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"
36 #include "pixblockdsp.h"
40 // The largest value that will not lead to overflow for 10-bit samples.
41 #define DNX10BIT_QMAT_SHIFT 18
42 #define RC_VARIANCE 1 // use variance or ssd for fast rc
43 #define LAMBDA_FRAC_BITS 10
45 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
46 static const AVOption options[] = {
47 { "nitris_compat", "encode with Avid Nitris compatibility",
48 offsetof(DNXHDEncContext, nitris_compat), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, VE },
49 { "ibias", "intra quant bias",
50 offsetof(DNXHDEncContext, intra_quant_bias), AV_OPT_TYPE_INT,
51 { .i64 = 0 }, INT_MIN, INT_MAX, VE },
52 { "profile", NULL, offsetof(DNXHDEncContext, profile), AV_OPT_TYPE_INT,
53 { .i64 = FF_PROFILE_DNXHD },
54 FF_PROFILE_DNXHD, FF_PROFILE_DNXHR_444, VE, "profile" },
55 { "dnxhd", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_PROFILE_DNXHD },
56 0, 0, VE, "profile" },
57 { "dnxhr_444", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_PROFILE_DNXHR_444 },
58 0, 0, VE, "profile" },
59 { "dnxhr_hqx", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_PROFILE_DNXHR_HQX },
60 0, 0, VE, "profile" },
61 { "dnxhr_hq", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_PROFILE_DNXHR_HQ },
62 0, 0, VE, "profile" },
63 { "dnxhr_sq", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_PROFILE_DNXHR_SQ },
64 0, 0, VE, "profile" },
65 { "dnxhr_lb", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_PROFILE_DNXHR_LB },
66 0, 0, VE, "profile" },
70 static const AVClass dnxhd_class = {
71 .class_name = "dnxhd",
72 .item_name = av_default_item_name,
74 .version = LIBAVUTIL_VERSION_INT,
77 static void dnxhd_8bit_get_pixels_8x4_sym(int16_t *av_restrict block,
78 const uint8_t *pixels,
82 for (i = 0; i < 4; i++) {
94 memcpy(block, block - 8, sizeof(*block) * 8);
95 memcpy(block + 8, block - 16, sizeof(*block) * 8);
96 memcpy(block + 16, block - 24, sizeof(*block) * 8);
97 memcpy(block + 24, block - 32, sizeof(*block) * 8);
100 static av_always_inline
101 void dnxhd_10bit_get_pixels_8x4_sym(int16_t *av_restrict block,
102 const uint8_t *pixels,
105 memcpy(block + 0 * 8, pixels + 0 * line_size, 8 * sizeof(*block));
106 memcpy(block + 7 * 8, pixels + 0 * line_size, 8 * sizeof(*block));
107 memcpy(block + 1 * 8, pixels + 1 * line_size, 8 * sizeof(*block));
108 memcpy(block + 6 * 8, pixels + 1 * line_size, 8 * sizeof(*block));
109 memcpy(block + 2 * 8, pixels + 2 * line_size, 8 * sizeof(*block));
110 memcpy(block + 5 * 8, pixels + 2 * line_size, 8 * sizeof(*block));
111 memcpy(block + 3 * 8, pixels + 3 * line_size, 8 * sizeof(*block));
112 memcpy(block + 4 * 8, pixels + 3 * line_size, 8 * sizeof(*block));
115 static int dnxhd_10bit_dct_quantize(MpegEncContext *ctx, int16_t *block,
116 int n, int qscale, int *overflow)
118 const uint8_t *scantable= ctx->intra_scantable.scantable;
119 const int *qmat = n<4 ? ctx->q_intra_matrix[qscale] : ctx->q_chroma_intra_matrix[qscale];
120 int last_non_zero = 0;
123 ctx->fdsp.fdct(block);
125 // Divide by 4 with rounding, to compensate scaling of DCT coefficients
126 block[0] = (block[0] + 2) >> 2;
128 for (i = 1; i < 64; ++i) {
129 int j = scantable[i];
130 int sign = FF_SIGNBIT(block[j]);
131 int level = (block[j] ^ sign) - sign;
132 level = level * qmat[j] >> DNX10BIT_QMAT_SHIFT;
133 block[j] = (level ^ sign) - sign;
138 /* we need this permutation so that we correct the IDCT, we only permute the !=0 elements */
139 if (ctx->idsp.perm_type != FF_IDCT_PERM_NONE)
140 ff_block_permute(block, ctx->idsp.idct_permutation,
141 scantable, last_non_zero);
143 return last_non_zero;
146 static av_cold int dnxhd_init_vlc(DNXHDEncContext *ctx)
148 int i, j, level, run;
149 int max_level = 1 << (ctx->cid_table->bit_depth + 2);
151 FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->vlc_codes,
152 max_level, 4 * sizeof(*ctx->vlc_codes), fail);
153 FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->vlc_bits,
154 max_level, 4 * sizeof(*ctx->vlc_bits), fail);
155 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_codes,
157 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_bits,
160 ctx->vlc_codes += max_level * 2;
161 ctx->vlc_bits += max_level * 2;
162 for (level = -max_level; level < max_level; level++) {
163 for (run = 0; run < 2; run++) {
164 int index = (level << 1) | run;
165 int sign, offset = 0, alevel = level;
167 MASK_ABS(sign, alevel);
169 offset = (alevel - 1) >> 6;
170 alevel -= offset << 6;
172 for (j = 0; j < 257; j++) {
173 if (ctx->cid_table->ac_info[2*j+0] >> 1 == alevel &&
174 (!offset || (ctx->cid_table->ac_info[2*j+1] & 1) && offset) &&
175 (!run || (ctx->cid_table->ac_info[2*j+1] & 2) && run)) {
176 av_assert1(!ctx->vlc_codes[index]);
178 ctx->vlc_codes[index] =
179 (ctx->cid_table->ac_codes[j] << 1) | (sign & 1);
180 ctx->vlc_bits[index] = ctx->cid_table->ac_bits[j] + 1;
182 ctx->vlc_codes[index] = ctx->cid_table->ac_codes[j];
183 ctx->vlc_bits[index] = ctx->cid_table->ac_bits[j];
188 av_assert0(!alevel || j < 257);
190 ctx->vlc_codes[index] =
191 (ctx->vlc_codes[index] << ctx->cid_table->index_bits) | offset;
192 ctx->vlc_bits[index] += ctx->cid_table->index_bits;
196 for (i = 0; i < 62; i++) {
197 int run = ctx->cid_table->run[i];
198 av_assert0(run < 63);
199 ctx->run_codes[run] = ctx->cid_table->run_codes[i];
200 ctx->run_bits[run] = ctx->cid_table->run_bits[i];
204 return AVERROR(ENOMEM);
207 static av_cold int dnxhd_init_qmat(DNXHDEncContext *ctx, int lbias, int cbias)
209 // init first elem to 1 to avoid div by 0 in convert_matrix
210 uint16_t weight_matrix[64] = { 1, }; // convert_matrix needs uint16_t*
212 const uint8_t *luma_weight_table = ctx->cid_table->luma_weight;
213 const uint8_t *chroma_weight_table = ctx->cid_table->chroma_weight;
215 FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l,
216 (ctx->m.avctx->qmax + 1), 64 * sizeof(int), fail);
217 FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c,
218 (ctx->m.avctx->qmax + 1), 64 * sizeof(int), fail);
219 FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l16,
220 (ctx->m.avctx->qmax + 1), 64 * 2 * sizeof(uint16_t),
222 FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c16,
223 (ctx->m.avctx->qmax + 1), 64 * 2 * sizeof(uint16_t),
226 if (ctx->cid_table->bit_depth == 8) {
227 for (i = 1; i < 64; i++) {
228 int j = ctx->m.idsp.idct_permutation[ff_zigzag_direct[i]];
229 weight_matrix[j] = ctx->cid_table->luma_weight[i];
231 ff_convert_matrix(&ctx->m, ctx->qmatrix_l, ctx->qmatrix_l16,
232 weight_matrix, ctx->intra_quant_bias, 1,
233 ctx->m.avctx->qmax, 1);
234 for (i = 1; i < 64; i++) {
235 int j = ctx->m.idsp.idct_permutation[ff_zigzag_direct[i]];
236 weight_matrix[j] = ctx->cid_table->chroma_weight[i];
238 ff_convert_matrix(&ctx->m, ctx->qmatrix_c, ctx->qmatrix_c16,
239 weight_matrix, ctx->intra_quant_bias, 1,
240 ctx->m.avctx->qmax, 1);
242 for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
243 for (i = 0; i < 64; i++) {
244 ctx->qmatrix_l[qscale][i] <<= 2;
245 ctx->qmatrix_c[qscale][i] <<= 2;
246 ctx->qmatrix_l16[qscale][0][i] <<= 2;
247 ctx->qmatrix_l16[qscale][1][i] <<= 2;
248 ctx->qmatrix_c16[qscale][0][i] <<= 2;
249 ctx->qmatrix_c16[qscale][1][i] <<= 2;
254 for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
255 for (i = 1; i < 64; i++) {
256 int j = ff_zigzag_direct[i];
258 /* The quantization formula from the VC-3 standard is:
259 * quantized = sign(block[i]) * floor(abs(block[i]/s) * p /
260 * (qscale * weight_table[i]))
261 * Where p is 32 for 8-bit samples and 8 for 10-bit ones.
262 * The s factor compensates scaling of DCT coefficients done by
263 * the DCT routines, and therefore is not present in standard.
264 * It's 8 for 8-bit samples and 4 for 10-bit ones.
265 * We want values of ctx->qtmatrix_l and ctx->qtmatrix_r to be:
266 * ((1 << DNX10BIT_QMAT_SHIFT) * (p / s)) /
267 * (qscale * weight_table[i])
268 * For 10-bit samples, p / s == 2 */
269 ctx->qmatrix_l[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) /
270 (qscale * luma_weight_table[i]);
271 ctx->qmatrix_c[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) /
272 (qscale * chroma_weight_table[i]);
277 ctx->m.q_chroma_intra_matrix16 = ctx->qmatrix_c16;
278 ctx->m.q_chroma_intra_matrix = ctx->qmatrix_c;
279 ctx->m.q_intra_matrix16 = ctx->qmatrix_l16;
280 ctx->m.q_intra_matrix = ctx->qmatrix_l;
284 return AVERROR(ENOMEM);
287 static av_cold int dnxhd_init_rc(DNXHDEncContext *ctx)
289 FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->mb_rc, (ctx->m.avctx->qmax + 1),
290 ctx->m.mb_num * sizeof(RCEntry), fail);
291 if (ctx->m.avctx->mb_decision != FF_MB_DECISION_RD)
292 FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->mb_cmp,
293 ctx->m.mb_num, sizeof(RCCMPEntry), fail);
294 ctx->frame_bits = (ctx->coding_unit_size -
295 ctx->data_offset - 4 - ctx->min_padding) * 8;
297 ctx->lambda = 2 << LAMBDA_FRAC_BITS; // qscale 2
300 return AVERROR(ENOMEM);
303 static int dnxhd_get_hr_frame_size(const CIDEntry* profile, int mb_num)
305 int result = mb_num * profile->packet_scale.num / profile->packet_scale.den;
306 result = (result + 2048) / 4096 * 4096;
307 return FFMAX(result, 8192);
309 static av_cold int dnxhd_encode_init(AVCodecContext *avctx)
311 DNXHDEncContext *ctx = avctx->priv_data;
312 int i, index, bit_depth, ret;
314 switch (avctx->pix_fmt) {
315 case AV_PIX_FMT_YUV422P:
318 case AV_PIX_FMT_YUV422P10:
322 av_log(avctx, AV_LOG_ERROR,
323 "pixel format is incompatible with DNxHD\n");
324 return AVERROR(EINVAL);
327 if (ctx->profile == FF_PROFILE_DNXHR_HQX ||
328 ctx->profile == FF_PROFILE_DNXHR_HQX) {
329 avpriv_report_missing_feature(avctx,
330 "dnxhr_444 or dnxhr_hqx profile");
331 return AVERROR_PATCHWELCOME;
334 avctx->profile = ctx->profile;
335 ctx->cid = ff_dnxhd_find_cid(avctx, bit_depth);
337 av_log(avctx, AV_LOG_ERROR,
338 "video parameters incompatible with DNxHD. Valid DNxHD profiles:\n");
339 ff_dnxhd_print_profiles(avctx, AV_LOG_ERROR);
340 return AVERROR(EINVAL);
342 av_log(avctx, AV_LOG_DEBUG, "cid %d\n", ctx->cid);
344 if (ctx->cid >= 1270 && ctx->cid <= 1274)
345 avctx->codec_tag = MKTAG('A','V','d','h');
347 if (avctx->width < 256 || avctx->height < 120) {
348 av_log(avctx, AV_LOG_ERROR,
349 "Input dimensions too small, input must be at least 256x120\n");
350 return AVERROR(EINVAL);
353 index = ff_dnxhd_get_cid_table(ctx->cid);
354 av_assert0(index >= 0);
356 ctx->cid_table = &ff_dnxhd_cid_table[index];
358 ctx->m.avctx = avctx;
362 avctx->bits_per_raw_sample = ctx->cid_table->bit_depth;
364 ff_blockdsp_init(&ctx->bdsp, avctx);
365 ff_fdctdsp_init(&ctx->m.fdsp, avctx);
366 ff_mpv_idct_init(&ctx->m);
367 ff_mpegvideoencdsp_init(&ctx->m.mpvencdsp, avctx);
368 ff_pixblockdsp_init(&ctx->m.pdsp, avctx);
369 ff_dct_encode_init(&ctx->m);
371 if (!ctx->m.dct_quantize)
372 ctx->m.dct_quantize = ff_dct_quantize_c;
374 if (ctx->cid_table->bit_depth == 10) {
375 ctx->m.dct_quantize = dnxhd_10bit_dct_quantize;
376 ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym;
377 ctx->block_width_l2 = 4;
379 ctx->get_pixels_8x4_sym = dnxhd_8bit_get_pixels_8x4_sym;
380 ctx->block_width_l2 = 3;
384 ff_dnxhdenc_init_x86(ctx);
386 ctx->m.mb_height = (avctx->height + 15) / 16;
387 ctx->m.mb_width = (avctx->width + 15) / 16;
389 if (avctx->flags & AV_CODEC_FLAG_INTERLACED_DCT) {
391 ctx->m.mb_height /= 2;
394 ctx->m.mb_num = ctx->m.mb_height * ctx->m.mb_width;
396 if (ctx->cid_table->frame_size == DNXHD_VARIABLE) {
397 ctx->frame_size = dnxhd_get_hr_frame_size(ctx->cid_table,
399 ctx->coding_unit_size = ctx->frame_size;
401 ctx->frame_size = ctx->cid_table->frame_size;
402 ctx->coding_unit_size = ctx->cid_table->coding_unit_size;
405 if (ctx->m.mb_height > 68)
406 ctx->data_offset = 0x170 + (ctx->m.mb_height << 2);
408 ctx->data_offset = 0x280;
410 #if FF_API_QUANT_BIAS
411 FF_DISABLE_DEPRECATION_WARNINGS
412 if (avctx->intra_quant_bias != FF_DEFAULT_QUANT_BIAS)
413 ctx->intra_quant_bias = avctx->intra_quant_bias;
414 FF_ENABLE_DEPRECATION_WARNINGS
416 // XXX tune lbias/cbias
417 if ((ret = dnxhd_init_qmat(ctx, ctx->intra_quant_bias, 0)) < 0)
420 /* Avid Nitris hardware decoder requires a minimum amount of padding
421 * in the coding unit payload */
422 if (ctx->nitris_compat)
423 ctx->min_padding = 1600;
425 if ((ret = dnxhd_init_vlc(ctx)) < 0)
427 if ((ret = dnxhd_init_rc(ctx)) < 0)
430 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_size,
431 ctx->m.mb_height * sizeof(uint32_t), fail);
432 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_offs,
433 ctx->m.mb_height * sizeof(uint32_t), fail);
434 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_bits,
435 ctx->m.mb_num * sizeof(uint16_t), fail);
436 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_qscale,
437 ctx->m.mb_num * sizeof(uint8_t), fail);
439 #if FF_API_CODED_FRAME
440 FF_DISABLE_DEPRECATION_WARNINGS
441 avctx->coded_frame->key_frame = 1;
442 avctx->coded_frame->pict_type = AV_PICTURE_TYPE_I;
443 FF_ENABLE_DEPRECATION_WARNINGS
446 if (avctx->thread_count > MAX_THREADS) {
447 av_log(avctx, AV_LOG_ERROR, "too many threads\n");
448 return AVERROR(EINVAL);
451 if (avctx->qmax <= 1) {
452 av_log(avctx, AV_LOG_ERROR, "qmax must be at least 2\n");
453 return AVERROR(EINVAL);
456 ctx->thread[0] = ctx;
457 for (i = 1; i < avctx->thread_count; i++) {
458 ctx->thread[i] = av_malloc(sizeof(DNXHDEncContext));
459 memcpy(ctx->thread[i], ctx, sizeof(DNXHDEncContext));
463 fail: // for FF_ALLOCZ_OR_GOTO
464 return AVERROR(ENOMEM);
467 static int dnxhd_write_header(AVCodecContext *avctx, uint8_t *buf)
469 DNXHDEncContext *ctx = avctx->priv_data;
471 memset(buf, 0, ctx->data_offset);
474 AV_WB16(buf + 0x02, ctx->data_offset);
475 if (ctx->cid >= 1270 && ctx->cid <= 1274)
480 buf[5] = ctx->interlaced ? ctx->cur_field + 2 : 0x01;
481 buf[6] = 0x80; // crc flag off
482 buf[7] = 0xa0; // reserved
483 AV_WB16(buf + 0x18, avctx->height >> ctx->interlaced); // ALPF
484 AV_WB16(buf + 0x1a, avctx->width); // SPL
485 AV_WB16(buf + 0x1d, avctx->height >> ctx->interlaced); // NAL
487 buf[0x21] = ctx->cid_table->bit_depth == 10 ? 0x58 : 0x38;
488 buf[0x22] = 0x88 + (ctx->interlaced << 2);
489 AV_WB32(buf + 0x28, ctx->cid); // CID
490 buf[0x2c] = ctx->interlaced ? 0 : 0x80;
492 buf[0x5f] = 0x01; // UDL
494 buf[0x167] = 0x02; // reserved
495 AV_WB16(buf + 0x16a, ctx->m.mb_height * 4 + 4); // MSIPS
496 buf[0x16d] = ctx->m.mb_height; // Ns
497 buf[0x16f] = 0x10; // reserved
499 ctx->msip = buf + 0x170;
503 static av_always_inline void dnxhd_encode_dc(DNXHDEncContext *ctx, int diff)
507 nbits = av_log2_16bit(-2 * diff);
510 nbits = av_log2_16bit(2 * diff);
512 put_bits(&ctx->m.pb, ctx->cid_table->dc_bits[nbits] + nbits,
513 (ctx->cid_table->dc_codes[nbits] << nbits) +
514 av_mod_uintp2(diff, nbits));
517 static av_always_inline
518 void dnxhd_encode_block(DNXHDEncContext *ctx, int16_t *block,
519 int last_index, int n)
521 int last_non_zero = 0;
524 dnxhd_encode_dc(ctx, block[0] - ctx->m.last_dc[n]);
525 ctx->m.last_dc[n] = block[0];
527 for (i = 1; i <= last_index; i++) {
528 j = ctx->m.intra_scantable.permutated[i];
531 int run_level = i - last_non_zero - 1;
532 int rlevel = (slevel << 1) | !!run_level;
533 put_bits(&ctx->m.pb, ctx->vlc_bits[rlevel], ctx->vlc_codes[rlevel]);
535 put_bits(&ctx->m.pb, ctx->run_bits[run_level],
536 ctx->run_codes[run_level]);
540 put_bits(&ctx->m.pb, ctx->vlc_bits[0], ctx->vlc_codes[0]); // EOB
543 static av_always_inline
544 void dnxhd_unquantize_c(DNXHDEncContext *ctx, int16_t *block, int n,
545 int qscale, int last_index)
547 const uint8_t *weight_matrix;
551 weight_matrix = (n & 2) ? ctx->cid_table->chroma_weight
552 : ctx->cid_table->luma_weight;
554 for (i = 1; i <= last_index; i++) {
555 int j = ctx->m.intra_scantable.permutated[i];
559 level = (1 - 2 * level) * qscale * weight_matrix[i];
560 if (ctx->cid_table->bit_depth == 10) {
561 if (weight_matrix[i] != 8)
565 if (weight_matrix[i] != 32)
571 level = (2 * level + 1) * qscale * weight_matrix[i];
572 if (ctx->cid_table->bit_depth == 10) {
573 if (weight_matrix[i] != 8)
577 if (weight_matrix[i] != 32)
587 static av_always_inline int dnxhd_ssd_block(int16_t *qblock, int16_t *block)
591 for (i = 0; i < 64; i++)
592 score += (block[i] - qblock[i]) * (block[i] - qblock[i]);
596 static av_always_inline
597 int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, int16_t *block, int last_index)
599 int last_non_zero = 0;
602 for (i = 1; i <= last_index; i++) {
603 j = ctx->m.intra_scantable.permutated[i];
606 int run_level = i - last_non_zero - 1;
607 bits += ctx->vlc_bits[(level << 1) |
608 !!run_level] + ctx->run_bits[run_level];
615 static av_always_inline
616 void dnxhd_get_blocks(DNXHDEncContext *ctx, int mb_x, int mb_y)
618 const int bs = ctx->block_width_l2;
619 const int bw = 1 << bs;
620 const uint8_t *ptr_y = ctx->thread[0]->src[0] +
621 ((mb_y << 4) * ctx->m.linesize) + (mb_x << bs + 1);
622 const uint8_t *ptr_u = ctx->thread[0]->src[1] +
623 ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs);
624 const uint8_t *ptr_v = ctx->thread[0]->src[2] +
625 ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs);
626 PixblockDSPContext *pdsp = &ctx->m.pdsp;
628 pdsp->get_pixels(ctx->blocks[0], ptr_y, ctx->m.linesize);
629 pdsp->get_pixels(ctx->blocks[1], ptr_y + bw, ctx->m.linesize);
630 pdsp->get_pixels(ctx->blocks[2], ptr_u, ctx->m.uvlinesize);
631 pdsp->get_pixels(ctx->blocks[3], ptr_v, ctx->m.uvlinesize);
633 if (mb_y + 1 == ctx->m.mb_height && (ctx->m.avctx->height % 16) != 0) {
634 if (ctx->interlaced) {
635 ctx->get_pixels_8x4_sym(ctx->blocks[4],
636 ptr_y + ctx->dct_y_offset,
638 ctx->get_pixels_8x4_sym(ctx->blocks[5],
639 ptr_y + ctx->dct_y_offset + bw,
641 ctx->get_pixels_8x4_sym(ctx->blocks[6],
642 ptr_u + ctx->dct_uv_offset,
644 ctx->get_pixels_8x4_sym(ctx->blocks[7],
645 ptr_v + ctx->dct_uv_offset,
648 ctx->bdsp.clear_block(ctx->blocks[4]);
649 ctx->bdsp.clear_block(ctx->blocks[5]);
650 ctx->bdsp.clear_block(ctx->blocks[6]);
651 ctx->bdsp.clear_block(ctx->blocks[7]);
654 pdsp->get_pixels(ctx->blocks[4],
655 ptr_y + ctx->dct_y_offset, ctx->m.linesize);
656 pdsp->get_pixels(ctx->blocks[5],
657 ptr_y + ctx->dct_y_offset + bw, ctx->m.linesize);
658 pdsp->get_pixels(ctx->blocks[6],
659 ptr_u + ctx->dct_uv_offset, ctx->m.uvlinesize);
660 pdsp->get_pixels(ctx->blocks[7],
661 ptr_v + ctx->dct_uv_offset, ctx->m.uvlinesize);
665 static av_always_inline
666 int dnxhd_switch_matrix(DNXHDEncContext *ctx, int i)
668 const static uint8_t component[8]={0,0,1,2,0,0,1,2};
672 static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg,
673 int jobnr, int threadnr)
675 DNXHDEncContext *ctx = avctx->priv_data;
676 int mb_y = jobnr, mb_x;
677 int qscale = ctx->qscale;
678 LOCAL_ALIGNED_16(int16_t, block, [64]);
679 ctx = ctx->thread[threadnr];
683 ctx->m.last_dc[2] = 1 << (ctx->cid_table->bit_depth + 2);
685 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
686 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
692 dnxhd_get_blocks(ctx, mb_x, mb_y);
694 for (i = 0; i < 8; i++) {
695 int16_t *src_block = ctx->blocks[i];
696 int overflow, nbits, diff, last_index;
697 int n = dnxhd_switch_matrix(ctx, i);
699 memcpy(block, src_block, 64 * sizeof(*block));
700 last_index = ctx->m.dct_quantize(&ctx->m, block, 4 & (2*i),
702 ac_bits += dnxhd_calc_ac_bits(ctx, block, last_index);
704 diff = block[0] - ctx->m.last_dc[n];
706 nbits = av_log2_16bit(-2 * diff);
708 nbits = av_log2_16bit(2 * diff);
710 av_assert1(nbits < ctx->cid_table->bit_depth + 4);
711 dc_bits += ctx->cid_table->dc_bits[nbits] + nbits;
713 ctx->m.last_dc[n] = block[0];
715 if (avctx->mb_decision == FF_MB_DECISION_RD || !RC_VARIANCE) {
716 dnxhd_unquantize_c(ctx, block, i, qscale, last_index);
717 ctx->m.idsp.idct(block);
718 ssd += dnxhd_ssd_block(block, src_block);
721 ctx->mb_rc[(qscale * ctx->m.mb_num) + mb].ssd = ssd;
722 ctx->mb_rc[(qscale * ctx->m.mb_num) + mb].bits = ac_bits + dc_bits + 12 +
723 8 * ctx->vlc_bits[0];
728 static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg,
729 int jobnr, int threadnr)
731 DNXHDEncContext *ctx = avctx->priv_data;
732 int mb_y = jobnr, mb_x;
733 ctx = ctx->thread[threadnr];
734 init_put_bits(&ctx->m.pb, (uint8_t *)arg + ctx->data_offset + ctx->slice_offs[jobnr],
735 ctx->slice_size[jobnr]);
739 ctx->m.last_dc[2] = 1 << (ctx->cid_table->bit_depth + 2);
740 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
741 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
742 int qscale = ctx->mb_qscale[mb];
745 put_bits(&ctx->m.pb, 12, qscale << 1);
747 dnxhd_get_blocks(ctx, mb_x, mb_y);
749 for (i = 0; i < 8; i++) {
750 int16_t *block = ctx->blocks[i];
751 int overflow, n = dnxhd_switch_matrix(ctx, i);
752 int last_index = ctx->m.dct_quantize(&ctx->m, block, 4 & (2*i),
755 dnxhd_encode_block(ctx, block, last_index, n);
756 // STOP_TIMER("encode_block");
759 if (put_bits_count(&ctx->m.pb) & 31)
760 put_bits(&ctx->m.pb, 32 - (put_bits_count(&ctx->m.pb) & 31), 0);
761 flush_put_bits(&ctx->m.pb);
765 static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx)
769 for (mb_y = 0; mb_y < ctx->m.mb_height; mb_y++) {
771 ctx->slice_offs[mb_y] = offset;
772 ctx->slice_size[mb_y] = 0;
773 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
774 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
775 ctx->slice_size[mb_y] += ctx->mb_bits[mb];
777 ctx->slice_size[mb_y] = (ctx->slice_size[mb_y] + 31) & ~31;
778 ctx->slice_size[mb_y] >>= 3;
779 thread_size = ctx->slice_size[mb_y];
780 offset += thread_size;
784 static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg,
785 int jobnr, int threadnr)
787 DNXHDEncContext *ctx = avctx->priv_data;
788 int mb_y = jobnr, mb_x, x, y;
789 int partial_last_row = (mb_y == ctx->m.mb_height - 1) &&
790 ((avctx->height >> ctx->interlaced) & 0xF);
792 ctx = ctx->thread[threadnr];
793 if (ctx->cid_table->bit_depth == 8) {
794 uint8_t *pix = ctx->thread[0]->src[0] + ((mb_y << 4) * ctx->m.linesize);
795 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x, pix += 16) {
796 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
800 if (!partial_last_row && mb_x * 16 <= avctx->width - 16 && (avctx->width % 16) == 0) {
801 sum = ctx->m.mpvencdsp.pix_sum(pix, ctx->m.linesize);
802 varc = ctx->m.mpvencdsp.pix_norm1(pix, ctx->m.linesize);
804 int bw = FFMIN(avctx->width - 16 * mb_x, 16);
805 int bh = FFMIN((avctx->height >> ctx->interlaced) - 16 * mb_y, 16);
807 for (y = 0; y < bh; y++) {
808 for (x = 0; x < bw; x++) {
809 uint8_t val = pix[x + y * ctx->m.linesize];
815 varc = (varc - (((unsigned) sum * sum) >> 8) + 128) >> 8;
817 ctx->mb_cmp[mb].value = varc;
818 ctx->mb_cmp[mb].mb = mb;
821 const int linesize = ctx->m.linesize >> 1;
822 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x) {
823 uint16_t *pix = (uint16_t *)ctx->thread[0]->src[0] +
824 ((mb_y << 4) * linesize) + (mb_x << 4);
825 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
828 int bw = FFMIN(avctx->width - 16 * mb_x, 16);
829 int bh = FFMIN((avctx->height >> ctx->interlaced) - 16 * mb_y, 16);
832 // Macroblocks are 16x16 pixels, unlike DCT blocks which are 8x8.
833 for (i = 0; i < bh; ++i) {
834 for (j = 0; j < bw; ++j) {
835 // Turn 16-bit pixels into 10-bit ones.
836 const int sample = (unsigned) pix[j] >> 6;
838 sqsum += sample * sample;
839 // 2^10 * 2^10 * 16 * 16 = 2^28, which is less than INT_MAX
843 mean = sum >> 8; // 16*16 == 2^8
845 ctx->mb_cmp[mb].value = sqmean - mean * mean;
846 ctx->mb_cmp[mb].mb = mb;
852 static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx)
854 int lambda, up_step, down_step;
855 int last_lower = INT_MAX, last_higher = 0;
858 for (q = 1; q < avctx->qmax; q++) {
860 avctx->execute2(avctx, dnxhd_calc_bits_thread,
861 NULL, NULL, ctx->m.mb_height);
863 up_step = down_step = 2 << LAMBDA_FRAC_BITS;
864 lambda = ctx->lambda;
869 if (lambda == last_higher) {
871 end = 1; // need to set final qscales/bits
873 for (y = 0; y < ctx->m.mb_height; y++) {
874 for (x = 0; x < ctx->m.mb_width; x++) {
875 unsigned min = UINT_MAX;
877 int mb = y * ctx->m.mb_width + x;
879 for (q = 1; q < avctx->qmax; q++) {
880 int i = (q*ctx->m.mb_num) + mb;
881 unsigned score = ctx->mb_rc[i].bits * lambda +
882 ((unsigned) ctx->mb_rc[i].ssd << LAMBDA_FRAC_BITS);
889 bits += ctx->mb_rc[rc].bits;
890 ctx->mb_qscale[mb] = qscale;
891 ctx->mb_bits[mb] = ctx->mb_rc[rc].bits;
893 bits = (bits + 31) & ~31; // padding
894 if (bits > ctx->frame_bits)
897 // ff_dlog(ctx->m.avctx,
898 // "lambda %d, up %u, down %u, bits %d, frame %d\n",
899 // lambda, last_higher, last_lower, bits, ctx->frame_bits);
901 if (bits > ctx->frame_bits)
902 return AVERROR(EINVAL);
905 if (bits < ctx->frame_bits) {
906 last_lower = FFMIN(lambda, last_lower);
907 if (last_higher != 0)
908 lambda = (lambda+last_higher)>>1;
911 down_step = FFMIN((int64_t)down_step*5, INT_MAX);
912 up_step = 1<<LAMBDA_FRAC_BITS;
913 lambda = FFMAX(1, lambda);
914 if (lambda == last_lower)
917 last_higher = FFMAX(lambda, last_higher);
918 if (last_lower != INT_MAX)
919 lambda = (lambda+last_lower)>>1;
920 else if ((int64_t)lambda + up_step > INT_MAX)
921 return AVERROR(EINVAL);
924 up_step = FFMIN((int64_t)up_step*5, INT_MAX);
925 down_step = 1<<LAMBDA_FRAC_BITS;
928 //ff_dlog(ctx->m.avctx, "out lambda %d\n", lambda);
929 ctx->lambda = lambda;
933 static int dnxhd_find_qscale(DNXHDEncContext *ctx)
939 int last_lower = INT_MAX;
943 qscale = ctx->qscale;
946 ctx->qscale = qscale;
947 // XXX avoid recalculating bits
948 ctx->m.avctx->execute2(ctx->m.avctx, dnxhd_calc_bits_thread,
949 NULL, NULL, ctx->m.mb_height);
950 for (y = 0; y < ctx->m.mb_height; y++) {
951 for (x = 0; x < ctx->m.mb_width; x++)
952 bits += ctx->mb_rc[(qscale*ctx->m.mb_num) + (y*ctx->m.mb_width+x)].bits;
953 bits = (bits+31)&~31; // padding
954 if (bits > ctx->frame_bits)
957 // ff_dlog(ctx->m.avctx,
958 // "%d, qscale %d, bits %d, frame %d, higher %d, lower %d\n",
959 // ctx->m.avctx->frame_number, qscale, bits, ctx->frame_bits,
960 // last_higher, last_lower);
961 if (bits < ctx->frame_bits) {
964 if (last_higher == qscale - 1) {
965 qscale = last_higher;
968 last_lower = FFMIN(qscale, last_lower);
969 if (last_higher != 0)
970 qscale = (qscale + last_higher) >> 1;
972 qscale -= down_step++;
977 if (last_lower == qscale + 1)
979 last_higher = FFMAX(qscale, last_higher);
980 if (last_lower != INT_MAX)
981 qscale = (qscale + last_lower) >> 1;
985 if (qscale >= ctx->m.avctx->qmax)
986 return AVERROR(EINVAL);
989 //ff_dlog(ctx->m.avctx, "out qscale %d\n", qscale);
990 ctx->qscale = qscale;
994 #define BUCKET_BITS 8
995 #define RADIX_PASSES 4
996 #define NBUCKETS (1 << BUCKET_BITS)
998 static inline int get_bucket(int value, int shift)
1001 value &= NBUCKETS - 1;
1002 return NBUCKETS - 1 - value;
1005 static void radix_count(const RCCMPEntry *data, int size,
1006 int buckets[RADIX_PASSES][NBUCKETS])
1009 memset(buckets, 0, sizeof(buckets[0][0]) * RADIX_PASSES * NBUCKETS);
1010 for (i = 0; i < size; i++) {
1011 int v = data[i].value;
1012 for (j = 0; j < RADIX_PASSES; j++) {
1013 buckets[j][get_bucket(v, 0)]++;
1018 for (j = 0; j < RADIX_PASSES; j++) {
1020 for (i = NBUCKETS - 1; i >= 0; i--)
1021 buckets[j][i] = offset -= buckets[j][i];
1022 av_assert1(!buckets[j][0]);
1026 static void radix_sort_pass(RCCMPEntry *dst, const RCCMPEntry *data,
1027 int size, int buckets[NBUCKETS], int pass)
1029 int shift = pass * BUCKET_BITS;
1031 for (i = 0; i < size; i++) {
1032 int v = get_bucket(data[i].value, shift);
1033 int pos = buckets[v]++;
1038 static void radix_sort(RCCMPEntry *data, int size)
1040 int buckets[RADIX_PASSES][NBUCKETS];
1041 RCCMPEntry *tmp = av_malloc_array(size, sizeof(*tmp));
1042 radix_count(data, size, buckets);
1043 radix_sort_pass(tmp, data, size, buckets[0], 0);
1044 radix_sort_pass(data, tmp, size, buckets[1], 1);
1045 if (buckets[2][NBUCKETS - 1] || buckets[3][NBUCKETS - 1]) {
1046 radix_sort_pass(tmp, data, size, buckets[2], 2);
1047 radix_sort_pass(data, tmp, size, buckets[3], 3);
1052 static int dnxhd_encode_fast(AVCodecContext *avctx, DNXHDEncContext *ctx)
1056 if ((ret = dnxhd_find_qscale(ctx)) < 0)
1058 for (y = 0; y < ctx->m.mb_height; y++) {
1059 for (x = 0; x < ctx->m.mb_width; x++) {
1060 int mb = y * ctx->m.mb_width + x;
1061 int rc = (ctx->qscale * ctx->m.mb_num ) + mb;
1063 ctx->mb_qscale[mb] = ctx->qscale;
1064 ctx->mb_bits[mb] = ctx->mb_rc[rc].bits;
1065 max_bits += ctx->mb_rc[rc].bits;
1067 delta_bits = ctx->mb_rc[rc].bits -
1068 ctx->mb_rc[rc + ctx->m.mb_num].bits;
1069 ctx->mb_cmp[mb].mb = mb;
1070 ctx->mb_cmp[mb].value =
1071 delta_bits ? ((ctx->mb_rc[rc].ssd -
1072 ctx->mb_rc[rc + ctx->m.mb_num].ssd) * 100) /
1074 : INT_MIN; // avoid increasing qscale
1077 max_bits += 31; // worst padding
1081 avctx->execute2(avctx, dnxhd_mb_var_thread,
1082 NULL, NULL, ctx->m.mb_height);
1083 radix_sort(ctx->mb_cmp, ctx->m.mb_num);
1084 for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) {
1085 int mb = ctx->mb_cmp[x].mb;
1086 int rc = (ctx->qscale * ctx->m.mb_num ) + mb;
1087 max_bits -= ctx->mb_rc[rc].bits -
1088 ctx->mb_rc[rc + ctx->m.mb_num].bits;
1089 ctx->mb_qscale[mb] = ctx->qscale + 1;
1090 ctx->mb_bits[mb] = ctx->mb_rc[rc + ctx->m.mb_num].bits;
1096 static void dnxhd_load_picture(DNXHDEncContext *ctx, const AVFrame *frame)
1100 for (i = 0; i < ctx->m.avctx->thread_count; i++) {
1101 ctx->thread[i]->m.linesize = frame->linesize[0] << ctx->interlaced;
1102 ctx->thread[i]->m.uvlinesize = frame->linesize[1] << ctx->interlaced;
1103 ctx->thread[i]->dct_y_offset = ctx->m.linesize *8;
1104 ctx->thread[i]->dct_uv_offset = ctx->m.uvlinesize*8;
1107 #if FF_API_CODED_FRAME
1108 FF_DISABLE_DEPRECATION_WARNINGS
1109 ctx->m.avctx->coded_frame->interlaced_frame = frame->interlaced_frame;
1110 FF_ENABLE_DEPRECATION_WARNINGS
1112 ctx->cur_field = frame->interlaced_frame && !frame->top_field_first;
1115 static int dnxhd_encode_picture(AVCodecContext *avctx, AVPacket *pkt,
1116 const AVFrame *frame, int *got_packet)
1118 DNXHDEncContext *ctx = avctx->priv_data;
1119 int first_field = 1;
1123 if ((ret = ff_alloc_packet2(avctx, pkt, ctx->frame_size, 0)) < 0)
1127 dnxhd_load_picture(ctx, frame);
1130 for (i = 0; i < 3; i++) {
1131 ctx->src[i] = frame->data[i];
1132 if (ctx->interlaced && ctx->cur_field)
1133 ctx->src[i] += frame->linesize[i];
1136 dnxhd_write_header(avctx, buf);
1138 if (avctx->mb_decision == FF_MB_DECISION_RD)
1139 ret = dnxhd_encode_rdo(avctx, ctx);
1141 ret = dnxhd_encode_fast(avctx, ctx);
1143 av_log(avctx, AV_LOG_ERROR,
1144 "picture could not fit ratecontrol constraints, increase qmax\n");
1148 dnxhd_setup_threads_slices(ctx);
1151 for (i = 0; i < ctx->m.mb_height; i++) {
1152 AV_WB32(ctx->msip + i * 4, offset);
1153 offset += ctx->slice_size[i];
1154 av_assert1(!(ctx->slice_size[i] & 3));
1157 avctx->execute2(avctx, dnxhd_encode_thread, buf, NULL, ctx->m.mb_height);
1159 av_assert1(ctx->data_offset + offset + 4 <= ctx->coding_unit_size);
1160 memset(buf + ctx->data_offset + offset, 0,
1161 ctx->coding_unit_size - 4 - offset - ctx->data_offset);
1163 AV_WB32(buf + ctx->coding_unit_size - 4, 0x600DC0DE); // EOF
1165 if (ctx->interlaced && first_field) {
1167 ctx->cur_field ^= 1;
1168 buf += ctx->coding_unit_size;
1169 goto encode_coding_unit;
1172 #if FF_API_CODED_FRAME
1173 FF_DISABLE_DEPRECATION_WARNINGS
1174 avctx->coded_frame->quality = ctx->qscale * FF_QP2LAMBDA;
1175 FF_ENABLE_DEPRECATION_WARNINGS
1178 ff_side_data_set_encoder_stats(pkt, ctx->qscale * FF_QP2LAMBDA, NULL, 0, AV_PICTURE_TYPE_I);
1180 pkt->flags |= AV_PKT_FLAG_KEY;
1185 static av_cold int dnxhd_encode_end(AVCodecContext *avctx)
1187 DNXHDEncContext *ctx = avctx->priv_data;
1188 int max_level = 1 << (ctx->cid_table->bit_depth + 2);
1191 av_free(ctx->vlc_codes - max_level * 2);
1192 av_free(ctx->vlc_bits - max_level * 2);
1193 av_freep(&ctx->run_codes);
1194 av_freep(&ctx->run_bits);
1196 av_freep(&ctx->mb_bits);
1197 av_freep(&ctx->mb_qscale);
1198 av_freep(&ctx->mb_rc);
1199 av_freep(&ctx->mb_cmp);
1200 av_freep(&ctx->slice_size);
1201 av_freep(&ctx->slice_offs);
1203 av_freep(&ctx->qmatrix_c);
1204 av_freep(&ctx->qmatrix_l);
1205 av_freep(&ctx->qmatrix_c16);
1206 av_freep(&ctx->qmatrix_l16);
1208 for (i = 1; i < avctx->thread_count; i++)
1209 av_freep(&ctx->thread[i]);
1214 static const AVCodecDefault dnxhd_defaults[] = {
1215 { "qmax", "1024" }, /* Maximum quantization scale factor allowed for VC-3 */
1219 AVCodec ff_dnxhd_encoder = {
1221 .long_name = NULL_IF_CONFIG_SMALL("VC3/DNxHD"),
1222 .type = AVMEDIA_TYPE_VIDEO,
1223 .id = AV_CODEC_ID_DNXHD,
1224 .priv_data_size = sizeof(DNXHDEncContext),
1225 .init = dnxhd_encode_init,
1226 .encode2 = dnxhd_encode_picture,
1227 .close = dnxhd_encode_end,
1228 .capabilities = AV_CODEC_CAP_SLICE_THREADS,
1229 .pix_fmts = (const enum AVPixelFormat[]) {
1231 AV_PIX_FMT_YUV422P10,
1234 .priv_class = &dnxhd_class,
1235 .defaults = dnxhd_defaults,
1236 .profiles = NULL_IF_CONFIG_SMALL(ff_dnxhd_profiles),