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"
34 #include "mpegvideo.h"
35 #include "pixblockdsp.h"
36 #include "packet_internal.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_444(MpegEncContext *ctx, int16_t *block,
116 int n, int qscale, int *overflow)
118 int i, j, level, last_non_zero, start_i;
120 const uint8_t *scantable= ctx->intra_scantable.scantable;
123 unsigned int threshold1, threshold2;
125 ctx->fdsp.fdct(block);
127 block[0] = (block[0] + 2) >> 2;
130 qmat = n < 4 ? ctx->q_intra_matrix[qscale] : ctx->q_chroma_intra_matrix[qscale];
131 bias= ctx->intra_quant_bias * (1 << (16 - 8));
132 threshold1 = (1 << 16) - bias - 1;
133 threshold2 = (threshold1 << 1);
135 for (i = 63; i >= start_i; i--) {
137 level = block[j] * qmat[j];
139 if (((unsigned)(level + threshold1)) > threshold2) {
147 for (i = start_i; i <= last_non_zero; i++) {
149 level = block[j] * qmat[j];
151 if (((unsigned)(level + threshold1)) > threshold2) {
153 level = (bias + level) >> 16;
156 level = (bias - level) >> 16;
164 *overflow = ctx->max_qcoeff < max; //overflow might have happened
166 /* we need this permutation so that we correct the IDCT, we only permute the !=0 elements */
167 if (ctx->idsp.perm_type != FF_IDCT_PERM_NONE)
168 ff_block_permute(block, ctx->idsp.idct_permutation,
169 scantable, last_non_zero);
171 return last_non_zero;
174 static int dnxhd_10bit_dct_quantize(MpegEncContext *ctx, int16_t *block,
175 int n, int qscale, int *overflow)
177 const uint8_t *scantable= ctx->intra_scantable.scantable;
178 const int *qmat = n<4 ? ctx->q_intra_matrix[qscale] : ctx->q_chroma_intra_matrix[qscale];
179 int last_non_zero = 0;
182 ctx->fdsp.fdct(block);
184 // Divide by 4 with rounding, to compensate scaling of DCT coefficients
185 block[0] = (block[0] + 2) >> 2;
187 for (i = 1; i < 64; ++i) {
188 int j = scantable[i];
189 int sign = FF_SIGNBIT(block[j]);
190 int level = (block[j] ^ sign) - sign;
191 level = level * qmat[j] >> DNX10BIT_QMAT_SHIFT;
192 block[j] = (level ^ sign) - sign;
197 /* we need this permutation so that we correct the IDCT, we only permute the !=0 elements */
198 if (ctx->idsp.perm_type != FF_IDCT_PERM_NONE)
199 ff_block_permute(block, ctx->idsp.idct_permutation,
200 scantable, last_non_zero);
202 return last_non_zero;
205 static av_cold int dnxhd_init_vlc(DNXHDEncContext *ctx)
207 int i, j, level, run;
208 int max_level = 1 << (ctx->bit_depth + 2);
210 if (!FF_ALLOCZ_TYPED_ARRAY(ctx->orig_vlc_codes, max_level * 4) ||
211 !FF_ALLOCZ_TYPED_ARRAY(ctx->orig_vlc_bits, max_level * 4) ||
212 !(ctx->run_codes = av_mallocz(63 * 2)) ||
213 !(ctx->run_bits = av_mallocz(63)))
214 return AVERROR(ENOMEM);
215 ctx->vlc_codes = ctx->orig_vlc_codes + max_level * 2;
216 ctx->vlc_bits = ctx->orig_vlc_bits + max_level * 2;
217 for (level = -max_level; level < max_level; level++) {
218 for (run = 0; run < 2; run++) {
219 int index = level * (1 << 1) | run;
220 int sign, offset = 0, alevel = level;
222 MASK_ABS(sign, alevel);
224 offset = (alevel - 1) >> 6;
225 alevel -= offset << 6;
227 for (j = 0; j < 257; j++) {
228 if (ctx->cid_table->ac_info[2*j+0] >> 1 == alevel &&
229 (!offset || (ctx->cid_table->ac_info[2*j+1] & 1) && offset) &&
230 (!run || (ctx->cid_table->ac_info[2*j+1] & 2) && run)) {
231 av_assert1(!ctx->vlc_codes[index]);
233 ctx->vlc_codes[index] =
234 (ctx->cid_table->ac_codes[j] << 1) | (sign & 1);
235 ctx->vlc_bits[index] = ctx->cid_table->ac_bits[j] + 1;
237 ctx->vlc_codes[index] = ctx->cid_table->ac_codes[j];
238 ctx->vlc_bits[index] = ctx->cid_table->ac_bits[j];
243 av_assert0(!alevel || j < 257);
245 ctx->vlc_codes[index] =
246 (ctx->vlc_codes[index] << ctx->cid_table->index_bits) | offset;
247 ctx->vlc_bits[index] += ctx->cid_table->index_bits;
251 for (i = 0; i < 62; i++) {
252 int run = ctx->cid_table->run[i];
253 av_assert0(run < 63);
254 ctx->run_codes[run] = ctx->cid_table->run_codes[i];
255 ctx->run_bits[run] = ctx->cid_table->run_bits[i];
260 static av_cold int dnxhd_init_qmat(DNXHDEncContext *ctx, int lbias, int cbias)
262 // init first elem to 1 to avoid div by 0 in convert_matrix
263 uint16_t weight_matrix[64] = { 1, }; // convert_matrix needs uint16_t*
265 const uint8_t *luma_weight_table = ctx->cid_table->luma_weight;
266 const uint8_t *chroma_weight_table = ctx->cid_table->chroma_weight;
268 if (!FF_ALLOCZ_TYPED_ARRAY(ctx->qmatrix_l, ctx->m.avctx->qmax + 1) ||
269 !FF_ALLOCZ_TYPED_ARRAY(ctx->qmatrix_c, ctx->m.avctx->qmax + 1) ||
270 !FF_ALLOCZ_TYPED_ARRAY(ctx->qmatrix_l16, ctx->m.avctx->qmax + 1) ||
271 !FF_ALLOCZ_TYPED_ARRAY(ctx->qmatrix_c16, ctx->m.avctx->qmax + 1))
272 return AVERROR(ENOMEM);
274 if (ctx->bit_depth == 8) {
275 for (i = 1; i < 64; i++) {
276 int j = ctx->m.idsp.idct_permutation[ff_zigzag_direct[i]];
277 weight_matrix[j] = ctx->cid_table->luma_weight[i];
279 ff_convert_matrix(&ctx->m, ctx->qmatrix_l, ctx->qmatrix_l16,
280 weight_matrix, ctx->intra_quant_bias, 1,
281 ctx->m.avctx->qmax, 1);
282 for (i = 1; i < 64; i++) {
283 int j = ctx->m.idsp.idct_permutation[ff_zigzag_direct[i]];
284 weight_matrix[j] = ctx->cid_table->chroma_weight[i];
286 ff_convert_matrix(&ctx->m, ctx->qmatrix_c, ctx->qmatrix_c16,
287 weight_matrix, ctx->intra_quant_bias, 1,
288 ctx->m.avctx->qmax, 1);
290 for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
291 for (i = 0; i < 64; i++) {
292 ctx->qmatrix_l[qscale][i] <<= 2;
293 ctx->qmatrix_c[qscale][i] <<= 2;
294 ctx->qmatrix_l16[qscale][0][i] <<= 2;
295 ctx->qmatrix_l16[qscale][1][i] <<= 2;
296 ctx->qmatrix_c16[qscale][0][i] <<= 2;
297 ctx->qmatrix_c16[qscale][1][i] <<= 2;
302 for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
303 for (i = 1; i < 64; i++) {
304 int j = ff_zigzag_direct[i];
306 /* The quantization formula from the VC-3 standard is:
307 * quantized = sign(block[i]) * floor(abs(block[i]/s) * p /
308 * (qscale * weight_table[i]))
309 * Where p is 32 for 8-bit samples and 8 for 10-bit ones.
310 * The s factor compensates scaling of DCT coefficients done by
311 * the DCT routines, and therefore is not present in standard.
312 * It's 8 for 8-bit samples and 4 for 10-bit ones.
313 * We want values of ctx->qtmatrix_l and ctx->qtmatrix_r to be:
314 * ((1 << DNX10BIT_QMAT_SHIFT) * (p / s)) /
315 * (qscale * weight_table[i])
316 * For 10-bit samples, p / s == 2 */
317 ctx->qmatrix_l[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) /
318 (qscale * luma_weight_table[i]);
319 ctx->qmatrix_c[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) /
320 (qscale * chroma_weight_table[i]);
325 ctx->m.q_chroma_intra_matrix16 = ctx->qmatrix_c16;
326 ctx->m.q_chroma_intra_matrix = ctx->qmatrix_c;
327 ctx->m.q_intra_matrix16 = ctx->qmatrix_l16;
328 ctx->m.q_intra_matrix = ctx->qmatrix_l;
333 static av_cold int dnxhd_init_rc(DNXHDEncContext *ctx)
335 if (!FF_ALLOCZ_TYPED_ARRAY(ctx->mb_rc, (ctx->m.avctx->qmax + 1) * ctx->m.mb_num))
336 return AVERROR(ENOMEM);
338 if (ctx->m.avctx->mb_decision != FF_MB_DECISION_RD) {
339 if (!FF_ALLOCZ_TYPED_ARRAY(ctx->mb_cmp, ctx->m.mb_num) ||
340 !FF_ALLOCZ_TYPED_ARRAY(ctx->mb_cmp_tmp, ctx->m.mb_num))
341 return AVERROR(ENOMEM);
343 ctx->frame_bits = (ctx->coding_unit_size -
344 ctx->data_offset - 4 - ctx->min_padding) * 8;
346 ctx->lambda = 2 << LAMBDA_FRAC_BITS; // qscale 2
350 static av_cold int dnxhd_encode_init(AVCodecContext *avctx)
352 DNXHDEncContext *ctx = avctx->priv_data;
355 switch (avctx->pix_fmt) {
356 case AV_PIX_FMT_YUV422P:
359 case AV_PIX_FMT_YUV422P10:
360 case AV_PIX_FMT_YUV444P10:
361 case AV_PIX_FMT_GBRP10:
365 av_log(avctx, AV_LOG_ERROR,
366 "pixel format is incompatible with DNxHD\n");
367 return AVERROR(EINVAL);
370 if ((ctx->profile == FF_PROFILE_DNXHR_444 && (avctx->pix_fmt != AV_PIX_FMT_YUV444P10 &&
371 avctx->pix_fmt != AV_PIX_FMT_GBRP10)) ||
372 (ctx->profile != FF_PROFILE_DNXHR_444 && (avctx->pix_fmt == AV_PIX_FMT_YUV444P10 ||
373 avctx->pix_fmt == AV_PIX_FMT_GBRP10))) {
374 av_log(avctx, AV_LOG_ERROR,
375 "pixel format is incompatible with DNxHD profile\n");
376 return AVERROR(EINVAL);
379 if (ctx->profile == FF_PROFILE_DNXHR_HQX && avctx->pix_fmt != AV_PIX_FMT_YUV422P10) {
380 av_log(avctx, AV_LOG_ERROR,
381 "pixel format is incompatible with DNxHR HQX profile\n");
382 return AVERROR(EINVAL);
385 if ((ctx->profile == FF_PROFILE_DNXHR_LB ||
386 ctx->profile == FF_PROFILE_DNXHR_SQ ||
387 ctx->profile == FF_PROFILE_DNXHR_HQ) && avctx->pix_fmt != AV_PIX_FMT_YUV422P) {
388 av_log(avctx, AV_LOG_ERROR,
389 "pixel format is incompatible with DNxHR LB/SQ/HQ profile\n");
390 return AVERROR(EINVAL);
393 ctx->is_444 = ctx->profile == FF_PROFILE_DNXHR_444;
394 avctx->profile = ctx->profile;
395 ctx->cid = ff_dnxhd_find_cid(avctx, ctx->bit_depth);
397 av_log(avctx, AV_LOG_ERROR,
398 "video parameters incompatible with DNxHD. Valid DNxHD profiles:\n");
399 ff_dnxhd_print_profiles(avctx, AV_LOG_ERROR);
400 return AVERROR(EINVAL);
402 av_log(avctx, AV_LOG_DEBUG, "cid %d\n", ctx->cid);
404 if (ctx->cid >= 1270 && ctx->cid <= 1274)
405 avctx->codec_tag = MKTAG('A','V','d','h');
407 if (avctx->width < 256 || avctx->height < 120) {
408 av_log(avctx, AV_LOG_ERROR,
409 "Input dimensions too small, input must be at least 256x120\n");
410 return AVERROR(EINVAL);
413 index = ff_dnxhd_get_cid_table(ctx->cid);
414 av_assert0(index >= 0);
416 ctx->cid_table = &ff_dnxhd_cid_table[index];
418 ctx->m.avctx = avctx;
422 avctx->bits_per_raw_sample = ctx->bit_depth;
424 ff_blockdsp_init(&ctx->bdsp, avctx);
425 ff_fdctdsp_init(&ctx->m.fdsp, avctx);
426 ff_mpv_idct_init(&ctx->m);
427 ff_mpegvideoencdsp_init(&ctx->m.mpvencdsp, avctx);
428 ff_pixblockdsp_init(&ctx->m.pdsp, avctx);
429 ff_dct_encode_init(&ctx->m);
431 if (ctx->profile != FF_PROFILE_DNXHD)
432 ff_videodsp_init(&ctx->m.vdsp, ctx->bit_depth);
434 if (!ctx->m.dct_quantize)
435 ctx->m.dct_quantize = ff_dct_quantize_c;
437 if (ctx->is_444 || ctx->profile == FF_PROFILE_DNXHR_HQX) {
438 ctx->m.dct_quantize = dnxhd_10bit_dct_quantize_444;
439 ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym;
440 ctx->block_width_l2 = 4;
441 } else if (ctx->bit_depth == 10) {
442 ctx->m.dct_quantize = dnxhd_10bit_dct_quantize;
443 ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym;
444 ctx->block_width_l2 = 4;
446 ctx->get_pixels_8x4_sym = dnxhd_8bit_get_pixels_8x4_sym;
447 ctx->block_width_l2 = 3;
451 ff_dnxhdenc_init_x86(ctx);
453 ctx->m.mb_height = (avctx->height + 15) / 16;
454 ctx->m.mb_width = (avctx->width + 15) / 16;
456 if (avctx->flags & AV_CODEC_FLAG_INTERLACED_DCT) {
458 ctx->m.mb_height /= 2;
461 if (ctx->interlaced && ctx->profile != FF_PROFILE_DNXHD) {
462 av_log(avctx, AV_LOG_ERROR,
463 "Interlaced encoding is not supported for DNxHR profiles.\n");
464 return AVERROR(EINVAL);
467 ctx->m.mb_num = ctx->m.mb_height * ctx->m.mb_width;
469 if (ctx->cid_table->frame_size == DNXHD_VARIABLE) {
470 ctx->frame_size = avpriv_dnxhd_get_hr_frame_size(ctx->cid,
471 avctx->width, avctx->height);
472 av_assert0(ctx->frame_size >= 0);
473 ctx->coding_unit_size = ctx->frame_size;
475 ctx->frame_size = ctx->cid_table->frame_size;
476 ctx->coding_unit_size = ctx->cid_table->coding_unit_size;
479 if (ctx->m.mb_height > 68)
480 ctx->data_offset = 0x170 + (ctx->m.mb_height << 2);
482 ctx->data_offset = 0x280;
484 // XXX tune lbias/cbias
485 if ((ret = dnxhd_init_qmat(ctx, ctx->intra_quant_bias, 0)) < 0)
488 /* Avid Nitris hardware decoder requires a minimum amount of padding
489 * in the coding unit payload */
490 if (ctx->nitris_compat)
491 ctx->min_padding = 1600;
493 if ((ret = dnxhd_init_vlc(ctx)) < 0)
495 if ((ret = dnxhd_init_rc(ctx)) < 0)
498 if (!FF_ALLOCZ_TYPED_ARRAY(ctx->slice_size, ctx->m.mb_height) ||
499 !FF_ALLOCZ_TYPED_ARRAY(ctx->slice_offs, ctx->m.mb_height) ||
500 !FF_ALLOCZ_TYPED_ARRAY(ctx->mb_bits, ctx->m.mb_num) ||
501 !FF_ALLOCZ_TYPED_ARRAY(ctx->mb_qscale, ctx->m.mb_num))
502 return AVERROR(ENOMEM);
503 #if FF_API_CODED_FRAME
504 FF_DISABLE_DEPRECATION_WARNINGS
505 avctx->coded_frame->key_frame = 1;
506 avctx->coded_frame->pict_type = AV_PICTURE_TYPE_I;
507 FF_ENABLE_DEPRECATION_WARNINGS
510 if (avctx->active_thread_type == FF_THREAD_SLICE) {
511 if (avctx->thread_count > MAX_THREADS) {
512 av_log(avctx, AV_LOG_ERROR, "too many threads\n");
513 return AVERROR(EINVAL);
517 if (avctx->qmax <= 1) {
518 av_log(avctx, AV_LOG_ERROR, "qmax must be at least 2\n");
519 return AVERROR(EINVAL);
522 ctx->thread[0] = ctx;
523 if (avctx->active_thread_type == FF_THREAD_SLICE) {
524 for (i = 1; i < avctx->thread_count; i++) {
525 ctx->thread[i] = av_malloc(sizeof(DNXHDEncContext));
527 return AVERROR(ENOMEM);
528 memcpy(ctx->thread[i], ctx, sizeof(DNXHDEncContext));
535 static int dnxhd_write_header(AVCodecContext *avctx, uint8_t *buf)
537 DNXHDEncContext *ctx = avctx->priv_data;
539 memset(buf, 0, ctx->data_offset);
542 AV_WB16(buf + 0x02, ctx->data_offset);
543 if (ctx->cid >= 1270 && ctx->cid <= 1274)
548 buf[5] = ctx->interlaced ? ctx->cur_field + 2 : 0x01;
549 buf[6] = 0x80; // crc flag off
550 buf[7] = 0xa0; // reserved
551 AV_WB16(buf + 0x18, avctx->height >> ctx->interlaced); // ALPF
552 AV_WB16(buf + 0x1a, avctx->width); // SPL
553 AV_WB16(buf + 0x1d, avctx->height >> ctx->interlaced); // NAL
555 buf[0x21] = ctx->bit_depth == 10 ? 0x58 : 0x38;
556 buf[0x22] = 0x88 + (ctx->interlaced << 2);
557 AV_WB32(buf + 0x28, ctx->cid); // CID
558 buf[0x2c] = (!ctx->interlaced << 7) | (ctx->is_444 << 6) | (avctx->pix_fmt == AV_PIX_FMT_YUV444P10);
560 buf[0x5f] = 0x01; // UDL
562 buf[0x167] = 0x02; // reserved
563 AV_WB16(buf + 0x16a, ctx->m.mb_height * 4 + 4); // MSIPS
564 AV_WB16(buf + 0x16c, ctx->m.mb_height); // Ns
565 buf[0x16f] = 0x10; // reserved
567 ctx->msip = buf + 0x170;
571 static av_always_inline void dnxhd_encode_dc(DNXHDEncContext *ctx, int diff)
575 nbits = av_log2_16bit(-2 * diff);
578 nbits = av_log2_16bit(2 * diff);
580 put_bits(&ctx->m.pb, ctx->cid_table->dc_bits[nbits] + nbits,
581 (ctx->cid_table->dc_codes[nbits] << nbits) +
582 av_mod_uintp2(diff, nbits));
585 static av_always_inline
586 void dnxhd_encode_block(DNXHDEncContext *ctx, int16_t *block,
587 int last_index, int n)
589 int last_non_zero = 0;
592 dnxhd_encode_dc(ctx, block[0] - ctx->m.last_dc[n]);
593 ctx->m.last_dc[n] = block[0];
595 for (i = 1; i <= last_index; i++) {
596 j = ctx->m.intra_scantable.permutated[i];
599 int run_level = i - last_non_zero - 1;
600 int rlevel = slevel * (1 << 1) | !!run_level;
601 put_bits(&ctx->m.pb, ctx->vlc_bits[rlevel], ctx->vlc_codes[rlevel]);
603 put_bits(&ctx->m.pb, ctx->run_bits[run_level],
604 ctx->run_codes[run_level]);
608 put_bits(&ctx->m.pb, ctx->vlc_bits[0], ctx->vlc_codes[0]); // EOB
611 static av_always_inline
612 void dnxhd_unquantize_c(DNXHDEncContext *ctx, int16_t *block, int n,
613 int qscale, int last_index)
615 const uint8_t *weight_matrix;
620 weight_matrix = ((n % 6) < 2) ? ctx->cid_table->luma_weight
621 : ctx->cid_table->chroma_weight;
623 weight_matrix = (n & 2) ? ctx->cid_table->chroma_weight
624 : ctx->cid_table->luma_weight;
627 for (i = 1; i <= last_index; i++) {
628 int j = ctx->m.intra_scantable.permutated[i];
632 level = (1 - 2 * level) * qscale * weight_matrix[i];
633 if (ctx->bit_depth == 10) {
634 if (weight_matrix[i] != 8)
638 if (weight_matrix[i] != 32)
644 level = (2 * level + 1) * qscale * weight_matrix[i];
645 if (ctx->bit_depth == 10) {
646 if (weight_matrix[i] != 8)
650 if (weight_matrix[i] != 32)
660 static av_always_inline int dnxhd_ssd_block(int16_t *qblock, int16_t *block)
664 for (i = 0; i < 64; i++)
665 score += (block[i] - qblock[i]) * (block[i] - qblock[i]);
669 static av_always_inline
670 int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, int16_t *block, int last_index)
672 int last_non_zero = 0;
675 for (i = 1; i <= last_index; i++) {
676 j = ctx->m.intra_scantable.permutated[i];
679 int run_level = i - last_non_zero - 1;
680 bits += ctx->vlc_bits[level * (1 << 1) |
681 !!run_level] + ctx->run_bits[run_level];
688 static av_always_inline
689 void dnxhd_get_blocks(DNXHDEncContext *ctx, int mb_x, int mb_y)
691 const int bs = ctx->block_width_l2;
692 const int bw = 1 << bs;
693 int dct_y_offset = ctx->dct_y_offset;
694 int dct_uv_offset = ctx->dct_uv_offset;
695 int linesize = ctx->m.linesize;
696 int uvlinesize = ctx->m.uvlinesize;
697 const uint8_t *ptr_y = ctx->thread[0]->src[0] +
698 ((mb_y << 4) * ctx->m.linesize) + (mb_x << bs + 1);
699 const uint8_t *ptr_u = ctx->thread[0]->src[1] +
700 ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs + ctx->is_444);
701 const uint8_t *ptr_v = ctx->thread[0]->src[2] +
702 ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs + ctx->is_444);
703 PixblockDSPContext *pdsp = &ctx->m.pdsp;
704 VideoDSPContext *vdsp = &ctx->m.vdsp;
706 if (ctx->bit_depth != 10 && vdsp->emulated_edge_mc && ((mb_x << 4) + 16 > ctx->m.avctx->width ||
707 (mb_y << 4) + 16 > ctx->m.avctx->height)) {
708 int y_w = ctx->m.avctx->width - (mb_x << 4);
709 int y_h = ctx->m.avctx->height - (mb_y << 4);
710 int uv_w = (y_w + 1) / 2;
715 vdsp->emulated_edge_mc(&ctx->edge_buf_y[0], ptr_y,
716 linesize, ctx->m.linesize,
719 vdsp->emulated_edge_mc(&ctx->edge_buf_uv[0][0], ptr_u,
720 uvlinesize, ctx->m.uvlinesize,
723 vdsp->emulated_edge_mc(&ctx->edge_buf_uv[1][0], ptr_v,
724 uvlinesize, ctx->m.uvlinesize,
728 dct_y_offset = bw * linesize;
729 dct_uv_offset = bw * uvlinesize;
730 ptr_y = &ctx->edge_buf_y[0];
731 ptr_u = &ctx->edge_buf_uv[0][0];
732 ptr_v = &ctx->edge_buf_uv[1][0];
733 } else if (ctx->bit_depth == 10 && vdsp->emulated_edge_mc && ((mb_x << 4) + 16 > ctx->m.avctx->width ||
734 (mb_y << 4) + 16 > ctx->m.avctx->height)) {
735 int y_w = ctx->m.avctx->width - (mb_x << 4);
736 int y_h = ctx->m.avctx->height - (mb_y << 4);
737 int uv_w = ctx->is_444 ? y_w : (y_w + 1) / 2;
740 uvlinesize = 16 + 16 * ctx->is_444;
742 vdsp->emulated_edge_mc(&ctx->edge_buf_y[0], ptr_y,
743 linesize, ctx->m.linesize,
746 vdsp->emulated_edge_mc(&ctx->edge_buf_uv[0][0], ptr_u,
747 uvlinesize, ctx->m.uvlinesize,
750 vdsp->emulated_edge_mc(&ctx->edge_buf_uv[1][0], ptr_v,
751 uvlinesize, ctx->m.uvlinesize,
755 dct_y_offset = bw * linesize / 2;
756 dct_uv_offset = bw * uvlinesize / 2;
757 ptr_y = &ctx->edge_buf_y[0];
758 ptr_u = &ctx->edge_buf_uv[0][0];
759 ptr_v = &ctx->edge_buf_uv[1][0];
763 pdsp->get_pixels(ctx->blocks[0], ptr_y, linesize);
764 pdsp->get_pixels(ctx->blocks[1], ptr_y + bw, linesize);
765 pdsp->get_pixels(ctx->blocks[2], ptr_u, uvlinesize);
766 pdsp->get_pixels(ctx->blocks[3], ptr_v, uvlinesize);
768 if (mb_y + 1 == ctx->m.mb_height && ctx->m.avctx->height == 1080) {
769 if (ctx->interlaced) {
770 ctx->get_pixels_8x4_sym(ctx->blocks[4],
771 ptr_y + dct_y_offset,
773 ctx->get_pixels_8x4_sym(ctx->blocks[5],
774 ptr_y + dct_y_offset + bw,
776 ctx->get_pixels_8x4_sym(ctx->blocks[6],
777 ptr_u + dct_uv_offset,
779 ctx->get_pixels_8x4_sym(ctx->blocks[7],
780 ptr_v + dct_uv_offset,
783 ctx->bdsp.clear_block(ctx->blocks[4]);
784 ctx->bdsp.clear_block(ctx->blocks[5]);
785 ctx->bdsp.clear_block(ctx->blocks[6]);
786 ctx->bdsp.clear_block(ctx->blocks[7]);
789 pdsp->get_pixels(ctx->blocks[4],
790 ptr_y + dct_y_offset, linesize);
791 pdsp->get_pixels(ctx->blocks[5],
792 ptr_y + dct_y_offset + bw, linesize);
793 pdsp->get_pixels(ctx->blocks[6],
794 ptr_u + dct_uv_offset, uvlinesize);
795 pdsp->get_pixels(ctx->blocks[7],
796 ptr_v + dct_uv_offset, uvlinesize);
799 pdsp->get_pixels(ctx->blocks[0], ptr_y, linesize);
800 pdsp->get_pixels(ctx->blocks[1], ptr_y + bw, linesize);
801 pdsp->get_pixels(ctx->blocks[6], ptr_y + dct_y_offset, linesize);
802 pdsp->get_pixels(ctx->blocks[7], ptr_y + dct_y_offset + bw, linesize);
804 pdsp->get_pixels(ctx->blocks[2], ptr_u, uvlinesize);
805 pdsp->get_pixels(ctx->blocks[3], ptr_u + bw, uvlinesize);
806 pdsp->get_pixels(ctx->blocks[8], ptr_u + dct_uv_offset, uvlinesize);
807 pdsp->get_pixels(ctx->blocks[9], ptr_u + dct_uv_offset + bw, uvlinesize);
809 pdsp->get_pixels(ctx->blocks[4], ptr_v, uvlinesize);
810 pdsp->get_pixels(ctx->blocks[5], ptr_v + bw, uvlinesize);
811 pdsp->get_pixels(ctx->blocks[10], ptr_v + dct_uv_offset, uvlinesize);
812 pdsp->get_pixels(ctx->blocks[11], ptr_v + dct_uv_offset + bw, uvlinesize);
816 static av_always_inline
817 int dnxhd_switch_matrix(DNXHDEncContext *ctx, int i)
824 const static uint8_t component[8]={0,0,1,2,0,0,1,2};
830 static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg,
831 int jobnr, int threadnr)
833 DNXHDEncContext *ctx = avctx->priv_data;
834 int mb_y = jobnr, mb_x;
835 int qscale = ctx->qscale;
836 LOCAL_ALIGNED_16(int16_t, block, [64]);
837 ctx = ctx->thread[threadnr];
841 ctx->m.last_dc[2] = 1 << (ctx->bit_depth + 2);
843 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
844 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
850 dnxhd_get_blocks(ctx, mb_x, mb_y);
852 for (i = 0; i < 8 + 4 * ctx->is_444; i++) {
853 int16_t *src_block = ctx->blocks[i];
854 int overflow, nbits, diff, last_index;
855 int n = dnxhd_switch_matrix(ctx, i);
857 memcpy(block, src_block, 64 * sizeof(*block));
858 last_index = ctx->m.dct_quantize(&ctx->m, block,
859 ctx->is_444 ? 4 * (n > 0): 4 & (2*i),
861 ac_bits += dnxhd_calc_ac_bits(ctx, block, last_index);
863 diff = block[0] - ctx->m.last_dc[n];
865 nbits = av_log2_16bit(-2 * diff);
867 nbits = av_log2_16bit(2 * diff);
869 av_assert1(nbits < ctx->bit_depth + 4);
870 dc_bits += ctx->cid_table->dc_bits[nbits] + nbits;
872 ctx->m.last_dc[n] = block[0];
874 if (avctx->mb_decision == FF_MB_DECISION_RD || !RC_VARIANCE) {
875 dnxhd_unquantize_c(ctx, block, i, qscale, last_index);
876 ctx->m.idsp.idct(block);
877 ssd += dnxhd_ssd_block(block, src_block);
880 ctx->mb_rc[(qscale * ctx->m.mb_num) + mb].ssd = ssd;
881 ctx->mb_rc[(qscale * ctx->m.mb_num) + mb].bits = ac_bits + dc_bits + 12 +
882 (1 + ctx->is_444) * 8 * ctx->vlc_bits[0];
887 static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg,
888 int jobnr, int threadnr)
890 DNXHDEncContext *ctx = avctx->priv_data;
891 int mb_y = jobnr, mb_x;
892 ctx = ctx->thread[threadnr];
893 init_put_bits(&ctx->m.pb, (uint8_t *)arg + ctx->data_offset + ctx->slice_offs[jobnr],
894 ctx->slice_size[jobnr]);
898 ctx->m.last_dc[2] = 1 << (ctx->bit_depth + 2);
899 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
900 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
901 int qscale = ctx->mb_qscale[mb];
904 put_bits(&ctx->m.pb, 11, qscale);
905 put_bits(&ctx->m.pb, 1, avctx->pix_fmt == AV_PIX_FMT_YUV444P10);
907 dnxhd_get_blocks(ctx, mb_x, mb_y);
909 for (i = 0; i < 8 + 4 * ctx->is_444; i++) {
910 int16_t *block = ctx->blocks[i];
911 int overflow, n = dnxhd_switch_matrix(ctx, i);
912 int last_index = ctx->m.dct_quantize(&ctx->m, block,
913 ctx->is_444 ? (((i >> 1) % 3) < 1 ? 0 : 4): 4 & (2*i),
916 dnxhd_encode_block(ctx, block, last_index, n);
919 if (put_bits_count(&ctx->m.pb) & 31)
920 put_bits(&ctx->m.pb, 32 - (put_bits_count(&ctx->m.pb) & 31), 0);
921 flush_put_bits(&ctx->m.pb);
925 static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx)
929 for (mb_y = 0; mb_y < ctx->m.mb_height; mb_y++) {
931 ctx->slice_offs[mb_y] = offset;
932 ctx->slice_size[mb_y] = 0;
933 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
934 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
935 ctx->slice_size[mb_y] += ctx->mb_bits[mb];
937 ctx->slice_size[mb_y] = (ctx->slice_size[mb_y] + 31) & ~31;
938 ctx->slice_size[mb_y] >>= 3;
939 thread_size = ctx->slice_size[mb_y];
940 offset += thread_size;
944 static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg,
945 int jobnr, int threadnr)
947 DNXHDEncContext *ctx = avctx->priv_data;
948 int mb_y = jobnr, mb_x, x, y;
949 int partial_last_row = (mb_y == ctx->m.mb_height - 1) &&
950 ((avctx->height >> ctx->interlaced) & 0xF);
952 ctx = ctx->thread[threadnr];
953 if (ctx->bit_depth == 8) {
954 uint8_t *pix = ctx->thread[0]->src[0] + ((mb_y << 4) * ctx->m.linesize);
955 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x, pix += 16) {
956 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
960 if (!partial_last_row && mb_x * 16 <= avctx->width - 16 && (avctx->width % 16) == 0) {
961 sum = ctx->m.mpvencdsp.pix_sum(pix, ctx->m.linesize);
962 varc = ctx->m.mpvencdsp.pix_norm1(pix, ctx->m.linesize);
964 int bw = FFMIN(avctx->width - 16 * mb_x, 16);
965 int bh = FFMIN((avctx->height >> ctx->interlaced) - 16 * mb_y, 16);
967 for (y = 0; y < bh; y++) {
968 for (x = 0; x < bw; x++) {
969 uint8_t val = pix[x + y * ctx->m.linesize];
975 varc = (varc - (((unsigned) sum * sum) >> 8) + 128) >> 8;
977 ctx->mb_cmp[mb].value = varc;
978 ctx->mb_cmp[mb].mb = mb;
981 const int linesize = ctx->m.linesize >> 1;
982 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x) {
983 uint16_t *pix = (uint16_t *)ctx->thread[0]->src[0] +
984 ((mb_y << 4) * linesize) + (mb_x << 4);
985 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
988 int bw = FFMIN(avctx->width - 16 * mb_x, 16);
989 int bh = FFMIN((avctx->height >> ctx->interlaced) - 16 * mb_y, 16);
992 // Macroblocks are 16x16 pixels, unlike DCT blocks which are 8x8.
993 for (i = 0; i < bh; ++i) {
994 for (j = 0; j < bw; ++j) {
995 // Turn 16-bit pixels into 10-bit ones.
996 const int sample = (unsigned) pix[j] >> 6;
998 sqsum += sample * sample;
999 // 2^10 * 2^10 * 16 * 16 = 2^28, which is less than INT_MAX
1003 mean = sum >> 8; // 16*16 == 2^8
1004 sqmean = sqsum >> 8;
1005 ctx->mb_cmp[mb].value = sqmean - mean * mean;
1006 ctx->mb_cmp[mb].mb = mb;
1012 static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx)
1014 int lambda, up_step, down_step;
1015 int last_lower = INT_MAX, last_higher = 0;
1018 for (q = 1; q < avctx->qmax; q++) {
1020 avctx->execute2(avctx, dnxhd_calc_bits_thread,
1021 NULL, NULL, ctx->m.mb_height);
1023 up_step = down_step = 2 << LAMBDA_FRAC_BITS;
1024 lambda = ctx->lambda;
1029 if (lambda == last_higher) {
1031 end = 1; // need to set final qscales/bits
1033 for (y = 0; y < ctx->m.mb_height; y++) {
1034 for (x = 0; x < ctx->m.mb_width; x++) {
1035 unsigned min = UINT_MAX;
1037 int mb = y * ctx->m.mb_width + x;
1039 for (q = 1; q < avctx->qmax; q++) {
1040 int i = (q*ctx->m.mb_num) + mb;
1041 unsigned score = ctx->mb_rc[i].bits * lambda +
1042 ((unsigned) ctx->mb_rc[i].ssd << LAMBDA_FRAC_BITS);
1049 bits += ctx->mb_rc[rc].bits;
1050 ctx->mb_qscale[mb] = qscale;
1051 ctx->mb_bits[mb] = ctx->mb_rc[rc].bits;
1053 bits = (bits + 31) & ~31; // padding
1054 if (bits > ctx->frame_bits)
1058 if (bits > ctx->frame_bits)
1059 return AVERROR(EINVAL);
1062 if (bits < ctx->frame_bits) {
1063 last_lower = FFMIN(lambda, last_lower);
1064 if (last_higher != 0)
1065 lambda = (lambda+last_higher)>>1;
1067 lambda -= down_step;
1068 down_step = FFMIN((int64_t)down_step*5, INT_MAX);
1069 up_step = 1<<LAMBDA_FRAC_BITS;
1070 lambda = FFMAX(1, lambda);
1071 if (lambda == last_lower)
1074 last_higher = FFMAX(lambda, last_higher);
1075 if (last_lower != INT_MAX)
1076 lambda = (lambda+last_lower)>>1;
1077 else if ((int64_t)lambda + up_step > INT_MAX)
1078 return AVERROR(EINVAL);
1081 up_step = FFMIN((int64_t)up_step*5, INT_MAX);
1082 down_step = 1<<LAMBDA_FRAC_BITS;
1085 ctx->lambda = lambda;
1089 static int dnxhd_find_qscale(DNXHDEncContext *ctx)
1094 int last_higher = 0;
1095 int last_lower = INT_MAX;
1099 qscale = ctx->qscale;
1102 ctx->qscale = qscale;
1103 // XXX avoid recalculating bits
1104 ctx->m.avctx->execute2(ctx->m.avctx, dnxhd_calc_bits_thread,
1105 NULL, NULL, ctx->m.mb_height);
1106 for (y = 0; y < ctx->m.mb_height; y++) {
1107 for (x = 0; x < ctx->m.mb_width; x++)
1108 bits += ctx->mb_rc[(qscale*ctx->m.mb_num) + (y*ctx->m.mb_width+x)].bits;
1109 bits = (bits+31)&~31; // padding
1110 if (bits > ctx->frame_bits)
1113 if (bits < ctx->frame_bits) {
1116 if (last_higher == qscale - 1) {
1117 qscale = last_higher;
1120 last_lower = FFMIN(qscale, last_lower);
1121 if (last_higher != 0)
1122 qscale = (qscale + last_higher) >> 1;
1124 qscale -= down_step++;
1129 if (last_lower == qscale + 1)
1131 last_higher = FFMAX(qscale, last_higher);
1132 if (last_lower != INT_MAX)
1133 qscale = (qscale + last_lower) >> 1;
1135 qscale += up_step++;
1137 if (qscale >= ctx->m.avctx->qmax)
1138 return AVERROR(EINVAL);
1141 ctx->qscale = qscale;
1145 #define BUCKET_BITS 8
1146 #define RADIX_PASSES 4
1147 #define NBUCKETS (1 << BUCKET_BITS)
1149 static inline int get_bucket(int value, int shift)
1152 value &= NBUCKETS - 1;
1153 return NBUCKETS - 1 - value;
1156 static void radix_count(const RCCMPEntry *data, int size,
1157 int buckets[RADIX_PASSES][NBUCKETS])
1160 memset(buckets, 0, sizeof(buckets[0][0]) * RADIX_PASSES * NBUCKETS);
1161 for (i = 0; i < size; i++) {
1162 int v = data[i].value;
1163 for (j = 0; j < RADIX_PASSES; j++) {
1164 buckets[j][get_bucket(v, 0)]++;
1169 for (j = 0; j < RADIX_PASSES; j++) {
1171 for (i = NBUCKETS - 1; i >= 0; i--)
1172 buckets[j][i] = offset -= buckets[j][i];
1173 av_assert1(!buckets[j][0]);
1177 static void radix_sort_pass(RCCMPEntry *dst, const RCCMPEntry *data,
1178 int size, int buckets[NBUCKETS], int pass)
1180 int shift = pass * BUCKET_BITS;
1182 for (i = 0; i < size; i++) {
1183 int v = get_bucket(data[i].value, shift);
1184 int pos = buckets[v]++;
1189 static void radix_sort(RCCMPEntry *data, RCCMPEntry *tmp, int size)
1191 int buckets[RADIX_PASSES][NBUCKETS];
1192 radix_count(data, size, buckets);
1193 radix_sort_pass(tmp, data, size, buckets[0], 0);
1194 radix_sort_pass(data, tmp, size, buckets[1], 1);
1195 if (buckets[2][NBUCKETS - 1] || buckets[3][NBUCKETS - 1]) {
1196 radix_sort_pass(tmp, data, size, buckets[2], 2);
1197 radix_sort_pass(data, tmp, size, buckets[3], 3);
1201 static int dnxhd_encode_fast(AVCodecContext *avctx, DNXHDEncContext *ctx)
1205 if ((ret = dnxhd_find_qscale(ctx)) < 0)
1207 for (y = 0; y < ctx->m.mb_height; y++) {
1208 for (x = 0; x < ctx->m.mb_width; x++) {
1209 int mb = y * ctx->m.mb_width + x;
1210 int rc = (ctx->qscale * ctx->m.mb_num ) + mb;
1212 ctx->mb_qscale[mb] = ctx->qscale;
1213 ctx->mb_bits[mb] = ctx->mb_rc[rc].bits;
1214 max_bits += ctx->mb_rc[rc].bits;
1216 delta_bits = ctx->mb_rc[rc].bits -
1217 ctx->mb_rc[rc + ctx->m.mb_num].bits;
1218 ctx->mb_cmp[mb].mb = mb;
1219 ctx->mb_cmp[mb].value =
1220 delta_bits ? ((ctx->mb_rc[rc].ssd -
1221 ctx->mb_rc[rc + ctx->m.mb_num].ssd) * 100) /
1223 : INT_MIN; // avoid increasing qscale
1226 max_bits += 31; // worst padding
1230 avctx->execute2(avctx, dnxhd_mb_var_thread,
1231 NULL, NULL, ctx->m.mb_height);
1232 radix_sort(ctx->mb_cmp, ctx->mb_cmp_tmp, ctx->m.mb_num);
1233 for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) {
1234 int mb = ctx->mb_cmp[x].mb;
1235 int rc = (ctx->qscale * ctx->m.mb_num ) + mb;
1236 max_bits -= ctx->mb_rc[rc].bits -
1237 ctx->mb_rc[rc + ctx->m.mb_num].bits;
1238 ctx->mb_qscale[mb] = ctx->qscale + 1;
1239 ctx->mb_bits[mb] = ctx->mb_rc[rc + ctx->m.mb_num].bits;
1245 static void dnxhd_load_picture(DNXHDEncContext *ctx, const AVFrame *frame)
1249 for (i = 0; i < ctx->m.avctx->thread_count; i++) {
1250 ctx->thread[i]->m.linesize = frame->linesize[0] << ctx->interlaced;
1251 ctx->thread[i]->m.uvlinesize = frame->linesize[1] << ctx->interlaced;
1252 ctx->thread[i]->dct_y_offset = ctx->m.linesize *8;
1253 ctx->thread[i]->dct_uv_offset = ctx->m.uvlinesize*8;
1256 #if FF_API_CODED_FRAME
1257 FF_DISABLE_DEPRECATION_WARNINGS
1258 ctx->m.avctx->coded_frame->interlaced_frame = frame->interlaced_frame;
1259 FF_ENABLE_DEPRECATION_WARNINGS
1261 ctx->cur_field = frame->interlaced_frame && !frame->top_field_first;
1264 static int dnxhd_encode_picture(AVCodecContext *avctx, AVPacket *pkt,
1265 const AVFrame *frame, int *got_packet)
1267 DNXHDEncContext *ctx = avctx->priv_data;
1268 int first_field = 1;
1272 if ((ret = ff_alloc_packet2(avctx, pkt, ctx->frame_size, 0)) < 0)
1276 dnxhd_load_picture(ctx, frame);
1279 for (i = 0; i < 3; i++) {
1280 ctx->src[i] = frame->data[i];
1281 if (ctx->interlaced && ctx->cur_field)
1282 ctx->src[i] += frame->linesize[i];
1285 dnxhd_write_header(avctx, buf);
1287 if (avctx->mb_decision == FF_MB_DECISION_RD)
1288 ret = dnxhd_encode_rdo(avctx, ctx);
1290 ret = dnxhd_encode_fast(avctx, ctx);
1292 av_log(avctx, AV_LOG_ERROR,
1293 "picture could not fit ratecontrol constraints, increase qmax\n");
1297 dnxhd_setup_threads_slices(ctx);
1300 for (i = 0; i < ctx->m.mb_height; i++) {
1301 AV_WB32(ctx->msip + i * 4, offset);
1302 offset += ctx->slice_size[i];
1303 av_assert1(!(ctx->slice_size[i] & 3));
1306 avctx->execute2(avctx, dnxhd_encode_thread, buf, NULL, ctx->m.mb_height);
1308 av_assert1(ctx->data_offset + offset + 4 <= ctx->coding_unit_size);
1309 memset(buf + ctx->data_offset + offset, 0,
1310 ctx->coding_unit_size - 4 - offset - ctx->data_offset);
1312 AV_WB32(buf + ctx->coding_unit_size - 4, 0x600DC0DE); // EOF
1314 if (ctx->interlaced && first_field) {
1316 ctx->cur_field ^= 1;
1317 buf += ctx->coding_unit_size;
1318 goto encode_coding_unit;
1321 #if FF_API_CODED_FRAME
1322 FF_DISABLE_DEPRECATION_WARNINGS
1323 avctx->coded_frame->quality = ctx->qscale * FF_QP2LAMBDA;
1324 FF_ENABLE_DEPRECATION_WARNINGS
1327 ff_side_data_set_encoder_stats(pkt, ctx->qscale * FF_QP2LAMBDA, NULL, 0, AV_PICTURE_TYPE_I);
1329 pkt->flags |= AV_PKT_FLAG_KEY;
1334 static av_cold int dnxhd_encode_end(AVCodecContext *avctx)
1336 DNXHDEncContext *ctx = avctx->priv_data;
1339 av_freep(&ctx->orig_vlc_codes);
1340 av_freep(&ctx->orig_vlc_bits);
1341 av_freep(&ctx->run_codes);
1342 av_freep(&ctx->run_bits);
1344 av_freep(&ctx->mb_bits);
1345 av_freep(&ctx->mb_qscale);
1346 av_freep(&ctx->mb_rc);
1347 av_freep(&ctx->mb_cmp);
1348 av_freep(&ctx->mb_cmp_tmp);
1349 av_freep(&ctx->slice_size);
1350 av_freep(&ctx->slice_offs);
1352 av_freep(&ctx->qmatrix_c);
1353 av_freep(&ctx->qmatrix_l);
1354 av_freep(&ctx->qmatrix_c16);
1355 av_freep(&ctx->qmatrix_l16);
1357 if (avctx->active_thread_type == FF_THREAD_SLICE) {
1358 for (i = 1; i < avctx->thread_count; i++)
1359 av_freep(&ctx->thread[i]);
1365 static const AVCodecDefault dnxhd_defaults[] = {
1366 { "qmax", "1024" }, /* Maximum quantization scale factor allowed for VC-3 */
1370 AVCodec ff_dnxhd_encoder = {
1372 .long_name = NULL_IF_CONFIG_SMALL("VC3/DNxHD"),
1373 .type = AVMEDIA_TYPE_VIDEO,
1374 .id = AV_CODEC_ID_DNXHD,
1375 .priv_data_size = sizeof(DNXHDEncContext),
1376 .init = dnxhd_encode_init,
1377 .encode2 = dnxhd_encode_picture,
1378 .close = dnxhd_encode_end,
1379 .capabilities = AV_CODEC_CAP_SLICE_THREADS | AV_CODEC_CAP_FRAME_THREADS,
1380 .caps_internal = FF_CODEC_CAP_INIT_CLEANUP,
1381 .pix_fmts = (const enum AVPixelFormat[]) {
1383 AV_PIX_FMT_YUV422P10,
1384 AV_PIX_FMT_YUV444P10,
1388 .priv_class = &dnxhd_class,
1389 .defaults = dnxhd_defaults,
1390 .profiles = NULL_IF_CONFIG_SMALL(ff_dnxhd_profiles),