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
26 #include "libavutil/attributes.h"
27 #include "libavutil/internal.h"
28 #include "libavutil/opt.h"
29 #include "libavutil/timer.h"
34 #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 dnxhd_class = {
51 .class_name = "dnxhd",
52 .item_name = av_default_item_name,
54 .version = LIBAVUTIL_VERSION_INT,
57 static void dnxhd_8bit_get_pixels_8x4_sym(int16_t *av_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 *av_restrict block,
82 const uint8_t *pixels,
86 const uint16_t* pixels16 = (const uint16_t*)pixels;
89 for (i = 0; i < 4; i++) {
90 block[0] = pixels16[0]; block[1] = pixels16[1];
91 block[2] = pixels16[2]; block[3] = pixels16[3];
92 block[4] = pixels16[4]; block[5] = pixels16[5];
93 block[6] = pixels16[6]; block[7] = pixels16[7];
94 pixels16 += line_size;
97 memcpy(block, block - 8, sizeof(*block) * 8);
98 memcpy(block + 8, block - 16, sizeof(*block) * 8);
99 memcpy(block + 16, block - 24, sizeof(*block) * 8);
100 memcpy(block + 24, block - 32, sizeof(*block) * 8);
103 static int dnxhd_10bit_dct_quantize(MpegEncContext *ctx, int16_t *block,
104 int n, int qscale, int *overflow)
106 const uint8_t *scantable= ctx->intra_scantable.scantable;
107 const int *qmat = n<4 ? ctx->q_intra_matrix[qscale] : ctx->q_chroma_intra_matrix[qscale];
108 int last_non_zero = 0;
111 ctx->dsp.fdct(block);
113 // Divide by 4 with rounding, to compensate scaling of DCT coefficients
114 block[0] = (block[0] + 2) >> 2;
116 for (i = 1; i < 64; ++i) {
117 int j = scantable[i];
118 int sign = block[j] >> 31;
119 int level = (block[j] ^ sign) - sign;
120 level = level * qmat[j] >> DNX10BIT_QMAT_SHIFT;
121 block[j] = (level ^ sign) - sign;
126 return last_non_zero;
129 static av_cold int dnxhd_init_vlc(DNXHDEncContext *ctx)
131 int i, j, level, run;
132 int max_level = 1 << (ctx->cid_table->bit_depth + 2);
134 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->vlc_codes,
135 max_level * 4 * sizeof(*ctx->vlc_codes), fail);
136 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->vlc_bits,
137 max_level * 4 * sizeof(*ctx->vlc_bits), fail);
138 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_codes,
140 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_bits,
143 ctx->vlc_codes += max_level * 2;
144 ctx->vlc_bits += max_level * 2;
145 for (level = -max_level; level < max_level; level++) {
146 for (run = 0; run < 2; run++) {
147 int index = (level << 1) | run;
148 int sign, offset = 0, alevel = level;
150 MASK_ABS(sign, alevel);
152 offset = (alevel - 1) >> 6;
153 alevel -= offset << 6;
155 for (j = 0; j < 257; j++) {
156 if (ctx->cid_table->ac_level[j] >> 1 == alevel &&
157 (!offset || (ctx->cid_table->ac_flags[j] & 1) && offset) &&
158 (!run || (ctx->cid_table->ac_flags[j] & 2) && run)) {
159 av_assert1(!ctx->vlc_codes[index]);
161 ctx->vlc_codes[index] =
162 (ctx->cid_table->ac_codes[j] << 1) | (sign & 1);
163 ctx->vlc_bits[index] = ctx->cid_table->ac_bits[j] + 1;
165 ctx->vlc_codes[index] = ctx->cid_table->ac_codes[j];
166 ctx->vlc_bits[index] = ctx->cid_table->ac_bits[j];
171 av_assert0(!alevel || j < 257);
173 ctx->vlc_codes[index] =
174 (ctx->vlc_codes[index] << ctx->cid_table->index_bits) | offset;
175 ctx->vlc_bits[index] += ctx->cid_table->index_bits;
179 for (i = 0; i < 62; i++) {
180 int run = ctx->cid_table->run[i];
181 av_assert0(run < 63);
182 ctx->run_codes[run] = ctx->cid_table->run_codes[i];
183 ctx->run_bits[run] = ctx->cid_table->run_bits[i];
187 return AVERROR(ENOMEM);
190 static av_cold int dnxhd_init_qmat(DNXHDEncContext *ctx, int lbias, int cbias)
192 // init first elem to 1 to avoid div by 0 in convert_matrix
193 uint16_t weight_matrix[64] = { 1, }; // convert_matrix needs uint16_t*
195 const uint8_t *luma_weight_table = ctx->cid_table->luma_weight;
196 const uint8_t *chroma_weight_table = ctx->cid_table->chroma_weight;
198 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l,
199 (ctx->m.avctx->qmax + 1) * 64 * sizeof(int), fail);
200 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c,
201 (ctx->m.avctx->qmax + 1) * 64 * sizeof(int), fail);
202 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l16,
203 (ctx->m.avctx->qmax + 1) * 64 * 2 * sizeof(uint16_t),
205 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c16,
206 (ctx->m.avctx->qmax + 1) * 64 * 2 * sizeof(uint16_t),
209 if (ctx->cid_table->bit_depth == 8) {
210 for (i = 1; i < 64; i++) {
211 int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]];
212 weight_matrix[j] = ctx->cid_table->luma_weight[i];
214 ff_convert_matrix(&ctx->m.dsp, ctx->qmatrix_l, ctx->qmatrix_l16,
215 weight_matrix, ctx->m.intra_quant_bias, 1,
216 ctx->m.avctx->qmax, 1);
217 for (i = 1; i < 64; i++) {
218 int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]];
219 weight_matrix[j] = ctx->cid_table->chroma_weight[i];
221 ff_convert_matrix(&ctx->m.dsp, ctx->qmatrix_c, ctx->qmatrix_c16,
222 weight_matrix, ctx->m.intra_quant_bias, 1,
223 ctx->m.avctx->qmax, 1);
225 for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
226 for (i = 0; i < 64; i++) {
227 ctx->qmatrix_l[qscale][i] <<= 2;
228 ctx->qmatrix_c[qscale][i] <<= 2;
229 ctx->qmatrix_l16[qscale][0][i] <<= 2;
230 ctx->qmatrix_l16[qscale][1][i] <<= 2;
231 ctx->qmatrix_c16[qscale][0][i] <<= 2;
232 ctx->qmatrix_c16[qscale][1][i] <<= 2;
237 for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
238 for (i = 1; i < 64; i++) {
239 int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]];
241 /* The quantization formula from the VC-3 standard is:
242 * quantized = sign(block[i]) * floor(abs(block[i]/s) * p /
243 * (qscale * weight_table[i]))
244 * Where p is 32 for 8-bit samples and 8 for 10-bit ones.
245 * The s factor compensates scaling of DCT coefficients done by
246 * the DCT routines, and therefore is not present in standard.
247 * It's 8 for 8-bit samples and 4 for 10-bit ones.
248 * We want values of ctx->qtmatrix_l and ctx->qtmatrix_r to be:
249 * ((1 << DNX10BIT_QMAT_SHIFT) * (p / s)) /
250 * (qscale * weight_table[i])
251 * For 10-bit samples, p / s == 2 */
252 ctx->qmatrix_l[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) /
253 (qscale * luma_weight_table[i]);
254 ctx->qmatrix_c[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) /
255 (qscale * chroma_weight_table[i]);
260 ctx->m.q_chroma_intra_matrix16 = ctx->qmatrix_c16;
261 ctx->m.q_chroma_intra_matrix = ctx->qmatrix_c;
262 ctx->m.q_intra_matrix16 = ctx->qmatrix_l16;
263 ctx->m.q_intra_matrix = ctx->qmatrix_l;
267 return AVERROR(ENOMEM);
270 static av_cold int dnxhd_init_rc(DNXHDEncContext *ctx)
272 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_rc, 8160 * (ctx->m.avctx->qmax + 1) * sizeof(RCEntry), fail);
273 if (ctx->m.avctx->mb_decision != FF_MB_DECISION_RD)
274 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_cmp,
275 ctx->m.mb_num * sizeof(RCCMPEntry), fail);
277 ctx->frame_bits = (ctx->cid_table->coding_unit_size -
278 640 - 4 - ctx->min_padding) * 8;
280 ctx->lambda = 2 << LAMBDA_FRAC_BITS; // qscale 2
283 return AVERROR(ENOMEM);
286 static av_cold int dnxhd_encode_init(AVCodecContext *avctx)
288 DNXHDEncContext *ctx = avctx->priv_data;
289 int i, index, bit_depth, ret;
291 switch (avctx->pix_fmt) {
292 case AV_PIX_FMT_YUV422P:
295 case AV_PIX_FMT_YUV422P10:
299 av_log(avctx, AV_LOG_ERROR,
300 "pixel format is incompatible with DNxHD\n");
301 return AVERROR(EINVAL);
304 ctx->cid = ff_dnxhd_find_cid(avctx, bit_depth);
306 av_log(avctx, AV_LOG_ERROR,
307 "video parameters incompatible with DNxHD. Valid DNxHD profiles:\n");
308 ff_dnxhd_print_profiles(avctx, AV_LOG_ERROR);
309 return AVERROR(EINVAL);
311 av_log(avctx, AV_LOG_DEBUG, "cid %d\n", ctx->cid);
313 index = ff_dnxhd_get_cid_table(ctx->cid);
314 av_assert0(index >= 0);
315 ctx->cid_table = &ff_dnxhd_cid_table[index];
317 ctx->m.avctx = avctx;
321 avctx->bits_per_raw_sample = ctx->cid_table->bit_depth;
323 ff_dct_common_init(&ctx->m);
324 ff_dct_encode_init(&ctx->m);
326 if (!ctx->m.dct_quantize)
327 ctx->m.dct_quantize = ff_dct_quantize_c;
329 if (ctx->cid_table->bit_depth == 10) {
330 ctx->m.dct_quantize = dnxhd_10bit_dct_quantize;
331 ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym;
332 ctx->block_width_l2 = 4;
334 ctx->get_pixels_8x4_sym = dnxhd_8bit_get_pixels_8x4_sym;
335 ctx->block_width_l2 = 3;
339 ff_dnxhdenc_init_x86(ctx);
341 ctx->m.mb_height = (avctx->height + 15) / 16;
342 ctx->m.mb_width = (avctx->width + 15) / 16;
344 if (avctx->flags & CODEC_FLAG_INTERLACED_DCT) {
346 ctx->m.mb_height /= 2;
349 ctx->m.mb_num = ctx->m.mb_height * ctx->m.mb_width;
351 if (avctx->intra_quant_bias != FF_DEFAULT_QUANT_BIAS)
352 ctx->m.intra_quant_bias = avctx->intra_quant_bias;
353 // XXX tune lbias/cbias
354 if ((ret = dnxhd_init_qmat(ctx, ctx->m.intra_quant_bias, 0)) < 0)
357 /* Avid Nitris hardware decoder requires a minimum amount of padding
358 * in the coding unit payload */
359 if (ctx->nitris_compat)
360 ctx->min_padding = 1600;
362 if ((ret = dnxhd_init_vlc(ctx)) < 0)
364 if ((ret = dnxhd_init_rc(ctx)) < 0)
367 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_size,
368 ctx->m.mb_height * sizeof(uint32_t), fail);
369 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_offs,
370 ctx->m.mb_height * sizeof(uint32_t), fail);
371 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_bits,
372 ctx->m.mb_num * sizeof(uint16_t), fail);
373 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_qscale,
374 ctx->m.mb_num * sizeof(uint8_t), fail);
376 avctx->coded_frame = av_frame_alloc();
377 if (!avctx->coded_frame)
378 return AVERROR(ENOMEM);
380 avctx->coded_frame->key_frame = 1;
381 avctx->coded_frame->pict_type = AV_PICTURE_TYPE_I;
383 if (avctx->thread_count > MAX_THREADS) {
384 av_log(avctx, AV_LOG_ERROR, "too many threads\n");
385 return AVERROR(EINVAL);
388 if (avctx->qmax <= 1) {
389 av_log(avctx, AV_LOG_ERROR, "qmax must be at least 2\n");
390 return AVERROR(EINVAL);
393 ctx->thread[0] = ctx;
394 for (i = 1; i < avctx->thread_count; i++) {
395 ctx->thread[i] = av_malloc(sizeof(DNXHDEncContext));
396 memcpy(ctx->thread[i], ctx, sizeof(DNXHDEncContext));
400 fail: // for FF_ALLOCZ_OR_GOTO
401 return AVERROR(ENOMEM);
404 static int dnxhd_write_header(AVCodecContext *avctx, uint8_t *buf)
406 DNXHDEncContext *ctx = avctx->priv_data;
407 static const uint8_t header_prefix[5] = { 0x00, 0x00, 0x02, 0x80, 0x01 };
411 memcpy(buf, header_prefix, 5);
412 buf[5] = ctx->interlaced ? ctx->cur_field + 2 : 0x01;
413 buf[6] = 0x80; // crc flag off
414 buf[7] = 0xa0; // reserved
415 AV_WB16(buf + 0x18, avctx->height >> ctx->interlaced); // ALPF
416 AV_WB16(buf + 0x1a, avctx->width); // SPL
417 AV_WB16(buf + 0x1d, avctx->height >> ctx->interlaced); // NAL
419 buf[0x21] = ctx->cid_table->bit_depth == 10 ? 0x58 : 0x38;
420 buf[0x22] = 0x88 + (ctx->interlaced << 2);
421 AV_WB32(buf + 0x28, ctx->cid); // CID
422 buf[0x2c] = ctx->interlaced ? 0 : 0x80;
424 buf[0x5f] = 0x01; // UDL
426 buf[0x167] = 0x02; // reserved
427 AV_WB16(buf + 0x16a, ctx->m.mb_height * 4 + 4); // MSIPS
428 buf[0x16d] = ctx->m.mb_height; // Ns
429 buf[0x16f] = 0x10; // reserved
431 ctx->msip = buf + 0x170;
435 static av_always_inline void dnxhd_encode_dc(DNXHDEncContext *ctx, int diff)
439 nbits = av_log2_16bit(-2 * diff);
442 nbits = av_log2_16bit(2 * diff);
444 put_bits(&ctx->m.pb, ctx->cid_table->dc_bits[nbits] + nbits,
445 (ctx->cid_table->dc_codes[nbits] << nbits) +
446 (diff & ((1 << nbits) - 1)));
449 static av_always_inline
450 void dnxhd_encode_block(DNXHDEncContext *ctx, int16_t *block,
451 int last_index, int n)
453 int last_non_zero = 0;
456 dnxhd_encode_dc(ctx, block[0] - ctx->m.last_dc[n]);
457 ctx->m.last_dc[n] = block[0];
459 for (i = 1; i <= last_index; i++) {
460 j = ctx->m.intra_scantable.permutated[i];
463 int run_level = i - last_non_zero - 1;
464 int rlevel = (slevel << 1) | !!run_level;
465 put_bits(&ctx->m.pb, ctx->vlc_bits[rlevel], ctx->vlc_codes[rlevel]);
467 put_bits(&ctx->m.pb, ctx->run_bits[run_level],
468 ctx->run_codes[run_level]);
472 put_bits(&ctx->m.pb, ctx->vlc_bits[0], ctx->vlc_codes[0]); // EOB
475 static av_always_inline
476 void dnxhd_unquantize_c(DNXHDEncContext *ctx, int16_t *block, int n,
477 int qscale, int last_index)
479 const uint8_t *weight_matrix;
483 weight_matrix = (n & 2) ? ctx->cid_table->chroma_weight
484 : ctx->cid_table->luma_weight;
486 for (i = 1; i <= last_index; i++) {
487 int j = ctx->m.intra_scantable.permutated[i];
491 level = (1 - 2 * level) * qscale * weight_matrix[i];
492 if (ctx->cid_table->bit_depth == 10) {
493 if (weight_matrix[i] != 8)
497 if (weight_matrix[i] != 32)
503 level = (2 * level + 1) * qscale * weight_matrix[i];
504 if (ctx->cid_table->bit_depth == 10) {
505 if (weight_matrix[i] != 8)
509 if (weight_matrix[i] != 32)
519 static av_always_inline int dnxhd_ssd_block(int16_t *qblock, int16_t *block)
523 for (i = 0; i < 64; i++)
524 score += (block[i] - qblock[i]) * (block[i] - qblock[i]);
528 static av_always_inline
529 int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, int16_t *block, int last_index)
531 int last_non_zero = 0;
534 for (i = 1; i <= last_index; i++) {
535 j = ctx->m.intra_scantable.permutated[i];
538 int run_level = i - last_non_zero - 1;
539 bits += ctx->vlc_bits[(level << 1) |
540 !!run_level] + ctx->run_bits[run_level];
547 static av_always_inline
548 void dnxhd_get_blocks(DNXHDEncContext *ctx, int mb_x, int mb_y)
550 const int bs = ctx->block_width_l2;
551 const int bw = 1 << bs;
552 const uint8_t *ptr_y = ctx->thread[0]->src[0] +
553 ((mb_y << 4) * ctx->m.linesize) + (mb_x << bs + 1);
554 const uint8_t *ptr_u = ctx->thread[0]->src[1] +
555 ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs);
556 const uint8_t *ptr_v = ctx->thread[0]->src[2] +
557 ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs);
558 DSPContext *dsp = &ctx->m.dsp;
560 dsp->get_pixels(ctx->blocks[0], ptr_y, ctx->m.linesize);
561 dsp->get_pixels(ctx->blocks[1], ptr_y + bw, ctx->m.linesize);
562 dsp->get_pixels(ctx->blocks[2], ptr_u, ctx->m.uvlinesize);
563 dsp->get_pixels(ctx->blocks[3], ptr_v, ctx->m.uvlinesize);
565 if (mb_y + 1 == ctx->m.mb_height && ctx->m.avctx->height == 1080) {
566 if (ctx->interlaced) {
567 ctx->get_pixels_8x4_sym(ctx->blocks[4],
568 ptr_y + ctx->dct_y_offset,
570 ctx->get_pixels_8x4_sym(ctx->blocks[5],
571 ptr_y + ctx->dct_y_offset + bw,
573 ctx->get_pixels_8x4_sym(ctx->blocks[6],
574 ptr_u + ctx->dct_uv_offset,
576 ctx->get_pixels_8x4_sym(ctx->blocks[7],
577 ptr_v + ctx->dct_uv_offset,
580 dsp->clear_block(ctx->blocks[4]);
581 dsp->clear_block(ctx->blocks[5]);
582 dsp->clear_block(ctx->blocks[6]);
583 dsp->clear_block(ctx->blocks[7]);
586 dsp->get_pixels(ctx->blocks[4],
587 ptr_y + ctx->dct_y_offset, ctx->m.linesize);
588 dsp->get_pixels(ctx->blocks[5],
589 ptr_y + ctx->dct_y_offset + bw, ctx->m.linesize);
590 dsp->get_pixels(ctx->blocks[6],
591 ptr_u + ctx->dct_uv_offset, ctx->m.uvlinesize);
592 dsp->get_pixels(ctx->blocks[7],
593 ptr_v + ctx->dct_uv_offset, ctx->m.uvlinesize);
597 static av_always_inline
598 int dnxhd_switch_matrix(DNXHDEncContext *ctx, int i)
600 const static uint8_t component[8]={0,0,1,2,0,0,1,2};
604 static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg,
605 int jobnr, int threadnr)
607 DNXHDEncContext *ctx = avctx->priv_data;
608 int mb_y = jobnr, mb_x;
609 int qscale = ctx->qscale;
610 LOCAL_ALIGNED_16(int16_t, block, [64]);
611 ctx = ctx->thread[threadnr];
615 ctx->m.last_dc[2] = 1 << (ctx->cid_table->bit_depth + 2);
617 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
618 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
624 dnxhd_get_blocks(ctx, mb_x, mb_y);
626 for (i = 0; i < 8; i++) {
627 int16_t *src_block = ctx->blocks[i];
628 int overflow, nbits, diff, last_index;
629 int n = dnxhd_switch_matrix(ctx, i);
631 memcpy(block, src_block, 64 * sizeof(*block));
632 last_index = ctx->m.dct_quantize(&ctx->m, block, 4 & (2*i),
634 ac_bits += dnxhd_calc_ac_bits(ctx, block, last_index);
636 diff = block[0] - ctx->m.last_dc[n];
638 nbits = av_log2_16bit(-2 * diff);
640 nbits = av_log2_16bit(2 * diff);
642 av_assert1(nbits < ctx->cid_table->bit_depth + 4);
643 dc_bits += ctx->cid_table->dc_bits[nbits] + nbits;
645 ctx->m.last_dc[n] = block[0];
647 if (avctx->mb_decision == FF_MB_DECISION_RD || !RC_VARIANCE) {
648 dnxhd_unquantize_c(ctx, block, i, qscale, last_index);
649 ctx->m.dsp.idct(block);
650 ssd += dnxhd_ssd_block(block, src_block);
653 ctx->mb_rc[qscale][mb].ssd = ssd;
654 ctx->mb_rc[qscale][mb].bits = ac_bits + dc_bits + 12 +
655 8 * ctx->vlc_bits[0];
660 static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg,
661 int jobnr, int threadnr)
663 DNXHDEncContext *ctx = avctx->priv_data;
664 int mb_y = jobnr, mb_x;
665 ctx = ctx->thread[threadnr];
666 init_put_bits(&ctx->m.pb, (uint8_t *)arg + 640 + ctx->slice_offs[jobnr],
667 ctx->slice_size[jobnr]);
671 ctx->m.last_dc[2] = 1 << (ctx->cid_table->bit_depth + 2);
672 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
673 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
674 int qscale = ctx->mb_qscale[mb];
677 put_bits(&ctx->m.pb, 12, qscale << 1);
679 dnxhd_get_blocks(ctx, mb_x, mb_y);
681 for (i = 0; i < 8; i++) {
682 int16_t *block = ctx->blocks[i];
683 int overflow, n = dnxhd_switch_matrix(ctx, i);
684 int last_index = ctx->m.dct_quantize(&ctx->m, block, 4 & (2*i),
687 dnxhd_encode_block(ctx, block, last_index, n);
688 // STOP_TIMER("encode_block");
691 if (put_bits_count(&ctx->m.pb) & 31)
692 put_bits(&ctx->m.pb, 32 - (put_bits_count(&ctx->m.pb) & 31), 0);
693 flush_put_bits(&ctx->m.pb);
697 static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx)
701 for (mb_y = 0; mb_y < ctx->m.mb_height; mb_y++) {
703 ctx->slice_offs[mb_y] = offset;
704 ctx->slice_size[mb_y] = 0;
705 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
706 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
707 ctx->slice_size[mb_y] += ctx->mb_bits[mb];
709 ctx->slice_size[mb_y] = (ctx->slice_size[mb_y] + 31) & ~31;
710 ctx->slice_size[mb_y] >>= 3;
711 thread_size = ctx->slice_size[mb_y];
712 offset += thread_size;
716 static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg,
717 int jobnr, int threadnr)
719 DNXHDEncContext *ctx = avctx->priv_data;
720 int mb_y = jobnr, mb_x, x, y;
721 int partial_last_row = (mb_y == ctx->m.mb_height - 1) &&
722 ((avctx->height >> ctx->interlaced) & 0xF);
724 ctx = ctx->thread[threadnr];
725 if (ctx->cid_table->bit_depth == 8) {
726 uint8_t *pix = ctx->thread[0]->src[0] + ((mb_y << 4) * ctx->m.linesize);
727 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x, pix += 16) {
728 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
732 if (!partial_last_row && mb_x * 16 <= avctx->width - 16) {
733 sum = ctx->m.dsp.pix_sum(pix, ctx->m.linesize);
734 varc = ctx->m.dsp.pix_norm1(pix, ctx->m.linesize);
736 int bw = FFMIN(avctx->width - 16 * mb_x, 16);
737 int bh = FFMIN((avctx->height >> ctx->interlaced) - 16 * mb_y, 16);
739 for (y = 0; y < bh; y++) {
740 for (x = 0; x < bw; x++) {
741 uint8_t val = pix[x + y * ctx->m.linesize];
747 varc = (varc - (((unsigned) sum * sum) >> 8) + 128) >> 8;
749 ctx->mb_cmp[mb].value = varc;
750 ctx->mb_cmp[mb].mb = mb;
753 int const linesize = ctx->m.linesize >> 1;
754 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x) {
755 uint16_t *pix = (uint16_t *)ctx->thread[0]->src[0] +
756 ((mb_y << 4) * linesize) + (mb_x << 4);
757 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
762 // Macroblocks are 16x16 pixels, unlike DCT blocks which are 8x8.
763 for (i = 0; i < 16; ++i) {
764 for (j = 0; j < 16; ++j) {
765 // Turn 16-bit pixels into 10-bit ones.
766 int const sample = (unsigned) pix[j] >> 6;
768 sqsum += sample * sample;
769 // 2^10 * 2^10 * 16 * 16 = 2^28, which is less than INT_MAX
773 mean = sum >> 8; // 16*16 == 2^8
775 ctx->mb_cmp[mb].value = sqmean - mean * mean;
776 ctx->mb_cmp[mb].mb = mb;
782 static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx)
784 int lambda, up_step, down_step;
785 int last_lower = INT_MAX, last_higher = 0;
788 for (q = 1; q < avctx->qmax; q++) {
790 avctx->execute2(avctx, dnxhd_calc_bits_thread,
791 NULL, NULL, ctx->m.mb_height);
793 up_step = down_step = 2 << LAMBDA_FRAC_BITS;
794 lambda = ctx->lambda;
799 if (lambda == last_higher) {
801 end = 1; // need to set final qscales/bits
803 for (y = 0; y < ctx->m.mb_height; y++) {
804 for (x = 0; x < ctx->m.mb_width; x++) {
805 unsigned min = UINT_MAX;
807 int mb = y * ctx->m.mb_width + x;
808 for (q = 1; q < avctx->qmax; q++) {
809 unsigned score = ctx->mb_rc[q][mb].bits * lambda +
810 ((unsigned) ctx->mb_rc[q][mb].ssd << LAMBDA_FRAC_BITS);
816 bits += ctx->mb_rc[qscale][mb].bits;
817 ctx->mb_qscale[mb] = qscale;
818 ctx->mb_bits[mb] = ctx->mb_rc[qscale][mb].bits;
820 bits = (bits + 31) & ~31; // padding
821 if (bits > ctx->frame_bits)
824 // av_dlog(ctx->m.avctx,
825 // "lambda %d, up %u, down %u, bits %d, frame %d\n",
826 // lambda, last_higher, last_lower, bits, ctx->frame_bits);
828 if (bits > ctx->frame_bits)
829 return AVERROR(EINVAL);
832 if (bits < ctx->frame_bits) {
833 last_lower = FFMIN(lambda, last_lower);
834 if (last_higher != 0)
835 lambda = (lambda+last_higher)>>1;
838 down_step = FFMIN((int64_t)down_step*5, INT_MAX);
839 up_step = 1<<LAMBDA_FRAC_BITS;
840 lambda = FFMAX(1, lambda);
841 if (lambda == last_lower)
844 last_higher = FFMAX(lambda, last_higher);
845 if (last_lower != INT_MAX)
846 lambda = (lambda+last_lower)>>1;
847 else if ((int64_t)lambda + up_step > INT_MAX)
848 return AVERROR(EINVAL);
851 up_step = FFMIN((int64_t)up_step*5, INT_MAX);
852 down_step = 1<<LAMBDA_FRAC_BITS;
855 //av_dlog(ctx->m.avctx, "out lambda %d\n", lambda);
856 ctx->lambda = lambda;
860 static int dnxhd_find_qscale(DNXHDEncContext *ctx)
866 int last_lower = INT_MAX;
870 qscale = ctx->qscale;
873 ctx->qscale = qscale;
874 // XXX avoid recalculating bits
875 ctx->m.avctx->execute2(ctx->m.avctx, dnxhd_calc_bits_thread,
876 NULL, NULL, ctx->m.mb_height);
877 for (y = 0; y < ctx->m.mb_height; y++) {
878 for (x = 0; x < ctx->m.mb_width; x++)
879 bits += ctx->mb_rc[qscale][y*ctx->m.mb_width+x].bits;
880 bits = (bits+31)&~31; // padding
881 if (bits > ctx->frame_bits)
884 // av_dlog(ctx->m.avctx,
885 // "%d, qscale %d, bits %d, frame %d, higher %d, lower %d\n",
886 // ctx->m.avctx->frame_number, qscale, bits, ctx->frame_bits,
887 // last_higher, last_lower);
888 if (bits < ctx->frame_bits) {
891 if (last_higher == qscale - 1) {
892 qscale = last_higher;
895 last_lower = FFMIN(qscale, last_lower);
896 if (last_higher != 0)
897 qscale = (qscale + last_higher) >> 1;
899 qscale -= down_step++;
904 if (last_lower == qscale + 1)
906 last_higher = FFMAX(qscale, last_higher);
907 if (last_lower != INT_MAX)
908 qscale = (qscale + last_lower) >> 1;
912 if (qscale >= ctx->m.avctx->qmax)
913 return AVERROR(EINVAL);
916 //av_dlog(ctx->m.avctx, "out qscale %d\n", qscale);
917 ctx->qscale = qscale;
921 #define BUCKET_BITS 8
922 #define RADIX_PASSES 4
923 #define NBUCKETS (1 << BUCKET_BITS)
925 static inline int get_bucket(int value, int shift)
928 value &= NBUCKETS - 1;
929 return NBUCKETS - 1 - value;
932 static void radix_count(const RCCMPEntry *data, int size,
933 int buckets[RADIX_PASSES][NBUCKETS])
936 memset(buckets, 0, sizeof(buckets[0][0]) * RADIX_PASSES * NBUCKETS);
937 for (i = 0; i < size; i++) {
938 int v = data[i].value;
939 for (j = 0; j < RADIX_PASSES; j++) {
940 buckets[j][get_bucket(v, 0)]++;
945 for (j = 0; j < RADIX_PASSES; j++) {
947 for (i = NBUCKETS - 1; i >= 0; i--)
948 buckets[j][i] = offset -= buckets[j][i];
949 av_assert1(!buckets[j][0]);
953 static void radix_sort_pass(RCCMPEntry *dst, const RCCMPEntry *data,
954 int size, int buckets[NBUCKETS], int pass)
956 int shift = pass * BUCKET_BITS;
958 for (i = 0; i < size; i++) {
959 int v = get_bucket(data[i].value, shift);
960 int pos = buckets[v]++;
965 static void radix_sort(RCCMPEntry *data, int size)
967 int buckets[RADIX_PASSES][NBUCKETS];
968 RCCMPEntry *tmp = av_malloc_array(size, sizeof(*tmp));
969 radix_count(data, size, buckets);
970 radix_sort_pass(tmp, data, size, buckets[0], 0);
971 radix_sort_pass(data, tmp, size, buckets[1], 1);
972 if (buckets[2][NBUCKETS - 1] || buckets[3][NBUCKETS - 1]) {
973 radix_sort_pass(tmp, data, size, buckets[2], 2);
974 radix_sort_pass(data, tmp, size, buckets[3], 3);
979 static int dnxhd_encode_fast(AVCodecContext *avctx, DNXHDEncContext *ctx)
983 if ((ret = dnxhd_find_qscale(ctx)) < 0)
985 for (y = 0; y < ctx->m.mb_height; y++) {
986 for (x = 0; x < ctx->m.mb_width; x++) {
987 int mb = y * ctx->m.mb_width + x;
989 ctx->mb_qscale[mb] = ctx->qscale;
990 ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale][mb].bits;
991 max_bits += ctx->mb_rc[ctx->qscale][mb].bits;
993 delta_bits = ctx->mb_rc[ctx->qscale][mb].bits -
994 ctx->mb_rc[ctx->qscale + 1][mb].bits;
995 ctx->mb_cmp[mb].mb = mb;
996 ctx->mb_cmp[mb].value =
997 delta_bits ? ((ctx->mb_rc[ctx->qscale][mb].ssd -
998 ctx->mb_rc[ctx->qscale + 1][mb].ssd) * 100) /
1000 : INT_MIN; // avoid increasing qscale
1003 max_bits += 31; // worst padding
1007 avctx->execute2(avctx, dnxhd_mb_var_thread,
1008 NULL, NULL, ctx->m.mb_height);
1009 radix_sort(ctx->mb_cmp, ctx->m.mb_num);
1010 for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) {
1011 int mb = ctx->mb_cmp[x].mb;
1012 max_bits -= ctx->mb_rc[ctx->qscale][mb].bits -
1013 ctx->mb_rc[ctx->qscale + 1][mb].bits;
1014 ctx->mb_qscale[mb] = ctx->qscale + 1;
1015 ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale + 1][mb].bits;
1021 static void dnxhd_load_picture(DNXHDEncContext *ctx, const AVFrame *frame)
1025 for (i = 0; i < ctx->m.avctx->thread_count; i++) {
1026 ctx->thread[i]->m.linesize = frame->linesize[0] << ctx->interlaced;
1027 ctx->thread[i]->m.uvlinesize = frame->linesize[1] << ctx->interlaced;
1028 ctx->thread[i]->dct_y_offset = ctx->m.linesize *8;
1029 ctx->thread[i]->dct_uv_offset = ctx->m.uvlinesize*8;
1032 ctx->m.avctx->coded_frame->interlaced_frame = frame->interlaced_frame;
1033 ctx->cur_field = frame->interlaced_frame && !frame->top_field_first;
1036 static int dnxhd_encode_picture(AVCodecContext *avctx, AVPacket *pkt,
1037 const AVFrame *frame, int *got_packet)
1039 DNXHDEncContext *ctx = avctx->priv_data;
1040 int first_field = 1;
1044 if ((ret = ff_alloc_packet2(avctx, pkt, ctx->cid_table->frame_size)) < 0)
1048 dnxhd_load_picture(ctx, frame);
1051 for (i = 0; i < 3; i++) {
1052 ctx->src[i] = frame->data[i];
1053 if (ctx->interlaced && ctx->cur_field)
1054 ctx->src[i] += frame->linesize[i];
1057 dnxhd_write_header(avctx, buf);
1059 if (avctx->mb_decision == FF_MB_DECISION_RD)
1060 ret = dnxhd_encode_rdo(avctx, ctx);
1062 ret = dnxhd_encode_fast(avctx, ctx);
1064 av_log(avctx, AV_LOG_ERROR,
1065 "picture could not fit ratecontrol constraints, increase qmax\n");
1069 dnxhd_setup_threads_slices(ctx);
1072 for (i = 0; i < ctx->m.mb_height; i++) {
1073 AV_WB32(ctx->msip + i * 4, offset);
1074 offset += ctx->slice_size[i];
1075 av_assert1(!(ctx->slice_size[i] & 3));
1078 avctx->execute2(avctx, dnxhd_encode_thread, buf, NULL, ctx->m.mb_height);
1080 av_assert1(640 + offset + 4 <= ctx->cid_table->coding_unit_size);
1081 memset(buf + 640 + offset, 0,
1082 ctx->cid_table->coding_unit_size - 4 - offset - 640);
1084 AV_WB32(buf + ctx->cid_table->coding_unit_size - 4, 0x600DC0DE); // EOF
1086 if (ctx->interlaced && first_field) {
1088 ctx->cur_field ^= 1;
1089 buf += ctx->cid_table->coding_unit_size;
1090 goto encode_coding_unit;
1093 avctx->coded_frame->quality = ctx->qscale * FF_QP2LAMBDA;
1095 pkt->flags |= AV_PKT_FLAG_KEY;
1100 static av_cold int dnxhd_encode_end(AVCodecContext *avctx)
1102 DNXHDEncContext *ctx = avctx->priv_data;
1103 int max_level = 1 << (ctx->cid_table->bit_depth + 2);
1106 av_free(ctx->vlc_codes - max_level * 2);
1107 av_free(ctx->vlc_bits - max_level * 2);
1108 av_freep(&ctx->run_codes);
1109 av_freep(&ctx->run_bits);
1111 av_freep(&ctx->mb_bits);
1112 av_freep(&ctx->mb_qscale);
1113 av_freep(&ctx->mb_rc);
1114 av_freep(&ctx->mb_cmp);
1115 av_freep(&ctx->slice_size);
1116 av_freep(&ctx->slice_offs);
1118 av_freep(&ctx->qmatrix_c);
1119 av_freep(&ctx->qmatrix_l);
1120 av_freep(&ctx->qmatrix_c16);
1121 av_freep(&ctx->qmatrix_l16);
1123 for (i = 1; i < avctx->thread_count; i++)
1124 av_freep(&ctx->thread[i]);
1126 av_frame_free(&avctx->coded_frame);
1131 static const AVCodecDefault dnxhd_defaults[] = {
1132 { "qmax", "1024" }, /* Maximum quantization scale factor allowed for VC-3 */
1136 AVCodec ff_dnxhd_encoder = {
1138 .long_name = NULL_IF_CONFIG_SMALL("VC3/DNxHD"),
1139 .type = AVMEDIA_TYPE_VIDEO,
1140 .id = AV_CODEC_ID_DNXHD,
1141 .priv_data_size = sizeof(DNXHDEncContext),
1142 .init = dnxhd_encode_init,
1143 .encode2 = dnxhd_encode_picture,
1144 .close = dnxhd_encode_end,
1145 .capabilities = CODEC_CAP_SLICE_THREADS,
1146 .pix_fmts = (const enum AVPixelFormat[]) {
1148 AV_PIX_FMT_YUV422P10,
1151 .priv_class = &dnxhd_class,
1152 .defaults = dnxhd_defaults,