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"
35 #include "mpegvideo.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 dnxhd_class = {
52 .class_name = "dnxhd",
53 .item_name = av_default_item_name,
55 .version = LIBAVUTIL_VERSION_INT,
58 static void dnxhd_8bit_get_pixels_8x4_sym(int16_t *av_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 *av_restrict block,
83 const uint8_t *pixels,
87 const uint16_t* pixels16 = (const uint16_t*)pixels;
90 for (i = 0; i < 4; i++) {
91 block[0] = pixels16[0]; block[1] = pixels16[1];
92 block[2] = pixels16[2]; block[3] = pixels16[3];
93 block[4] = pixels16[4]; block[5] = pixels16[5];
94 block[6] = pixels16[6]; block[7] = pixels16[7];
95 pixels16 += line_size;
98 memcpy(block, block - 8, sizeof(*block) * 8);
99 memcpy(block + 8, block - 16, sizeof(*block) * 8);
100 memcpy(block + 16, block - 24, sizeof(*block) * 8);
101 memcpy(block + 24, block - 32, sizeof(*block) * 8);
104 static int dnxhd_10bit_dct_quantize(MpegEncContext *ctx, int16_t *block,
105 int n, int qscale, int *overflow)
107 const uint8_t *scantable= ctx->intra_scantable.scantable;
108 const int *qmat = n<4 ? ctx->q_intra_matrix[qscale] : ctx->q_chroma_intra_matrix[qscale];
109 int last_non_zero = 0;
112 ctx->dsp.fdct(block);
114 // Divide by 4 with rounding, to compensate scaling of DCT coefficients
115 block[0] = (block[0] + 2) >> 2;
117 for (i = 1; i < 64; ++i) {
118 int j = scantable[i];
119 int sign = block[j] >> 31;
120 int level = (block[j] ^ sign) - sign;
121 level = level * qmat[j] >> DNX10BIT_QMAT_SHIFT;
122 block[j] = (level ^ sign) - sign;
127 return last_non_zero;
130 static av_cold int dnxhd_init_vlc(DNXHDEncContext *ctx)
132 int i, j, level, run;
133 int max_level = 1 << (ctx->cid_table->bit_depth + 2);
135 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->vlc_codes,
136 max_level * 4 * sizeof(*ctx->vlc_codes), fail);
137 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->vlc_bits,
138 max_level * 4 * sizeof(*ctx->vlc_bits), fail);
139 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_codes,
141 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_bits,
144 ctx->vlc_codes += max_level * 2;
145 ctx->vlc_bits += max_level * 2;
146 for (level = -max_level; level < max_level; level++) {
147 for (run = 0; run < 2; run++) {
148 int index = (level << 1) | run;
149 int sign, offset = 0, alevel = level;
151 MASK_ABS(sign, alevel);
153 offset = (alevel - 1) >> 6;
154 alevel -= offset << 6;
156 for (j = 0; j < 257; j++) {
157 if (ctx->cid_table->ac_level[j] >> 1 == alevel &&
158 (!offset || (ctx->cid_table->ac_flags[j] & 1) && offset) &&
159 (!run || (ctx->cid_table->ac_flags[j] & 2) && run)) {
160 av_assert1(!ctx->vlc_codes[index]);
162 ctx->vlc_codes[index] =
163 (ctx->cid_table->ac_codes[j] << 1) | (sign & 1);
164 ctx->vlc_bits[index] = ctx->cid_table->ac_bits[j] + 1;
166 ctx->vlc_codes[index] = ctx->cid_table->ac_codes[j];
167 ctx->vlc_bits[index] = ctx->cid_table->ac_bits[j];
172 av_assert0(!alevel || j < 257);
174 ctx->vlc_codes[index] =
175 (ctx->vlc_codes[index] << ctx->cid_table->index_bits) | offset;
176 ctx->vlc_bits[index] += ctx->cid_table->index_bits;
180 for (i = 0; i < 62; i++) {
181 int run = ctx->cid_table->run[i];
182 av_assert0(run < 63);
183 ctx->run_codes[run] = ctx->cid_table->run_codes[i];
184 ctx->run_bits[run] = ctx->cid_table->run_bits[i];
188 return AVERROR(ENOMEM);
191 static av_cold int dnxhd_init_qmat(DNXHDEncContext *ctx, int lbias, int cbias)
193 // init first elem to 1 to avoid div by 0 in convert_matrix
194 uint16_t weight_matrix[64] = { 1, }; // convert_matrix needs uint16_t*
196 const uint8_t *luma_weight_table = ctx->cid_table->luma_weight;
197 const uint8_t *chroma_weight_table = ctx->cid_table->chroma_weight;
199 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l,
200 (ctx->m.avctx->qmax + 1) * 64 * sizeof(int), fail);
201 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c,
202 (ctx->m.avctx->qmax + 1) * 64 * sizeof(int), fail);
203 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l16,
204 (ctx->m.avctx->qmax + 1) * 64 * 2 * sizeof(uint16_t),
206 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c16,
207 (ctx->m.avctx->qmax + 1) * 64 * 2 * sizeof(uint16_t),
210 if (ctx->cid_table->bit_depth == 8) {
211 for (i = 1; i < 64; i++) {
212 int j = ctx->m.idsp.idct_permutation[ff_zigzag_direct[i]];
213 weight_matrix[j] = ctx->cid_table->luma_weight[i];
215 ff_convert_matrix(&ctx->m, ctx->qmatrix_l, ctx->qmatrix_l16,
216 weight_matrix, ctx->m.intra_quant_bias, 1,
217 ctx->m.avctx->qmax, 1);
218 for (i = 1; i < 64; i++) {
219 int j = ctx->m.idsp.idct_permutation[ff_zigzag_direct[i]];
220 weight_matrix[j] = ctx->cid_table->chroma_weight[i];
222 ff_convert_matrix(&ctx->m, ctx->qmatrix_c, ctx->qmatrix_c16,
223 weight_matrix, ctx->m.intra_quant_bias, 1,
224 ctx->m.avctx->qmax, 1);
226 for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
227 for (i = 0; i < 64; i++) {
228 ctx->qmatrix_l[qscale][i] <<= 2;
229 ctx->qmatrix_c[qscale][i] <<= 2;
230 ctx->qmatrix_l16[qscale][0][i] <<= 2;
231 ctx->qmatrix_l16[qscale][1][i] <<= 2;
232 ctx->qmatrix_c16[qscale][0][i] <<= 2;
233 ctx->qmatrix_c16[qscale][1][i] <<= 2;
238 for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
239 for (i = 1; i < 64; i++) {
240 int j = ctx->m.idsp.idct_permutation[ff_zigzag_direct[i]];
242 /* The quantization formula from the VC-3 standard is:
243 * quantized = sign(block[i]) * floor(abs(block[i]/s) * p /
244 * (qscale * weight_table[i]))
245 * Where p is 32 for 8-bit samples and 8 for 10-bit ones.
246 * The s factor compensates scaling of DCT coefficients done by
247 * the DCT routines, and therefore is not present in standard.
248 * It's 8 for 8-bit samples and 4 for 10-bit ones.
249 * We want values of ctx->qtmatrix_l and ctx->qtmatrix_r to be:
250 * ((1 << DNX10BIT_QMAT_SHIFT) * (p / s)) /
251 * (qscale * weight_table[i])
252 * For 10-bit samples, p / s == 2 */
253 ctx->qmatrix_l[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) /
254 (qscale * luma_weight_table[i]);
255 ctx->qmatrix_c[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) /
256 (qscale * chroma_weight_table[i]);
261 ctx->m.q_chroma_intra_matrix16 = ctx->qmatrix_c16;
262 ctx->m.q_chroma_intra_matrix = ctx->qmatrix_c;
263 ctx->m.q_intra_matrix16 = ctx->qmatrix_l16;
264 ctx->m.q_intra_matrix = ctx->qmatrix_l;
268 return AVERROR(ENOMEM);
271 static av_cold int dnxhd_init_rc(DNXHDEncContext *ctx)
273 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_rc, 8160 * (ctx->m.avctx->qmax + 1) * sizeof(RCEntry), fail);
274 if (ctx->m.avctx->mb_decision != FF_MB_DECISION_RD)
275 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_cmp,
276 ctx->m.mb_num * sizeof(RCCMPEntry), fail);
278 ctx->frame_bits = (ctx->cid_table->coding_unit_size -
279 640 - 4 - ctx->min_padding) * 8;
281 ctx->lambda = 2 << LAMBDA_FRAC_BITS; // qscale 2
284 return AVERROR(ENOMEM);
287 static av_cold int dnxhd_encode_init(AVCodecContext *avctx)
289 DNXHDEncContext *ctx = avctx->priv_data;
290 int i, index, bit_depth, ret;
292 switch (avctx->pix_fmt) {
293 case AV_PIX_FMT_YUV422P:
296 case AV_PIX_FMT_YUV422P10:
300 av_log(avctx, AV_LOG_ERROR,
301 "pixel format is incompatible with DNxHD\n");
302 return AVERROR(EINVAL);
305 ctx->cid = ff_dnxhd_find_cid(avctx, bit_depth);
307 av_log(avctx, AV_LOG_ERROR,
308 "video parameters incompatible with DNxHD. Valid DNxHD profiles:\n");
309 ff_dnxhd_print_profiles(avctx, AV_LOG_ERROR);
310 return AVERROR(EINVAL);
312 av_log(avctx, AV_LOG_DEBUG, "cid %d\n", ctx->cid);
314 index = ff_dnxhd_get_cid_table(ctx->cid);
315 av_assert0(index >= 0);
316 ctx->cid_table = &ff_dnxhd_cid_table[index];
318 ctx->m.avctx = avctx;
322 avctx->bits_per_raw_sample = ctx->cid_table->bit_depth;
324 ff_blockdsp_init(&ctx->bdsp, avctx);
325 ff_idctdsp_init(&ctx->m.idsp, avctx);
326 ff_dct_common_init(&ctx->m);
327 ff_dct_encode_init(&ctx->m);
329 if (!ctx->m.dct_quantize)
330 ctx->m.dct_quantize = ff_dct_quantize_c;
332 if (ctx->cid_table->bit_depth == 10) {
333 ctx->m.dct_quantize = dnxhd_10bit_dct_quantize;
334 ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym;
335 ctx->block_width_l2 = 4;
337 ctx->get_pixels_8x4_sym = dnxhd_8bit_get_pixels_8x4_sym;
338 ctx->block_width_l2 = 3;
342 ff_dnxhdenc_init_x86(ctx);
344 ctx->m.mb_height = (avctx->height + 15) / 16;
345 ctx->m.mb_width = (avctx->width + 15) / 16;
347 if (avctx->flags & CODEC_FLAG_INTERLACED_DCT) {
349 ctx->m.mb_height /= 2;
352 ctx->m.mb_num = ctx->m.mb_height * ctx->m.mb_width;
354 if (avctx->intra_quant_bias != FF_DEFAULT_QUANT_BIAS)
355 ctx->m.intra_quant_bias = avctx->intra_quant_bias;
356 // XXX tune lbias/cbias
357 if ((ret = dnxhd_init_qmat(ctx, ctx->m.intra_quant_bias, 0)) < 0)
360 /* Avid Nitris hardware decoder requires a minimum amount of padding
361 * in the coding unit payload */
362 if (ctx->nitris_compat)
363 ctx->min_padding = 1600;
365 if ((ret = dnxhd_init_vlc(ctx)) < 0)
367 if ((ret = dnxhd_init_rc(ctx)) < 0)
370 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_size,
371 ctx->m.mb_height * sizeof(uint32_t), fail);
372 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_offs,
373 ctx->m.mb_height * sizeof(uint32_t), fail);
374 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_bits,
375 ctx->m.mb_num * sizeof(uint16_t), fail);
376 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_qscale,
377 ctx->m.mb_num * sizeof(uint8_t), fail);
379 avctx->coded_frame = av_frame_alloc();
380 if (!avctx->coded_frame)
381 return AVERROR(ENOMEM);
383 avctx->coded_frame->key_frame = 1;
384 avctx->coded_frame->pict_type = AV_PICTURE_TYPE_I;
386 if (avctx->thread_count > MAX_THREADS) {
387 av_log(avctx, AV_LOG_ERROR, "too many threads\n");
388 return AVERROR(EINVAL);
391 if (avctx->qmax <= 1) {
392 av_log(avctx, AV_LOG_ERROR, "qmax must be at least 2\n");
393 return AVERROR(EINVAL);
396 ctx->thread[0] = ctx;
397 for (i = 1; i < avctx->thread_count; i++) {
398 ctx->thread[i] = av_malloc(sizeof(DNXHDEncContext));
399 memcpy(ctx->thread[i], ctx, sizeof(DNXHDEncContext));
403 fail: // for FF_ALLOCZ_OR_GOTO
404 return AVERROR(ENOMEM);
407 static int dnxhd_write_header(AVCodecContext *avctx, uint8_t *buf)
409 DNXHDEncContext *ctx = avctx->priv_data;
410 static const uint8_t header_prefix[5] = { 0x00, 0x00, 0x02, 0x80, 0x01 };
414 memcpy(buf, header_prefix, 5);
415 buf[5] = ctx->interlaced ? ctx->cur_field + 2 : 0x01;
416 buf[6] = 0x80; // crc flag off
417 buf[7] = 0xa0; // reserved
418 AV_WB16(buf + 0x18, avctx->height >> ctx->interlaced); // ALPF
419 AV_WB16(buf + 0x1a, avctx->width); // SPL
420 AV_WB16(buf + 0x1d, avctx->height >> ctx->interlaced); // NAL
422 buf[0x21] = ctx->cid_table->bit_depth == 10 ? 0x58 : 0x38;
423 buf[0x22] = 0x88 + (ctx->interlaced << 2);
424 AV_WB32(buf + 0x28, ctx->cid); // CID
425 buf[0x2c] = ctx->interlaced ? 0 : 0x80;
427 buf[0x5f] = 0x01; // UDL
429 buf[0x167] = 0x02; // reserved
430 AV_WB16(buf + 0x16a, ctx->m.mb_height * 4 + 4); // MSIPS
431 buf[0x16d] = ctx->m.mb_height; // Ns
432 buf[0x16f] = 0x10; // reserved
434 ctx->msip = buf + 0x170;
438 static av_always_inline void dnxhd_encode_dc(DNXHDEncContext *ctx, int diff)
442 nbits = av_log2_16bit(-2 * diff);
445 nbits = av_log2_16bit(2 * diff);
447 put_bits(&ctx->m.pb, ctx->cid_table->dc_bits[nbits] + nbits,
448 (ctx->cid_table->dc_codes[nbits] << nbits) +
449 (diff & ((1 << nbits) - 1)));
452 static av_always_inline
453 void dnxhd_encode_block(DNXHDEncContext *ctx, int16_t *block,
454 int last_index, int n)
456 int last_non_zero = 0;
459 dnxhd_encode_dc(ctx, block[0] - ctx->m.last_dc[n]);
460 ctx->m.last_dc[n] = block[0];
462 for (i = 1; i <= last_index; i++) {
463 j = ctx->m.intra_scantable.permutated[i];
466 int run_level = i - last_non_zero - 1;
467 int rlevel = (slevel << 1) | !!run_level;
468 put_bits(&ctx->m.pb, ctx->vlc_bits[rlevel], ctx->vlc_codes[rlevel]);
470 put_bits(&ctx->m.pb, ctx->run_bits[run_level],
471 ctx->run_codes[run_level]);
475 put_bits(&ctx->m.pb, ctx->vlc_bits[0], ctx->vlc_codes[0]); // EOB
478 static av_always_inline
479 void dnxhd_unquantize_c(DNXHDEncContext *ctx, int16_t *block, int n,
480 int qscale, int last_index)
482 const uint8_t *weight_matrix;
486 weight_matrix = (n & 2) ? ctx->cid_table->chroma_weight
487 : ctx->cid_table->luma_weight;
489 for (i = 1; i <= last_index; i++) {
490 int j = ctx->m.intra_scantable.permutated[i];
494 level = (1 - 2 * level) * qscale * weight_matrix[i];
495 if (ctx->cid_table->bit_depth == 10) {
496 if (weight_matrix[i] != 8)
500 if (weight_matrix[i] != 32)
506 level = (2 * level + 1) * qscale * weight_matrix[i];
507 if (ctx->cid_table->bit_depth == 10) {
508 if (weight_matrix[i] != 8)
512 if (weight_matrix[i] != 32)
522 static av_always_inline int dnxhd_ssd_block(int16_t *qblock, int16_t *block)
526 for (i = 0; i < 64; i++)
527 score += (block[i] - qblock[i]) * (block[i] - qblock[i]);
531 static av_always_inline
532 int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, int16_t *block, int last_index)
534 int last_non_zero = 0;
537 for (i = 1; i <= last_index; i++) {
538 j = ctx->m.intra_scantable.permutated[i];
541 int run_level = i - last_non_zero - 1;
542 bits += ctx->vlc_bits[(level << 1) |
543 !!run_level] + ctx->run_bits[run_level];
550 static av_always_inline
551 void dnxhd_get_blocks(DNXHDEncContext *ctx, int mb_x, int mb_y)
553 const int bs = ctx->block_width_l2;
554 const int bw = 1 << bs;
555 const uint8_t *ptr_y = ctx->thread[0]->src[0] +
556 ((mb_y << 4) * ctx->m.linesize) + (mb_x << bs + 1);
557 const uint8_t *ptr_u = ctx->thread[0]->src[1] +
558 ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs);
559 const uint8_t *ptr_v = ctx->thread[0]->src[2] +
560 ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs);
561 DSPContext *dsp = &ctx->m.dsp;
563 dsp->get_pixels(ctx->blocks[0], ptr_y, ctx->m.linesize);
564 dsp->get_pixels(ctx->blocks[1], ptr_y + bw, ctx->m.linesize);
565 dsp->get_pixels(ctx->blocks[2], ptr_u, ctx->m.uvlinesize);
566 dsp->get_pixels(ctx->blocks[3], ptr_v, ctx->m.uvlinesize);
568 if (mb_y + 1 == ctx->m.mb_height && ctx->m.avctx->height == 1080) {
569 if (ctx->interlaced) {
570 ctx->get_pixels_8x4_sym(ctx->blocks[4],
571 ptr_y + ctx->dct_y_offset,
573 ctx->get_pixels_8x4_sym(ctx->blocks[5],
574 ptr_y + ctx->dct_y_offset + bw,
576 ctx->get_pixels_8x4_sym(ctx->blocks[6],
577 ptr_u + ctx->dct_uv_offset,
579 ctx->get_pixels_8x4_sym(ctx->blocks[7],
580 ptr_v + ctx->dct_uv_offset,
583 ctx->bdsp.clear_block(ctx->blocks[4]);
584 ctx->bdsp.clear_block(ctx->blocks[5]);
585 ctx->bdsp.clear_block(ctx->blocks[6]);
586 ctx->bdsp.clear_block(ctx->blocks[7]);
589 dsp->get_pixels(ctx->blocks[4],
590 ptr_y + ctx->dct_y_offset, ctx->m.linesize);
591 dsp->get_pixels(ctx->blocks[5],
592 ptr_y + ctx->dct_y_offset + bw, ctx->m.linesize);
593 dsp->get_pixels(ctx->blocks[6],
594 ptr_u + ctx->dct_uv_offset, ctx->m.uvlinesize);
595 dsp->get_pixels(ctx->blocks[7],
596 ptr_v + ctx->dct_uv_offset, ctx->m.uvlinesize);
600 static av_always_inline
601 int dnxhd_switch_matrix(DNXHDEncContext *ctx, int i)
603 const static uint8_t component[8]={0,0,1,2,0,0,1,2};
607 static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg,
608 int jobnr, int threadnr)
610 DNXHDEncContext *ctx = avctx->priv_data;
611 int mb_y = jobnr, mb_x;
612 int qscale = ctx->qscale;
613 LOCAL_ALIGNED_16(int16_t, block, [64]);
614 ctx = ctx->thread[threadnr];
618 ctx->m.last_dc[2] = 1 << (ctx->cid_table->bit_depth + 2);
620 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
621 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
627 dnxhd_get_blocks(ctx, mb_x, mb_y);
629 for (i = 0; i < 8; i++) {
630 int16_t *src_block = ctx->blocks[i];
631 int overflow, nbits, diff, last_index;
632 int n = dnxhd_switch_matrix(ctx, i);
634 memcpy(block, src_block, 64 * sizeof(*block));
635 last_index = ctx->m.dct_quantize(&ctx->m, block, 4 & (2*i),
637 ac_bits += dnxhd_calc_ac_bits(ctx, block, last_index);
639 diff = block[0] - ctx->m.last_dc[n];
641 nbits = av_log2_16bit(-2 * diff);
643 nbits = av_log2_16bit(2 * diff);
645 av_assert1(nbits < ctx->cid_table->bit_depth + 4);
646 dc_bits += ctx->cid_table->dc_bits[nbits] + nbits;
648 ctx->m.last_dc[n] = block[0];
650 if (avctx->mb_decision == FF_MB_DECISION_RD || !RC_VARIANCE) {
651 dnxhd_unquantize_c(ctx, block, i, qscale, last_index);
652 ctx->m.idsp.idct(block);
653 ssd += dnxhd_ssd_block(block, src_block);
656 ctx->mb_rc[qscale][mb].ssd = ssd;
657 ctx->mb_rc[qscale][mb].bits = ac_bits + dc_bits + 12 +
658 8 * ctx->vlc_bits[0];
663 static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg,
664 int jobnr, int threadnr)
666 DNXHDEncContext *ctx = avctx->priv_data;
667 int mb_y = jobnr, mb_x;
668 ctx = ctx->thread[threadnr];
669 init_put_bits(&ctx->m.pb, (uint8_t *)arg + 640 + ctx->slice_offs[jobnr],
670 ctx->slice_size[jobnr]);
674 ctx->m.last_dc[2] = 1 << (ctx->cid_table->bit_depth + 2);
675 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
676 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
677 int qscale = ctx->mb_qscale[mb];
680 put_bits(&ctx->m.pb, 12, qscale << 1);
682 dnxhd_get_blocks(ctx, mb_x, mb_y);
684 for (i = 0; i < 8; i++) {
685 int16_t *block = ctx->blocks[i];
686 int overflow, n = dnxhd_switch_matrix(ctx, i);
687 int last_index = ctx->m.dct_quantize(&ctx->m, block, 4 & (2*i),
690 dnxhd_encode_block(ctx, block, last_index, n);
691 // STOP_TIMER("encode_block");
694 if (put_bits_count(&ctx->m.pb) & 31)
695 put_bits(&ctx->m.pb, 32 - (put_bits_count(&ctx->m.pb) & 31), 0);
696 flush_put_bits(&ctx->m.pb);
700 static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx)
704 for (mb_y = 0; mb_y < ctx->m.mb_height; mb_y++) {
706 ctx->slice_offs[mb_y] = offset;
707 ctx->slice_size[mb_y] = 0;
708 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
709 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
710 ctx->slice_size[mb_y] += ctx->mb_bits[mb];
712 ctx->slice_size[mb_y] = (ctx->slice_size[mb_y] + 31) & ~31;
713 ctx->slice_size[mb_y] >>= 3;
714 thread_size = ctx->slice_size[mb_y];
715 offset += thread_size;
719 static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg,
720 int jobnr, int threadnr)
722 DNXHDEncContext *ctx = avctx->priv_data;
723 int mb_y = jobnr, mb_x, x, y;
724 int partial_last_row = (mb_y == ctx->m.mb_height - 1) &&
725 ((avctx->height >> ctx->interlaced) & 0xF);
727 ctx = ctx->thread[threadnr];
728 if (ctx->cid_table->bit_depth == 8) {
729 uint8_t *pix = ctx->thread[0]->src[0] + ((mb_y << 4) * ctx->m.linesize);
730 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x, pix += 16) {
731 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
735 if (!partial_last_row && mb_x * 16 <= avctx->width - 16) {
736 sum = ctx->m.dsp.pix_sum(pix, ctx->m.linesize);
737 varc = ctx->m.dsp.pix_norm1(pix, ctx->m.linesize);
739 int bw = FFMIN(avctx->width - 16 * mb_x, 16);
740 int bh = FFMIN((avctx->height >> ctx->interlaced) - 16 * mb_y, 16);
742 for (y = 0; y < bh; y++) {
743 for (x = 0; x < bw; x++) {
744 uint8_t val = pix[x + y * ctx->m.linesize];
750 varc = (varc - (((unsigned) sum * sum) >> 8) + 128) >> 8;
752 ctx->mb_cmp[mb].value = varc;
753 ctx->mb_cmp[mb].mb = mb;
756 int const linesize = ctx->m.linesize >> 1;
757 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x) {
758 uint16_t *pix = (uint16_t *)ctx->thread[0]->src[0] +
759 ((mb_y << 4) * linesize) + (mb_x << 4);
760 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
765 // Macroblocks are 16x16 pixels, unlike DCT blocks which are 8x8.
766 for (i = 0; i < 16; ++i) {
767 for (j = 0; j < 16; ++j) {
768 // Turn 16-bit pixels into 10-bit ones.
769 int const sample = (unsigned) pix[j] >> 6;
771 sqsum += sample * sample;
772 // 2^10 * 2^10 * 16 * 16 = 2^28, which is less than INT_MAX
776 mean = sum >> 8; // 16*16 == 2^8
778 ctx->mb_cmp[mb].value = sqmean - mean * mean;
779 ctx->mb_cmp[mb].mb = mb;
785 static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx)
787 int lambda, up_step, down_step;
788 int last_lower = INT_MAX, last_higher = 0;
791 for (q = 1; q < avctx->qmax; q++) {
793 avctx->execute2(avctx, dnxhd_calc_bits_thread,
794 NULL, NULL, ctx->m.mb_height);
796 up_step = down_step = 2 << LAMBDA_FRAC_BITS;
797 lambda = ctx->lambda;
802 if (lambda == last_higher) {
804 end = 1; // need to set final qscales/bits
806 for (y = 0; y < ctx->m.mb_height; y++) {
807 for (x = 0; x < ctx->m.mb_width; x++) {
808 unsigned min = UINT_MAX;
810 int mb = y * ctx->m.mb_width + x;
811 for (q = 1; q < avctx->qmax; q++) {
812 unsigned score = ctx->mb_rc[q][mb].bits * lambda +
813 ((unsigned) ctx->mb_rc[q][mb].ssd << LAMBDA_FRAC_BITS);
819 bits += ctx->mb_rc[qscale][mb].bits;
820 ctx->mb_qscale[mb] = qscale;
821 ctx->mb_bits[mb] = ctx->mb_rc[qscale][mb].bits;
823 bits = (bits + 31) & ~31; // padding
824 if (bits > ctx->frame_bits)
827 // av_dlog(ctx->m.avctx,
828 // "lambda %d, up %u, down %u, bits %d, frame %d\n",
829 // lambda, last_higher, last_lower, bits, ctx->frame_bits);
831 if (bits > ctx->frame_bits)
832 return AVERROR(EINVAL);
835 if (bits < ctx->frame_bits) {
836 last_lower = FFMIN(lambda, last_lower);
837 if (last_higher != 0)
838 lambda = (lambda+last_higher)>>1;
841 down_step = FFMIN((int64_t)down_step*5, INT_MAX);
842 up_step = 1<<LAMBDA_FRAC_BITS;
843 lambda = FFMAX(1, lambda);
844 if (lambda == last_lower)
847 last_higher = FFMAX(lambda, last_higher);
848 if (last_lower != INT_MAX)
849 lambda = (lambda+last_lower)>>1;
850 else if ((int64_t)lambda + up_step > INT_MAX)
851 return AVERROR(EINVAL);
854 up_step = FFMIN((int64_t)up_step*5, INT_MAX);
855 down_step = 1<<LAMBDA_FRAC_BITS;
858 //av_dlog(ctx->m.avctx, "out lambda %d\n", lambda);
859 ctx->lambda = lambda;
863 static int dnxhd_find_qscale(DNXHDEncContext *ctx)
869 int last_lower = INT_MAX;
873 qscale = ctx->qscale;
876 ctx->qscale = qscale;
877 // XXX avoid recalculating bits
878 ctx->m.avctx->execute2(ctx->m.avctx, dnxhd_calc_bits_thread,
879 NULL, NULL, ctx->m.mb_height);
880 for (y = 0; y < ctx->m.mb_height; y++) {
881 for (x = 0; x < ctx->m.mb_width; x++)
882 bits += ctx->mb_rc[qscale][y*ctx->m.mb_width+x].bits;
883 bits = (bits+31)&~31; // padding
884 if (bits > ctx->frame_bits)
887 // av_dlog(ctx->m.avctx,
888 // "%d, qscale %d, bits %d, frame %d, higher %d, lower %d\n",
889 // ctx->m.avctx->frame_number, qscale, bits, ctx->frame_bits,
890 // last_higher, last_lower);
891 if (bits < ctx->frame_bits) {
894 if (last_higher == qscale - 1) {
895 qscale = last_higher;
898 last_lower = FFMIN(qscale, last_lower);
899 if (last_higher != 0)
900 qscale = (qscale + last_higher) >> 1;
902 qscale -= down_step++;
907 if (last_lower == qscale + 1)
909 last_higher = FFMAX(qscale, last_higher);
910 if (last_lower != INT_MAX)
911 qscale = (qscale + last_lower) >> 1;
915 if (qscale >= ctx->m.avctx->qmax)
916 return AVERROR(EINVAL);
919 //av_dlog(ctx->m.avctx, "out qscale %d\n", qscale);
920 ctx->qscale = qscale;
924 #define BUCKET_BITS 8
925 #define RADIX_PASSES 4
926 #define NBUCKETS (1 << BUCKET_BITS)
928 static inline int get_bucket(int value, int shift)
931 value &= NBUCKETS - 1;
932 return NBUCKETS - 1 - value;
935 static void radix_count(const RCCMPEntry *data, int size,
936 int buckets[RADIX_PASSES][NBUCKETS])
939 memset(buckets, 0, sizeof(buckets[0][0]) * RADIX_PASSES * NBUCKETS);
940 for (i = 0; i < size; i++) {
941 int v = data[i].value;
942 for (j = 0; j < RADIX_PASSES; j++) {
943 buckets[j][get_bucket(v, 0)]++;
948 for (j = 0; j < RADIX_PASSES; j++) {
950 for (i = NBUCKETS - 1; i >= 0; i--)
951 buckets[j][i] = offset -= buckets[j][i];
952 av_assert1(!buckets[j][0]);
956 static void radix_sort_pass(RCCMPEntry *dst, const RCCMPEntry *data,
957 int size, int buckets[NBUCKETS], int pass)
959 int shift = pass * BUCKET_BITS;
961 for (i = 0; i < size; i++) {
962 int v = get_bucket(data[i].value, shift);
963 int pos = buckets[v]++;
968 static void radix_sort(RCCMPEntry *data, int size)
970 int buckets[RADIX_PASSES][NBUCKETS];
971 RCCMPEntry *tmp = av_malloc_array(size, sizeof(*tmp));
972 radix_count(data, size, buckets);
973 radix_sort_pass(tmp, data, size, buckets[0], 0);
974 radix_sort_pass(data, tmp, size, buckets[1], 1);
975 if (buckets[2][NBUCKETS - 1] || buckets[3][NBUCKETS - 1]) {
976 radix_sort_pass(tmp, data, size, buckets[2], 2);
977 radix_sort_pass(data, tmp, size, buckets[3], 3);
982 static int dnxhd_encode_fast(AVCodecContext *avctx, DNXHDEncContext *ctx)
986 if ((ret = dnxhd_find_qscale(ctx)) < 0)
988 for (y = 0; y < ctx->m.mb_height; y++) {
989 for (x = 0; x < ctx->m.mb_width; x++) {
990 int mb = y * ctx->m.mb_width + x;
992 ctx->mb_qscale[mb] = ctx->qscale;
993 ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale][mb].bits;
994 max_bits += ctx->mb_rc[ctx->qscale][mb].bits;
996 delta_bits = ctx->mb_rc[ctx->qscale][mb].bits -
997 ctx->mb_rc[ctx->qscale + 1][mb].bits;
998 ctx->mb_cmp[mb].mb = mb;
999 ctx->mb_cmp[mb].value =
1000 delta_bits ? ((ctx->mb_rc[ctx->qscale][mb].ssd -
1001 ctx->mb_rc[ctx->qscale + 1][mb].ssd) * 100) /
1003 : INT_MIN; // avoid increasing qscale
1006 max_bits += 31; // worst padding
1010 avctx->execute2(avctx, dnxhd_mb_var_thread,
1011 NULL, NULL, ctx->m.mb_height);
1012 radix_sort(ctx->mb_cmp, ctx->m.mb_num);
1013 for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) {
1014 int mb = ctx->mb_cmp[x].mb;
1015 max_bits -= ctx->mb_rc[ctx->qscale][mb].bits -
1016 ctx->mb_rc[ctx->qscale + 1][mb].bits;
1017 ctx->mb_qscale[mb] = ctx->qscale + 1;
1018 ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale + 1][mb].bits;
1024 static void dnxhd_load_picture(DNXHDEncContext *ctx, const AVFrame *frame)
1028 for (i = 0; i < ctx->m.avctx->thread_count; i++) {
1029 ctx->thread[i]->m.linesize = frame->linesize[0] << ctx->interlaced;
1030 ctx->thread[i]->m.uvlinesize = frame->linesize[1] << ctx->interlaced;
1031 ctx->thread[i]->dct_y_offset = ctx->m.linesize *8;
1032 ctx->thread[i]->dct_uv_offset = ctx->m.uvlinesize*8;
1035 ctx->m.avctx->coded_frame->interlaced_frame = frame->interlaced_frame;
1036 ctx->cur_field = frame->interlaced_frame && !frame->top_field_first;
1039 static int dnxhd_encode_picture(AVCodecContext *avctx, AVPacket *pkt,
1040 const AVFrame *frame, int *got_packet)
1042 DNXHDEncContext *ctx = avctx->priv_data;
1043 int first_field = 1;
1047 if ((ret = ff_alloc_packet2(avctx, pkt, ctx->cid_table->frame_size)) < 0)
1051 dnxhd_load_picture(ctx, frame);
1054 for (i = 0; i < 3; i++) {
1055 ctx->src[i] = frame->data[i];
1056 if (ctx->interlaced && ctx->cur_field)
1057 ctx->src[i] += frame->linesize[i];
1060 dnxhd_write_header(avctx, buf);
1062 if (avctx->mb_decision == FF_MB_DECISION_RD)
1063 ret = dnxhd_encode_rdo(avctx, ctx);
1065 ret = dnxhd_encode_fast(avctx, ctx);
1067 av_log(avctx, AV_LOG_ERROR,
1068 "picture could not fit ratecontrol constraints, increase qmax\n");
1072 dnxhd_setup_threads_slices(ctx);
1075 for (i = 0; i < ctx->m.mb_height; i++) {
1076 AV_WB32(ctx->msip + i * 4, offset);
1077 offset += ctx->slice_size[i];
1078 av_assert1(!(ctx->slice_size[i] & 3));
1081 avctx->execute2(avctx, dnxhd_encode_thread, buf, NULL, ctx->m.mb_height);
1083 av_assert1(640 + offset + 4 <= ctx->cid_table->coding_unit_size);
1084 memset(buf + 640 + offset, 0,
1085 ctx->cid_table->coding_unit_size - 4 - offset - 640);
1087 AV_WB32(buf + ctx->cid_table->coding_unit_size - 4, 0x600DC0DE); // EOF
1089 if (ctx->interlaced && first_field) {
1091 ctx->cur_field ^= 1;
1092 buf += ctx->cid_table->coding_unit_size;
1093 goto encode_coding_unit;
1096 avctx->coded_frame->quality = ctx->qscale * FF_QP2LAMBDA;
1098 pkt->flags |= AV_PKT_FLAG_KEY;
1103 static av_cold int dnxhd_encode_end(AVCodecContext *avctx)
1105 DNXHDEncContext *ctx = avctx->priv_data;
1106 int max_level = 1 << (ctx->cid_table->bit_depth + 2);
1109 av_free(ctx->vlc_codes - max_level * 2);
1110 av_free(ctx->vlc_bits - max_level * 2);
1111 av_freep(&ctx->run_codes);
1112 av_freep(&ctx->run_bits);
1114 av_freep(&ctx->mb_bits);
1115 av_freep(&ctx->mb_qscale);
1116 av_freep(&ctx->mb_rc);
1117 av_freep(&ctx->mb_cmp);
1118 av_freep(&ctx->slice_size);
1119 av_freep(&ctx->slice_offs);
1121 av_freep(&ctx->qmatrix_c);
1122 av_freep(&ctx->qmatrix_l);
1123 av_freep(&ctx->qmatrix_c16);
1124 av_freep(&ctx->qmatrix_l16);
1126 for (i = 1; i < avctx->thread_count; i++)
1127 av_freep(&ctx->thread[i]);
1129 av_frame_free(&avctx->coded_frame);
1134 static const AVCodecDefault dnxhd_defaults[] = {
1135 { "qmax", "1024" }, /* Maximum quantization scale factor allowed for VC-3 */
1139 AVCodec ff_dnxhd_encoder = {
1141 .long_name = NULL_IF_CONFIG_SMALL("VC3/DNxHD"),
1142 .type = AVMEDIA_TYPE_VIDEO,
1143 .id = AV_CODEC_ID_DNXHD,
1144 .priv_data_size = sizeof(DNXHDEncContext),
1145 .init = dnxhd_encode_init,
1146 .encode2 = dnxhd_encode_picture,
1147 .close = dnxhd_encode_end,
1148 .capabilities = CODEC_CAP_SLICE_THREADS,
1149 .pix_fmts = (const enum AVPixelFormat[]) {
1151 AV_PIX_FMT_YUV422P10,
1154 .priv_class = &dnxhd_class,
1155 .defaults = dnxhd_defaults,