block, block - 8, sizeof(*block) * 8);
76 memcpy(block + 8, block - 16, sizeof(*block) * 8);
77 memcpy(block + 16, block - 24, sizeof(*block) * 8);
78 memcpy(block + 24, block - 32, sizeof(*block) * 8);
81 static av_always_inline
82 void dnxhd_10bit_get_pixels_8x4_sym(int16_t *restrict block,
83 const uint8_t *pixels,
90 for (i = 0; i < 4; i++) {
91 memcpy(block + i * 8, pixels + i * line_size, 8 * sizeof(*block));
92 memcpy(block - (i + 1) * 8, pixels + i * line_size, 8 * sizeof(*block));
96 static int dnxhd_10bit_dct_quantize(MpegEncContext *ctx, int16_t *block,
97 int n, int qscale, int *overflow)
99 const uint8_t *scantable= ctx->intra_scantable.scantable;
100 const int *qmat = ctx->q_intra_matrix[qscale];
101 int last_non_zero = 0;
104 ctx->fdsp.fdct(block);
106 // Divide by 4 with rounding, to compensate scaling of DCT coefficients
107 block[0] = (block[0] + 2) >> 2;
109 for (i = 1; i < 64; ++i) {
110 int j = scantable[i];
111 int sign = block[j] >> 31;
112 int level = (block[j] ^ sign) - sign;
113 level = level * qmat[j] >> DNX10BIT_QMAT_SHIFT;
114 block[j] = (level ^ sign) - sign;
119 return last_non_zero;
122 static av_cold int dnxhd_init_vlc(DNXHDEncContext *ctx)
124 int i, j, level, run;
125 int max_level = 1 << (ctx->cid_table->bit_depth + 2);
127 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->vlc_codes,
128 max_level * 4 * sizeof(*ctx->vlc_codes), fail);
129 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->vlc_bits,
130 max_level * 4 * sizeof(*ctx->vlc_bits), fail);
131 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_codes,
133 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_bits,
136 ctx->vlc_codes += max_level * 2;
137 ctx->vlc_bits += max_level * 2;
138 for (level = -max_level; level < max_level; level++) {
139 for (run = 0; run < 2; run++) {
140 int index = (level << 1) | run;
141 int sign, offset = 0, alevel = level;
143 MASK_ABS(sign, alevel);
145 offset = (alevel - 1) >> 6;
146 alevel -= offset << 6;
148 for (j = 0; j < 257; j++) {
149 if (ctx->cid_table->ac_level[j] == alevel &&
150 (!offset || (ctx->cid_table->ac_index_flag[j] && offset)) &&
151 (!run || (ctx->cid_table->ac_run_flag [j] && run))) {
152 assert(!ctx->vlc_codes[index]);
154 ctx->vlc_codes[index] =
155 (ctx->cid_table->ac_codes[j] << 1) | (sign & 1);
156 ctx->vlc_bits[index] = ctx->cid_table->ac_bits[j] + 1;
158 ctx->vlc_codes[index] = ctx->cid_table->ac_codes[j];
159 ctx->vlc_bits[index] = ctx->cid_table->ac_bits[j];
164 assert(!alevel || j < 257);
166 ctx->vlc_codes[index] =
167 (ctx->vlc_codes[index] << ctx->cid_table->index_bits) | offset;
168 ctx->vlc_bits[index] += ctx->cid_table->index_bits;
172 for (i = 0; i < 62; i++) {
173 int run = ctx->cid_table->run[i];
175 ctx->run_codes[run] = ctx->cid_table->run_codes[i];
176 ctx->run_bits[run] = ctx->cid_table->run_bits[i];
180 return AVERROR(ENOMEM);
183 static av_cold int dnxhd_init_qmat(DNXHDEncContext *ctx, int lbias, int cbias)
185 // init first elem to 1 to avoid div by 0 in convert_matrix
186 uint16_t weight_matrix[64] = { 1, }; // convert_matrix needs uint16_t*
188 const uint8_t *luma_weight_table = ctx->cid_table->luma_weight;
189 const uint8_t *chroma_weight_table = ctx->cid_table->chroma_weight;
191 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l,
192 (ctx->m.avctx->qmax + 1) * 64 * sizeof(int), fail);
193 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c,
194 (ctx->m.avctx->qmax + 1) * 64 * sizeof(int), fail);
195 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l16,
196 (ctx->m.avctx->qmax + 1) * 64 * 2 * sizeof(uint16_t),
198 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c16,
199 (ctx->m.avctx->qmax + 1) * 64 * 2 * sizeof(uint16_t),
202 if (ctx->cid_table->bit_depth == 8) {
203 for (i = 1; i < 64; i++) {
204 int j = ctx->m.idsp.idct_permutation[ff_zigzag_direct[i]];
205 weight_matrix[j] = ctx->cid_table->luma_weight[i];
207 ff_convert_matrix(&ctx->m, ctx->qmatrix_l, ctx->qmatrix_l16,
208 weight_matrix, ctx->m.intra_quant_bias, 1,
209 ctx->m.avctx->qmax, 1);
210 for (i = 1; i < 64; i++) {
211 int j = ctx->m.idsp.idct_permutation[ff_zigzag_direct[i]];
212 weight_matrix[j] = ctx->cid_table->chroma_weight[i];
214 ff_convert_matrix(&ctx->m, ctx->qmatrix_c, ctx->qmatrix_c16,
215 weight_matrix, ctx->m.intra_quant_bias, 1,
216 ctx->m.avctx->qmax, 1);
218 for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
219 for (i = 0; i < 64; i++) {
220 ctx->qmatrix_l[qscale][i] <<= 2;
221 ctx->qmatrix_c[qscale][i] <<= 2;
222 ctx->qmatrix_l16[qscale][0][i] <<= 2;
223 ctx->qmatrix_l16[qscale][1][i] <<= 2;
224 ctx->qmatrix_c16[qscale][0][i] <<= 2;
225 ctx->qmatrix_c16[qscale][1][i] <<= 2;
230 for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
231 for (i = 1; i < 64; i++) {
232 int j = ctx->m.idsp.idct_permutation[ff_zigzag_direct[i]];
234 /* The quantization formula from the VC-3 standard is:
235 * quantized = sign(block[i]) * floor(abs(block[i]/s) * p /
236 * (qscale * weight_table[i]))
237 * Where p is 32 for 8-bit samples and 8 for 10-bit ones.
238 * The s factor compensates scaling of DCT coefficients done by
239 * the DCT routines, and therefore is not present in standard.
240 * It's 8 for 8-bit samples and 4 for 10-bit ones.
241 * We want values of ctx->qtmatrix_l and ctx->qtmatrix_r to be:
242 * ((1 << DNX10BIT_QMAT_SHIFT) * (p / s)) /
243 * (qscale * weight_table[i])
244 * For 10-bit samples, p / s == 2 */
245 ctx->qmatrix_l[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) /
246 (qscale * luma_weight_table[i]);
247 ctx->qmatrix_c[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) /
248 (qscale * chroma_weight_table[i]);
255 return AVERROR(ENOMEM);
258 static av_cold int dnxhd_init_rc(DNXHDEncContext *ctx)
260 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_rc,
261 8160 * ctx->m.avctx->qmax * sizeof(RCEntry), fail);
262 if (ctx->m.avctx->mb_decision != FF_MB_DECISION_RD)
263 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_cmp,
264 ctx->m.mb_num * sizeof(RCCMPEntry), fail);
266 ctx->frame_bits = (ctx->cid_table->coding_unit_size -
267 640 - 4 - ctx->min_padding) * 8;
269 ctx->lambda = 2 << LAMBDA_FRAC_BITS; // qscale 2
272 return AVERROR(ENOMEM);
275 static av_cold int dnxhd_encode_init(AVCodecContext *avctx)
277 DNXHDEncContext *ctx = avctx->priv_data;
278 int i, index, bit_depth, ret;
280 switch (avctx->pix_fmt) {
281 case AV_PIX_FMT_YUV422P:
284 case AV_PIX_FMT_YUV422P10:
288 av_log(avctx, AV_LOG_ERROR,
289 "pixel format is incompatible with DNxHD\n");
290 return AVERROR(EINVAL);
293 ctx->cid = ff_dnxhd_find_cid(avctx, bit_depth);
295 av_log(avctx, AV_LOG_ERROR,
296 "video parameters incompatible with DNxHD\n");
297 return AVERROR(EINVAL);
299 av_log(avctx, AV_LOG_DEBUG, "cid %d\n", ctx->cid);
301 index = ff_dnxhd_get_cid_table(ctx->cid);
302 ctx->cid_table = &ff_dnxhd_cid_table[index];
304 ctx->m.avctx = avctx;
308 avctx->bits_per_raw_sample = ctx->cid_table->bit_depth;
310 ff_blockdsp_init(&ctx->bdsp, avctx);
311 ff_fdctdsp_init(&ctx->m.fdsp, avctx);
312 ff_idctdsp_init(&ctx->m.idsp, avctx);
313 ff_mpegvideoencdsp_init(&ctx->m.mpvencdsp, avctx);
314 ff_pixblockdsp_init(&ctx->m.pdsp, avctx);
315 ff_dct_common_init(&ctx->m);
316 if (!ctx->m.dct_quantize)
317 ctx->m.dct_quantize = ff_dct_quantize_c;
319 if (ctx->cid_table->bit_depth == 10) {
320 ctx->m.dct_quantize = dnxhd_10bit_dct_quantize;
321 ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym;
322 ctx->block_width_l2 = 4;
324 ctx->get_pixels_8x4_sym = dnxhd_8bit_get_pixels_8x4_sym;
325 ctx->block_width_l2 = 3;
329 ff_dnxhdenc_init_x86(ctx);
331 ctx->m.mb_height = (avctx->height + 15) / 16;
332 ctx->m.mb_width = (avctx->width + 15) / 16;
334 if (avctx->flags & CODEC_FLAG_INTERLACED_DCT) {
336 ctx->m.mb_height /= 2;
339 ctx->m.mb_num = ctx->m.mb_height * ctx->m.mb_width;
341 if (avctx->intra_quant_bias != FF_DEFAULT_QUANT_BIAS)
342 ctx->m.intra_quant_bias = avctx->intra_quant_bias;
343 // XXX tune lbias/cbias
344 if ((ret = dnxhd_init_qmat(ctx, ctx->m.intra_quant_bias, 0)) < 0)
347 /* Avid Nitris hardware decoder requires a minimum amount of padding
348 * in the coding unit payload */
349 if (ctx->nitris_compat)
350 ctx->min_padding = 1600;
352 if ((ret = dnxhd_init_vlc(ctx)) < 0)
354 if ((ret = dnxhd_init_rc(ctx)) < 0)
357 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_size,
358 ctx->m.mb_height * sizeof(uint32_t), fail);
359 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_offs,
360 ctx->m.mb_height * sizeof(uint32_t), fail);
361 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_bits,
362 ctx->m.mb_num * sizeof(uint16_t), fail);
363 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_qscale,
364 ctx->m.mb_num * sizeof(uint8_t), fail);
366 avctx->coded_frame = av_frame_alloc();
367 if (!avctx->coded_frame)
368 return AVERROR(ENOMEM);
370 avctx->coded_frame->key_frame = 1;
371 avctx->coded_frame->pict_type = AV_PICTURE_TYPE_I;
373 if (avctx->thread_count > MAX_THREADS) {
374 av_log(avctx, AV_LOG_ERROR, "too many threads\n");
375 return AVERROR(EINVAL);
378 ctx->thread[0] = ctx;
379 for (i = 1; i < avctx->thread_count; i++) {
380 ctx->thread[i] = av_malloc(sizeof(DNXHDEncContext));
381 memcpy(ctx->thread[i], ctx, sizeof(DNXHDEncContext));
385 fail: // for FF_ALLOCZ_OR_GOTO
386 return AVERROR(ENOMEM);
389 static int dnxhd_write_header(AVCodecContext *avctx, uint8_t *buf)
391 DNXHDEncContext *ctx = avctx->priv_data;
392 const uint8_t header_prefix[5] = { 0x00, 0x00, 0x02, 0x80, 0x01 };
396 memcpy(buf, header_prefix, 5);
397 buf[5] = ctx->interlaced ? ctx->cur_field + 2 : 0x01;
398 buf[6] = 0x80; // crc flag off
399 buf[7] = 0xa0; // reserved
400 AV_WB16(buf + 0x18, avctx->height >> ctx->interlaced); // ALPF
401 AV_WB16(buf + 0x1a, avctx->width); // SPL
402 AV_WB16(buf + 0x1d, avctx->height >> ctx->interlaced); // NAL
404 buf[0x21] = ctx->cid_table->bit_depth == 10 ? 0x58 : 0x38;
405 buf[0x22] = 0x88 + (ctx->interlaced << 2);
406 AV_WB32(buf + 0x28, ctx->cid); // CID
407 buf[0x2c] = ctx->interlaced ? 0 : 0x80;
409 buf[0x5f] = 0x01; // UDL
411 buf[0x167] = 0x02; // reserved
412 AV_WB16(buf + 0x16a, ctx->m.mb_height * 4 + 4); // MSIPS
413 buf[0x16d] = ctx->m.mb_height; // Ns
414 buf[0x16f] = 0x10; // reserved
416 ctx->msip = buf + 0x170;
420 static av_always_inline void dnxhd_encode_dc(DNXHDEncContext *ctx, int diff)
424 nbits = av_log2_16bit(-2 * diff);
427 nbits = av_log2_16bit(2 * diff);
429 put_bits(&ctx->m.pb, ctx->cid_table->dc_bits[nbits] + nbits,
430 (ctx->cid_table->dc_codes[nbits] << nbits) +
431 (diff & ((1 << nbits) - 1)));
434 static av_always_inline
435 void dnxhd_encode_block(DNXHDEncContext *ctx, int16_t *block,
436 int last_index, int n)
438 int last_non_zero = 0;
441 dnxhd_encode_dc(ctx, block[0] - ctx->m.last_dc[n]);
442 ctx->m.last_dc[n] = block[0];
444 for (i = 1; i <= last_index; i++) {
445 j = ctx->m.intra_scantable.permutated[i];
448 int run_level = i - last_non_zero - 1;
449 int rlevel = (slevel << 1) | !!run_level;
450 put_bits(&ctx->m.pb, ctx->vlc_bits[rlevel], ctx->vlc_codes[rlevel]);
452 put_bits(&ctx->m.pb, ctx->run_bits[run_level],
453 ctx->run_codes[run_level]);
457 put_bits(&ctx->m.pb, ctx->vlc_bits[0], ctx->vlc_codes[0]); // EOB
460 static av_always_inline
461 void dnxhd_unquantize_c(DNXHDEncContext *ctx, int16_t *block, int n,
462 int qscale, int last_index)
464 const uint8_t *weight_matrix;
468 weight_matrix = (n & 2) ? ctx->cid_table->chroma_weight
469 : ctx->cid_table->luma_weight;
471 for (i = 1; i <= last_index; i++) {
472 int j = ctx->m.intra_scantable.permutated[i];
476 level = (1 - 2 * level) * qscale * weight_matrix[i];
477 if (ctx->cid_table->bit_depth == 10) {
478 if (weight_matrix[i] != 8)
482 if (weight_matrix[i] != 32)
488 level = (2 * level + 1) * qscale * weight_matrix[i];
489 if (ctx->cid_table->bit_depth == 10) {
490 if (weight_matrix[i] != 8)
494 if (weight_matrix[i] != 32)
504 static av_always_inline int dnxhd_ssd_block(int16_t *qblock, int16_t *block)
508 for (i = 0; i < 64; i++)
509 score += (block[i] - qblock[i]) * (block[i] - qblock[i]);
513 static av_always_inline
514 int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, int16_t *block, int last_index)
516 int last_non_zero = 0;
519 for (i = 1; i <= last_index; i++) {
520 j = ctx->m.intra_scantable.permutated[i];
523 int run_level = i - last_non_zero - 1;
524 bits += ctx->vlc_bits[(level << 1) |
525 !!run_level] + ctx->run_bits[run_level];
532 static av_always_inline
533 void dnxhd_get_blocks(DNXHDEncContext *ctx, int mb_x, int mb_y)
535 const int bs = ctx->block_width_l2;
536 const int bw = 1 << bs;
537 const uint8_t *ptr_y = ctx->thread[0]->src[0] +
538 ((mb_y << 4) * ctx->m.linesize) + (mb_x << bs + 1);
539 const uint8_t *ptr_u = ctx->thread[0]->src[1] +
540 ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs);
541 const uint8_t *ptr_v = ctx->thread[0]->src[2] +
542 ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs);
543 PixblockDSPContext *pdsp = &ctx->m.pdsp;
545 pdsp->get_pixels(ctx->blocks[0], ptr_y, ctx->m.linesize);
546 pdsp->get_pixels(ctx->blocks[1], ptr_y + bw, ctx->m.linesize);
547 pdsp->get_pixels(ctx->blocks[2], ptr_u, ctx->m.uvlinesize);
548 pdsp->get_pixels(ctx->blocks[3], ptr_v, ctx->m.uvlinesize);
550 if (mb_y + 1 == ctx->m.mb_height && ctx->m.avctx->height == 1080) {
551 if (ctx->interlaced) {
552 ctx->get_pixels_8x4_sym(ctx->blocks[4],
553 ptr_y + ctx->dct_y_offset,
555 ctx->get_pixels_8x4_sym(ctx->blocks[5],
556 ptr_y + ctx->dct_y_offset + bw,
558 ctx->get_pixels_8x4_sym(ctx->blocks[6],
559 ptr_u + ctx->dct_uv_offset,
561 ctx->get_pixels_8x4_sym(ctx->blocks[7],
562 ptr_v + ctx->dct_uv_offset,
565 ctx->bdsp.clear_block(ctx->blocks[4]);
566 ctx->bdsp.clear_block(ctx->blocks[5]);
567 ctx->bdsp.clear_block(ctx->blocks[6]);
568 ctx->bdsp.clear_block(ctx->blocks[7]);
571 pdsp->get_pixels(ctx->blocks[4],
572 ptr_y + ctx->dct_y_offset, ctx->m.linesize);
573 pdsp->get_pixels(ctx->blocks[5],
574 ptr_y + ctx->dct_y_offset + bw, ctx->m.linesize);
575 pdsp->get_pixels(ctx->blocks[6],
576 ptr_u + ctx->dct_uv_offset, ctx->m.uvlinesize);
577 pdsp->get_pixels(ctx->blocks[7],
578 ptr_v + ctx->dct_uv_offset, ctx->m.uvlinesize);
582 static av_always_inline
583 int dnxhd_switch_matrix(DNXHDEncContext *ctx, int i)
586 ctx->m.q_intra_matrix16 = ctx->qmatrix_c16;
587 ctx->m.q_intra_matrix = ctx->qmatrix_c;
590 ctx->m.q_intra_matrix16 = ctx->qmatrix_l16;
591 ctx->m.q_intra_matrix = ctx->qmatrix_l;
596 static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg,
597 int jobnr, int threadnr)
599 DNXHDEncContext *ctx = avctx->priv_data;
600 int mb_y = jobnr, mb_x;
601 int qscale = ctx->qscale;
602 LOCAL_ALIGNED_16(int16_t, block, [64]);
603 ctx = ctx->thread[threadnr];
607 ctx->m.last_dc[2] = 1 << (ctx->cid_table->bit_depth + 2);
609 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
610 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
616 dnxhd_get_blocks(ctx, mb_x, mb_y);
618 for (i = 0; i < 8; i++) {
619 int16_t *src_block = ctx->blocks[i];
620 int overflow, nbits, diff, last_index;
621 int n = dnxhd_switch_matrix(ctx, i);
623 memcpy(block, src_block, 64 * sizeof(*block));
624 last_index = ctx->m.dct_quantize(&ctx->m, block, i,
626 ac_bits += dnxhd_calc_ac_bits(ctx, block, last_index);
628 diff = block[0] - ctx->m.last_dc[n];
630 nbits = av_log2_16bit(-2 * diff);
632 nbits = av_log2_16bit(2 * diff);
634 assert(nbits < ctx->cid_table->bit_depth + 4);
635 dc_bits += ctx->cid_table->dc_bits[nbits] + nbits;
637 ctx->m.last_dc[n] = block[0];
639 if (avctx->mb_decision == FF_MB_DECISION_RD || !RC_VARIANCE) {
640 dnxhd_unquantize_c(ctx, block, i, qscale, last_index);
641 ctx->m.idsp.idct(block);
642 ssd += dnxhd_ssd_block(block, src_block);
645 ctx->mb_rc[qscale][mb].ssd = ssd;
646 ctx->mb_rc[qscale][mb].bits = ac_bits + dc_bits + 12 +
647 8 * ctx->vlc_bits[0];
652 static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg,
653 int jobnr, int threadnr)
655 DNXHDEncContext *ctx = avctx->priv_data;
656 int mb_y = jobnr, mb_x;
657 ctx = ctx->thread[threadnr];
658 init_put_bits(&ctx->m.pb, (uint8_t *)arg + 640 + ctx->slice_offs[jobnr],
659 ctx->slice_size[jobnr]);
663 ctx->m.last_dc[2] = 1 << (ctx->cid_table->bit_depth + 2);
664 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
665 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
666 int qscale = ctx->mb_qscale[mb];
669 put_bits(&ctx->m.pb, 12, qscale << 1);
671 dnxhd_get_blocks(ctx, mb_x, mb_y);
673 for (i = 0; i < 8; i++) {
674 int16_t *block = ctx->blocks[i];
675 int overflow, n = dnxhd_switch_matrix(ctx, i);
676 int last_index = ctx->m.dct_quantize(&ctx->m, block, i,
679 dnxhd_encode_block(ctx, block, last_index, n);
680 // STOP_TIMER("encode_block");
683 if (put_bits_count(&ctx->m.pb) & 31)
684 put_bits(&ctx->m.pb, 32 - (put_bits_count(&ctx->m.pb) & 31), 0);
685 flush_put_bits(&ctx->m.pb);
689 static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx)
693 for (mb_y = 0; mb_y < ctx->m.mb_height; mb_y++) {
695 ctx->slice_offs[mb_y] = offset;
696 ctx->slice_size[mb_y] = 0;
697 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
698 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
699 ctx->slice_size[mb_y] += ctx->mb_bits[mb];
701 ctx->slice_size[mb_y] = (ctx->slice_size[mb_y] + 31) & ~31;
702 ctx->slice_size[mb_y] >>= 3;
703 thread_size = ctx->slice_size[mb_y];
704 offset += thread_size;
708 static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg,
709 int jobnr, int threadnr)
711 DNXHDEncContext *ctx = avctx->priv_data;
712 int mb_y = jobnr, mb_x, x, y;
713 int partial_last_row = (mb_y == ctx->m.mb_height - 1) &&
714 ((avctx->height >> ctx->interlaced) & 0xF);
716 ctx = ctx->thread[threadnr];
717 if (ctx->cid_table->bit_depth == 8) {
718 uint8_t *pix = ctx->thread[0]->src[0] + ((mb_y << 4) * ctx->m.linesize);
719 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x, pix += 16) {
720 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
724 if (!partial_last_row && mb_x * 16 <= avctx->width - 16) {
725 sum = ctx->m.mpvencdsp.pix_sum(pix, ctx->m.linesize);
726 varc = ctx->m.mpvencdsp.pix_norm1(pix, ctx->m.linesize);
728 int bw = FFMIN(avctx->width - 16 * mb_x, 16);
729 int bh = FFMIN((avctx->height >> ctx->interlaced) - 16 * mb_y, 16);
731 for (y = 0; y < bh; y++) {
732 for (x = 0; x < bw; x++) {
733 uint8_t val = pix[x + y * ctx->m.linesize];
739 varc = (varc - (((unsigned) sum * sum) >> 8) + 128) >> 8;
741 ctx->mb_cmp[mb].value = varc;
742 ctx->mb_cmp[mb].mb = mb;
745 int const linesize = ctx->m.linesize >> 1;
746 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x) {
747 uint16_t *pix = (uint16_t *)ctx->thread[0]->src[0] +
748 ((mb_y << 4) * linesize) + (mb_x << 4);
749 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
754 // Macroblocks are 16x16 pixels, unlike DCT blocks which are 8x8.
755 for (i = 0; i < 16; ++i) {
756 for (j = 0; j < 16; ++j) {
757 // Turn 16-bit pixels into 10-bit ones.
758 int const sample = (unsigned) pix[j] >> 6;
760 sqsum += sample * sample;
761 // 2^10 * 2^10 * 16 * 16 = 2^28, which is less than INT_MAX
765 mean = sum >> 8; // 16*16 == 2^8
767 ctx->mb_cmp[mb].value = sqmean - mean * mean;
768 ctx->mb_cmp[mb].mb = mb;
774 static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx)
776 int lambda, up_step, down_step;
777 int last_lower = INT_MAX, last_higher = 0;
780 for (q = 1; q < avctx->qmax; q++) {
782 avctx->execute2(avctx, dnxhd_calc_bits_thread,
783 NULL, NULL, ctx->m.mb_height);
785 up_step = down_step = 2 << LAMBDA_FRAC_BITS;
786 lambda = ctx->lambda;
791 if (lambda == last_higher) {
793 end = 1; // need to set final qscales/bits
795 for (y = 0; y < ctx->m.mb_height; y++) {
796 for (x = 0; x < ctx->m.mb_width; x++) {
797 unsigned min = UINT_MAX;
799 int mb = y * ctx->m.mb_width + x;
800 for (q = 1; q < avctx->qmax; q++) {
801 unsigned score = ctx->mb_rc[q][mb].bits * lambda +
802 ((unsigned) ctx->mb_rc[q][mb].ssd << LAMBDA_FRAC_BITS);
808 bits += ctx->mb_rc[qscale][mb].bits;
809 ctx->mb_qscale[mb] = qscale;
810 ctx->mb_bits[mb] = ctx->mb_rc[qscale][mb].bits;
812 bits = (bits + 31) & ~31; // padding
813 if (bits > ctx->frame_bits)
816 // av_dlog(ctx->m.avctx,
817 // "lambda %d, up %u, down %u, bits %d, frame %d\n",
818 // lambda, last_higher, last_lower, bits, ctx->frame_bits);
820 if (bits > ctx->frame_bits)
821 return AVERROR(EINVAL);
824 if (bits < ctx->frame_bits) {
825 last_lower = FFMIN(lambda, last_lower);
826 if (last_higher != 0)
827 lambda = (lambda+last_higher)>>1;
830 down_step = FFMIN((int64_t)down_step*5, INT_MAX);
831 up_step = 1<<LAMBDA_FRAC_BITS;
832 lambda = FFMAX(1, lambda);
833 if (lambda == last_lower)
836 last_higher = FFMAX(lambda, last_higher);
837 if (last_lower != INT_MAX)
838 lambda = (lambda+last_lower)>>1;
839 else if ((int64_t)lambda + up_step > INT_MAX)
840 return AVERROR(EINVAL);
843 up_step = FFMIN((int64_t)up_step*5, INT_MAX);
844 down_step = 1<<LAMBDA_FRAC_BITS;
847 //av_dlog(ctx->m.avctx, "out lambda %d\n", lambda);
848 ctx->lambda = lambda;
852 static int dnxhd_find_qscale(DNXHDEncContext *ctx)
858 int last_lower = INT_MAX;
862 qscale = ctx->qscale;
865 ctx->qscale = qscale;
866 // XXX avoid recalculating bits
867 ctx->m.avctx->execute2(ctx->m.avctx, dnxhd_calc_bits_thread,
868 NULL, NULL, ctx->m.mb_height);
869 for (y = 0; y < ctx->m.mb_height; y++) {
870 for (x = 0; x < ctx->m.mb_width; x++)
871 bits += ctx->mb_rc[qscale][y*ctx->m.mb_width+x].bits;
872 bits = (bits+31)&~31; // padding
873 if (bits > ctx->frame_bits)
876 // av_dlog(ctx->m.avctx,
877 // "%d, qscale %d, bits %d, frame %d, higher %d, lower %d\n",
878 // ctx->m.avctx->frame_number, qscale, bits, ctx->frame_bits,
879 // last_higher, last_lower);
880 if (bits < ctx->frame_bits) {
883 if (last_higher == qscale - 1) {
884 qscale = last_higher;
887 last_lower = FFMIN(qscale, last_lower);
888 if (last_higher != 0)
889 qscale = (qscale + last_higher) >> 1;
891 qscale -= down_step++;
896 if (last_lower == qscale + 1)
898 last_higher = FFMAX(qscale, last_higher);
899 if (last_lower != INT_MAX)
900 qscale = (qscale + last_lower) >> 1;
904 if (qscale >= ctx->m.avctx->qmax)
905 return AVERROR(EINVAL);
908 //av_dlog(ctx->m.avctx, "out qscale %d\n", qscale);
909 ctx->qscale = qscale;
913 #define BUCKET_BITS 8
914 #define RADIX_PASSES 4
915 #define NBUCKETS (1 << BUCKET_BITS)
917 static inline int get_bucket(int value, int shift)
920 value &= NBUCKETS - 1;
921 return NBUCKETS - 1 - value;
924 static void radix_count(const RCCMPEntry *data, int size,
925 int buckets[RADIX_PASSES][NBUCKETS])
928 memset(buckets, 0, sizeof(buckets[0][0]) * RADIX_PASSES * NBUCKETS);
929 for (i = 0; i < size; i++) {
930 int v = data[i].value;
931 for (j = 0; j < RADIX_PASSES; j++) {
932 buckets[j][get_bucket(v, 0)]++;
937 for (j = 0; j < RADIX_PASSES; j++) {
939 for (i = NBUCKETS - 1; i >= 0; i--)
940 buckets[j][i] = offset -= buckets[j][i];
941 assert(!buckets[j][0]);
945 static void radix_sort_pass(RCCMPEntry *dst, const RCCMPEntry *data,
946 int size, int buckets[NBUCKETS], int pass)
948 int shift = pass * BUCKET_BITS;
950 for (i = 0; i < size; i++) {
951 int v = get_bucket(data[i].value, shift);
952 int pos = buckets[v]++;
957 static void radix_sort(RCCMPEntry *data, int size)
959 int buckets[RADIX_PASSES][NBUCKETS];
960 RCCMPEntry *tmp = av_malloc(sizeof(*tmp) * size);
961 radix_count(data, size, buckets);
962 radix_sort_pass(tmp, data, size, buckets[0], 0);
963 radix_sort_pass(data, tmp, size, buckets[1], 1);
964 if (buckets[2][NBUCKETS - 1] || buckets[3][NBUCKETS - 1]) {
965 radix_sort_pass(tmp, data, size, buckets[2], 2);
966 radix_sort_pass(data, tmp, size, buckets[3], 3);
971 static int dnxhd_encode_fast(AVCodecContext *avctx, DNXHDEncContext *ctx)
975 if ((ret = dnxhd_find_qscale(ctx)) < 0)
977 for (y = 0; y < ctx->m.mb_height; y++) {
978 for (x = 0; x < ctx->m.mb_width; x++) {
979 int mb = y * ctx->m.mb_width + x;
981 ctx->mb_qscale[mb] = ctx->qscale;
982 ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale][mb].bits;
983 max_bits += ctx->mb_rc[ctx->qscale][mb].bits;
985 delta_bits = ctx->mb_rc[ctx->qscale][mb].bits -
986 ctx->mb_rc[ctx->qscale + 1][mb].bits;
987 ctx->mb_cmp[mb].mb = mb;
988 ctx->mb_cmp[mb].value =
989 delta_bits ? ((ctx->mb_rc[ctx->qscale][mb].ssd -
990 ctx->mb_rc[ctx->qscale + 1][mb].ssd) * 100) /
992 : INT_MIN; // avoid increasing qscale
995 max_bits += 31; // worst padding
999 avctx->execute2(avctx, dnxhd_mb_var_thread,
1000 NULL, NULL, ctx->m.mb_height);
1001 radix_sort(ctx->mb_cmp, ctx->m.mb_num);
1002 for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) {
1003 int mb = ctx->mb_cmp[x].mb;
1004 max_bits -= ctx->mb_rc[ctx->qscale][mb].bits -
1005 ctx->mb_rc[ctx->qscale + 1][mb].bits;
1006 ctx->mb_qscale[mb] = ctx->qscale + 1;
1007 ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale + 1][mb].bits;
1013 static void dnxhd_load_picture(DNXHDEncContext *ctx, const AVFrame *frame)
1017 for (i = 0; i < ctx->m.avctx->thread_count; i++) {
1018 ctx->thread[i]->m.linesize = frame->linesize[0] << ctx->interlaced;
1019 ctx->thread[i]->m.uvlinesize = frame->linesize[1] << ctx->interlaced;
1020 ctx->thread[i]->dct_y_offset = ctx->m.linesize *8;
1021 ctx->thread[i]->dct_uv_offset = ctx->m.uvlinesize*8;
1024 ctx->m.avctx->coded_frame->interlaced_frame = frame->interlaced_frame;
1025 ctx->cur_field = frame->interlaced_frame && !frame->top_field_first;
1028 static int dnxhd_encode_picture(AVCodecContext *avctx, AVPacket *pkt,
1029 const AVFrame *frame, int *got_packet)
1031 DNXHDEncContext *ctx = avctx->priv_data;
1032 int first_field = 1;
1036 if ((ret = ff_alloc_packet(pkt, ctx->cid_table->frame_size)) < 0) {
1037 av_log(avctx, AV_LOG_ERROR,
1038 "output buffer is too small to compress picture\n");
1043 dnxhd_load_picture(ctx, frame);
1046 for (i = 0; i < 3; i++) {
1047 ctx->src[i] = frame->data[i];
1048 if (ctx->interlaced && ctx->cur_field)
1049 ctx->src[i] += frame->linesize[i];
1052 dnxhd_write_header(avctx, buf);
1054 if (avctx->mb_decision == FF_MB_DECISION_RD)
1055 ret = dnxhd_encode_rdo(avctx, ctx);
1057 ret = dnxhd_encode_fast(avctx, ctx);
1059 av_log(avctx, AV_LOG_ERROR,
1060 "picture could not fit ratecontrol constraints, increase qmax\n");
1064 dnxhd_setup_threads_slices(ctx);
1067 for (i = 0; i < ctx->m.mb_height; i++) {
1068 AV_WB32(ctx->msip + i * 4, offset);
1069 offset += ctx->slice_size[i];
1070 assert(!(ctx->slice_size[i] & 3));
1073 avctx->execute2(avctx, dnxhd_encode_thread, buf, NULL, ctx->m.mb_height);
1075 assert(640 + offset + 4 <= ctx->cid_table->coding_unit_size);
1076 memset(buf + 640 + offset, 0,
1077 ctx->cid_table->coding_unit_size - 4 - offset - 640);
1079 AV_WB32(buf + ctx->cid_table->coding_unit_size - 4, 0x600DC0DE); // EOF
1081 if (ctx->interlaced && first_field) {
1083 ctx->cur_field ^= 1;
1084 buf += ctx->cid_table->coding_unit_size;
1085 goto encode_coding_unit;
1088 avctx->coded_frame->quality = ctx->qscale * FF_QP2LAMBDA;
1090 pkt->flags |= AV_PKT_FLAG_KEY;
1095 static av_cold int dnxhd_encode_end(AVCodecContext *avctx)
1097 DNXHDEncContext *ctx = avctx->priv_data;
1098 int max_level = 1 << (ctx->cid_table->bit_depth + 2);
1101 av_free(ctx->vlc_codes - max_level * 2);
1102 av_free(ctx->vlc_bits - max_level * 2);
1103 av_freep(&ctx->run_codes);
1104 av_freep(&ctx->run_bits);
1106 av_freep(&ctx->mb_bits);
1107 av_freep(&ctx->mb_qscale);
1108 av_freep(&ctx->mb_rc);
1109 av_freep(&ctx->mb_cmp);
1110 av_freep(&ctx->slice_size);
1111 av_freep(&ctx->slice_offs);
1113 av_freep(&ctx->qmatrix_c);
1114 av_freep(&ctx->qmatrix_l);
1115 av_freep(&ctx->qmatrix_c16);
1116 av_freep(&ctx->qmatrix_l16);
1118 for (i = 1; i < avctx->thread_count; i++)
1119 av_freep(&ctx->thread[i]);
1121 av_frame_free(&avctx->coded_frame);
1126 AVCodec ff_dnxhd_encoder = {
1128 .long_name = NULL_IF_CONFIG_SMALL("VC3/DNxHD"),
1129 .type = AVMEDIA_TYPE_VIDEO,
1130 .id = AV_CODEC_ID_DNXHD,
1131 .priv_data_size = sizeof(DNXHDEncContext),
1132 .init = dnxhd_encode_init,
1133 .encode2 = dnxhd_encode_picture,
1134 .close = dnxhd_encode_end,
1135 .capabilities = CODEC_CAP_SLICE_THREADS,
1136 .pix_fmts = (const enum AVPixelFormat[]) {
1138 AV_PIX_FMT_YUV422P10,
1141 .priv_class = &class,