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"
39 // The largest value that will not lead to overflow for 10-bit 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_BOOL, { .i64 = 0 }, 0, 1, VE },
48 { "ibias", "intra quant bias",
49 offsetof(DNXHDEncContext, intra_quant_bias), AV_OPT_TYPE_INT,
50 { .i64 = 0 }, INT_MIN, INT_MAX, VE },
51 { "profile", NULL, offsetof(DNXHDEncContext, profile), AV_OPT_TYPE_INT,
52 { .i64 = FF_PROFILE_DNXHD },
53 FF_PROFILE_DNXHD, FF_PROFILE_DNXHR_444, VE, "profile" },
54 { "dnxhd", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_PROFILE_DNXHD },
55 0, 0, VE, "profile" },
56 { "dnxhr_444", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_PROFILE_DNXHR_444 },
57 0, 0, VE, "profile" },
58 { "dnxhr_hqx", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_PROFILE_DNXHR_HQX },
59 0, 0, VE, "profile" },
60 { "dnxhr_hq", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_PROFILE_DNXHR_HQ },
61 0, 0, VE, "profile" },
62 { "dnxhr_sq", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_PROFILE_DNXHR_SQ },
63 0, 0, VE, "profile" },
64 { "dnxhr_lb", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_PROFILE_DNXHR_LB },
65 0, 0, VE, "profile" },
69 static const AVClass dnxhd_class = {
70 .class_name = "dnxhd",
71 .item_name = av_default_item_name,
73 .version = LIBAVUTIL_VERSION_INT,
76 static void dnxhd_8bit_get_pixels_8x4_sym(int16_t *av_restrict block,
77 const uint8_t *pixels,
81 for (i = 0; i < 4; i++) {
93 memcpy(block, block - 8, sizeof(*block) * 8);
94 memcpy(block + 8, block - 16, sizeof(*block) * 8);
95 memcpy(block + 16, block - 24, sizeof(*block) * 8);
96 memcpy(block + 24, block - 32, sizeof(*block) * 8);
99 static av_always_inline
100 void dnxhd_10bit_get_pixels_8x4_sym(int16_t *av_restrict block,
101 const uint8_t *pixels,
104 memcpy(block + 0 * 8, pixels + 0 * line_size, 8 * sizeof(*block));
105 memcpy(block + 7 * 8, pixels + 0 * line_size, 8 * sizeof(*block));
106 memcpy(block + 1 * 8, pixels + 1 * line_size, 8 * sizeof(*block));
107 memcpy(block + 6 * 8, pixels + 1 * line_size, 8 * sizeof(*block));
108 memcpy(block + 2 * 8, pixels + 2 * line_size, 8 * sizeof(*block));
109 memcpy(block + 5 * 8, pixels + 2 * line_size, 8 * sizeof(*block));
110 memcpy(block + 3 * 8, pixels + 3 * line_size, 8 * sizeof(*block));
111 memcpy(block + 4 * 8, pixels + 3 * line_size, 8 * sizeof(*block));
114 static int dnxhd_10bit_dct_quantize_444(MpegEncContext *ctx, int16_t *block,
115 int n, int qscale, int *overflow)
117 int i, j, level, last_non_zero, start_i;
119 const uint8_t *scantable= ctx->intra_scantable.scantable;
122 unsigned int threshold1, threshold2;
124 ctx->fdsp.fdct(block);
126 block[0] = (block[0] + 2) >> 2;
129 qmat = n < 4 ? ctx->q_intra_matrix[qscale] : ctx->q_chroma_intra_matrix[qscale];
130 bias= ctx->intra_quant_bias * (1 << (16 - 8));
131 threshold1 = (1 << 16) - bias - 1;
132 threshold2 = (threshold1 << 1);
134 for (i = 63; i >= start_i; i--) {
136 level = block[j] * qmat[j];
138 if (((unsigned)(level + threshold1)) > threshold2) {
146 for (i = start_i; i <= last_non_zero; i++) {
148 level = block[j] * qmat[j];
150 if (((unsigned)(level + threshold1)) > threshold2) {
152 level = (bias + level) >> 16;
155 level = (bias - level) >> 16;
163 *overflow = ctx->max_qcoeff < max; //overflow might have happened
165 /* we need this permutation so that we correct the IDCT, we only permute the !=0 elements */
166 if (ctx->idsp.perm_type != FF_IDCT_PERM_NONE)
167 ff_block_permute(block, ctx->idsp.idct_permutation,
168 scantable, last_non_zero);
170 return last_non_zero;
173 static int dnxhd_10bit_dct_quantize(MpegEncContext *ctx, int16_t *block,
174 int n, int qscale, int *overflow)
176 const uint8_t *scantable= ctx->intra_scantable.scantable;
177 const int *qmat = n<4 ? ctx->q_intra_matrix[qscale] : ctx->q_chroma_intra_matrix[qscale];
178 int last_non_zero = 0;
181 ctx->fdsp.fdct(block);
183 // Divide by 4 with rounding, to compensate scaling of DCT coefficients
184 block[0] = (block[0] + 2) >> 2;
186 for (i = 1; i < 64; ++i) {
187 int j = scantable[i];
188 int sign = FF_SIGNBIT(block[j]);
189 int level = (block[j] ^ sign) - sign;
190 level = level * qmat[j] >> DNX10BIT_QMAT_SHIFT;
191 block[j] = (level ^ sign) - sign;
196 /* we need this permutation so that we correct the IDCT, we only permute the !=0 elements */
197 if (ctx->idsp.perm_type != FF_IDCT_PERM_NONE)
198 ff_block_permute(block, ctx->idsp.idct_permutation,
199 scantable, last_non_zero);
201 return last_non_zero;
204 static av_cold int dnxhd_init_vlc(DNXHDEncContext *ctx)
206 int i, j, level, run;
207 int max_level = 1 << (ctx->bit_depth + 2);
209 FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->orig_vlc_codes,
210 max_level, 4 * sizeof(*ctx->orig_vlc_codes), fail);
211 FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->orig_vlc_bits,
212 max_level, 4 * sizeof(*ctx->orig_vlc_bits), fail);
213 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_codes,
215 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_bits,
218 ctx->vlc_codes = ctx->orig_vlc_codes + max_level * 2;
219 ctx->vlc_bits = ctx->orig_vlc_bits + max_level * 2;
220 for (level = -max_level; level < max_level; level++) {
221 for (run = 0; run < 2; run++) {
222 int index = level * (1 << 1) | run;
223 int sign, offset = 0, alevel = level;
225 MASK_ABS(sign, alevel);
227 offset = (alevel - 1) >> 6;
228 alevel -= offset << 6;
230 for (j = 0; j < 257; j++) {
231 if (ctx->cid_table->ac_info[2*j+0] >> 1 == alevel &&
232 (!offset || (ctx->cid_table->ac_info[2*j+1] & 1) && offset) &&
233 (!run || (ctx->cid_table->ac_info[2*j+1] & 2) && run)) {
234 av_assert1(!ctx->vlc_codes[index]);
236 ctx->vlc_codes[index] =
237 (ctx->cid_table->ac_codes[j] << 1) | (sign & 1);
238 ctx->vlc_bits[index] = ctx->cid_table->ac_bits[j] + 1;
240 ctx->vlc_codes[index] = ctx->cid_table->ac_codes[j];
241 ctx->vlc_bits[index] = ctx->cid_table->ac_bits[j];
246 av_assert0(!alevel || j < 257);
248 ctx->vlc_codes[index] =
249 (ctx->vlc_codes[index] << ctx->cid_table->index_bits) | offset;
250 ctx->vlc_bits[index] += ctx->cid_table->index_bits;
254 for (i = 0; i < 62; i++) {
255 int run = ctx->cid_table->run[i];
256 av_assert0(run < 63);
257 ctx->run_codes[run] = ctx->cid_table->run_codes[i];
258 ctx->run_bits[run] = ctx->cid_table->run_bits[i];
262 return AVERROR(ENOMEM);
265 static av_cold int dnxhd_init_qmat(DNXHDEncContext *ctx, int lbias, int cbias)
267 // init first elem to 1 to avoid div by 0 in convert_matrix
268 uint16_t weight_matrix[64] = { 1, }; // convert_matrix needs uint16_t*
270 const uint8_t *luma_weight_table = ctx->cid_table->luma_weight;
271 const uint8_t *chroma_weight_table = ctx->cid_table->chroma_weight;
273 FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l,
274 (ctx->m.avctx->qmax + 1), 64 * sizeof(int), fail);
275 FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c,
276 (ctx->m.avctx->qmax + 1), 64 * sizeof(int), fail);
277 FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l16,
278 (ctx->m.avctx->qmax + 1), 64 * 2 * sizeof(uint16_t),
280 FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c16,
281 (ctx->m.avctx->qmax + 1), 64 * 2 * sizeof(uint16_t),
284 if (ctx->bit_depth == 8) {
285 for (i = 1; i < 64; i++) {
286 int j = ctx->m.idsp.idct_permutation[ff_zigzag_direct[i]];
287 weight_matrix[j] = ctx->cid_table->luma_weight[i];
289 ff_convert_matrix(&ctx->m, ctx->qmatrix_l, ctx->qmatrix_l16,
290 weight_matrix, ctx->intra_quant_bias, 1,
291 ctx->m.avctx->qmax, 1);
292 for (i = 1; i < 64; i++) {
293 int j = ctx->m.idsp.idct_permutation[ff_zigzag_direct[i]];
294 weight_matrix[j] = ctx->cid_table->chroma_weight[i];
296 ff_convert_matrix(&ctx->m, ctx->qmatrix_c, ctx->qmatrix_c16,
297 weight_matrix, ctx->intra_quant_bias, 1,
298 ctx->m.avctx->qmax, 1);
300 for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
301 for (i = 0; i < 64; i++) {
302 ctx->qmatrix_l[qscale][i] <<= 2;
303 ctx->qmatrix_c[qscale][i] <<= 2;
304 ctx->qmatrix_l16[qscale][0][i] <<= 2;
305 ctx->qmatrix_l16[qscale][1][i] <<= 2;
306 ctx->qmatrix_c16[qscale][0][i] <<= 2;
307 ctx->qmatrix_c16[qscale][1][i] <<= 2;
312 for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
313 for (i = 1; i < 64; i++) {
314 int j = ff_zigzag_direct[i];
316 /* The quantization formula from the VC-3 standard is:
317 * quantized = sign(block[i]) * floor(abs(block[i]/s) * p /
318 * (qscale * weight_table[i]))
319 * Where p is 32 for 8-bit samples and 8 for 10-bit ones.
320 * The s factor compensates scaling of DCT coefficients done by
321 * the DCT routines, and therefore is not present in standard.
322 * It's 8 for 8-bit samples and 4 for 10-bit ones.
323 * We want values of ctx->qtmatrix_l and ctx->qtmatrix_r to be:
324 * ((1 << DNX10BIT_QMAT_SHIFT) * (p / s)) /
325 * (qscale * weight_table[i])
326 * For 10-bit samples, p / s == 2 */
327 ctx->qmatrix_l[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) /
328 (qscale * luma_weight_table[i]);
329 ctx->qmatrix_c[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) /
330 (qscale * chroma_weight_table[i]);
335 ctx->m.q_chroma_intra_matrix16 = ctx->qmatrix_c16;
336 ctx->m.q_chroma_intra_matrix = ctx->qmatrix_c;
337 ctx->m.q_intra_matrix16 = ctx->qmatrix_l16;
338 ctx->m.q_intra_matrix = ctx->qmatrix_l;
342 return AVERROR(ENOMEM);
345 static av_cold int dnxhd_init_rc(DNXHDEncContext *ctx)
347 FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->mb_rc, (ctx->m.avctx->qmax + 1),
348 ctx->m.mb_num * sizeof(RCEntry), fail);
349 if (ctx->m.avctx->mb_decision != FF_MB_DECISION_RD) {
350 FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->mb_cmp,
351 ctx->m.mb_num, sizeof(RCCMPEntry), fail);
352 FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->mb_cmp_tmp,
353 ctx->m.mb_num, sizeof(RCCMPEntry), fail);
355 ctx->frame_bits = (ctx->coding_unit_size -
356 ctx->data_offset - 4 - ctx->min_padding) * 8;
358 ctx->lambda = 2 << LAMBDA_FRAC_BITS; // qscale 2
361 return AVERROR(ENOMEM);
364 static av_cold int dnxhd_encode_init(AVCodecContext *avctx)
366 DNXHDEncContext *ctx = avctx->priv_data;
369 switch (avctx->pix_fmt) {
370 case AV_PIX_FMT_YUV422P:
373 case AV_PIX_FMT_YUV422P10:
374 case AV_PIX_FMT_YUV444P10:
375 case AV_PIX_FMT_GBRP10:
379 av_log(avctx, AV_LOG_ERROR,
380 "pixel format is incompatible with DNxHD\n");
381 return AVERROR(EINVAL);
384 if ((ctx->profile == FF_PROFILE_DNXHR_444 && (avctx->pix_fmt != AV_PIX_FMT_YUV444P10 &&
385 avctx->pix_fmt != AV_PIX_FMT_GBRP10)) ||
386 (ctx->profile != FF_PROFILE_DNXHR_444 && (avctx->pix_fmt == AV_PIX_FMT_YUV444P10 ||
387 avctx->pix_fmt == AV_PIX_FMT_GBRP10))) {
388 av_log(avctx, AV_LOG_ERROR,
389 "pixel format is incompatible with DNxHD profile\n");
390 return AVERROR(EINVAL);
393 if (ctx->profile == FF_PROFILE_DNXHR_HQX && avctx->pix_fmt != AV_PIX_FMT_YUV422P10) {
394 av_log(avctx, AV_LOG_ERROR,
395 "pixel format is incompatible with DNxHR HQX profile\n");
396 return AVERROR(EINVAL);
399 if ((ctx->profile == FF_PROFILE_DNXHR_LB ||
400 ctx->profile == FF_PROFILE_DNXHR_SQ ||
401 ctx->profile == FF_PROFILE_DNXHR_HQ) && avctx->pix_fmt != AV_PIX_FMT_YUV422P) {
402 av_log(avctx, AV_LOG_ERROR,
403 "pixel format is incompatible with DNxHR LB/SQ/HQ profile\n");
404 return AVERROR(EINVAL);
407 ctx->is_444 = ctx->profile == FF_PROFILE_DNXHR_444;
408 avctx->profile = ctx->profile;
409 ctx->cid = ff_dnxhd_find_cid(avctx, ctx->bit_depth);
411 av_log(avctx, AV_LOG_ERROR,
412 "video parameters incompatible with DNxHD. Valid DNxHD profiles:\n");
413 ff_dnxhd_print_profiles(avctx, AV_LOG_ERROR);
414 return AVERROR(EINVAL);
416 av_log(avctx, AV_LOG_DEBUG, "cid %d\n", ctx->cid);
418 if (ctx->cid >= 1270 && ctx->cid <= 1274)
419 avctx->codec_tag = MKTAG('A','V','d','h');
421 if (avctx->width < 256 || avctx->height < 120) {
422 av_log(avctx, AV_LOG_ERROR,
423 "Input dimensions too small, input must be at least 256x120\n");
424 return AVERROR(EINVAL);
427 index = ff_dnxhd_get_cid_table(ctx->cid);
428 av_assert0(index >= 0);
430 ctx->cid_table = &ff_dnxhd_cid_table[index];
432 ctx->m.avctx = avctx;
436 avctx->bits_per_raw_sample = ctx->bit_depth;
438 ff_blockdsp_init(&ctx->bdsp, avctx);
439 ff_fdctdsp_init(&ctx->m.fdsp, avctx);
440 ff_mpv_idct_init(&ctx->m);
441 ff_mpegvideoencdsp_init(&ctx->m.mpvencdsp, avctx);
442 ff_pixblockdsp_init(&ctx->m.pdsp, avctx);
443 ff_dct_encode_init(&ctx->m);
445 if (ctx->profile != FF_PROFILE_DNXHD)
446 ff_videodsp_init(&ctx->m.vdsp, ctx->bit_depth);
448 if (!ctx->m.dct_quantize)
449 ctx->m.dct_quantize = ff_dct_quantize_c;
451 if (ctx->is_444 || ctx->profile == FF_PROFILE_DNXHR_HQX) {
452 ctx->m.dct_quantize = dnxhd_10bit_dct_quantize_444;
453 ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym;
454 ctx->block_width_l2 = 4;
455 } else if (ctx->bit_depth == 10) {
456 ctx->m.dct_quantize = dnxhd_10bit_dct_quantize;
457 ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym;
458 ctx->block_width_l2 = 4;
460 ctx->get_pixels_8x4_sym = dnxhd_8bit_get_pixels_8x4_sym;
461 ctx->block_width_l2 = 3;
465 ff_dnxhdenc_init_x86(ctx);
467 ctx->m.mb_height = (avctx->height + 15) / 16;
468 ctx->m.mb_width = (avctx->width + 15) / 16;
470 if (avctx->flags & AV_CODEC_FLAG_INTERLACED_DCT) {
472 ctx->m.mb_height /= 2;
475 if (ctx->interlaced && ctx->profile != FF_PROFILE_DNXHD) {
476 av_log(avctx, AV_LOG_ERROR,
477 "Interlaced encoding is not supported for DNxHR profiles.\n");
478 return AVERROR(EINVAL);
481 ctx->m.mb_num = ctx->m.mb_height * ctx->m.mb_width;
483 if (ctx->cid_table->frame_size == DNXHD_VARIABLE) {
484 ctx->frame_size = avpriv_dnxhd_get_hr_frame_size(ctx->cid,
485 avctx->width, avctx->height);
486 av_assert0(ctx->frame_size >= 0);
487 ctx->coding_unit_size = ctx->frame_size;
489 ctx->frame_size = ctx->cid_table->frame_size;
490 ctx->coding_unit_size = ctx->cid_table->coding_unit_size;
493 if (ctx->m.mb_height > 68)
494 ctx->data_offset = 0x170 + (ctx->m.mb_height << 2);
496 ctx->data_offset = 0x280;
498 // XXX tune lbias/cbias
499 if ((ret = dnxhd_init_qmat(ctx, ctx->intra_quant_bias, 0)) < 0)
502 /* Avid Nitris hardware decoder requires a minimum amount of padding
503 * in the coding unit payload */
504 if (ctx->nitris_compat)
505 ctx->min_padding = 1600;
507 if ((ret = dnxhd_init_vlc(ctx)) < 0)
509 if ((ret = dnxhd_init_rc(ctx)) < 0)
512 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_size,
513 ctx->m.mb_height * sizeof(uint32_t), fail);
514 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_offs,
515 ctx->m.mb_height * sizeof(uint32_t), fail);
516 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_bits,
517 ctx->m.mb_num * sizeof(uint16_t), fail);
518 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_qscale,
519 ctx->m.mb_num * sizeof(uint8_t), fail);
521 #if FF_API_CODED_FRAME
522 FF_DISABLE_DEPRECATION_WARNINGS
523 avctx->coded_frame->key_frame = 1;
524 avctx->coded_frame->pict_type = AV_PICTURE_TYPE_I;
525 FF_ENABLE_DEPRECATION_WARNINGS
528 if (avctx->active_thread_type == FF_THREAD_SLICE) {
529 if (avctx->thread_count > MAX_THREADS) {
530 av_log(avctx, AV_LOG_ERROR, "too many threads\n");
531 return AVERROR(EINVAL);
535 if (avctx->qmax <= 1) {
536 av_log(avctx, AV_LOG_ERROR, "qmax must be at least 2\n");
537 return AVERROR(EINVAL);
540 ctx->thread[0] = ctx;
541 if (avctx->active_thread_type == FF_THREAD_SLICE) {
542 for (i = 1; i < avctx->thread_count; i++) {
543 ctx->thread[i] = av_malloc(sizeof(DNXHDEncContext));
546 memcpy(ctx->thread[i], ctx, sizeof(DNXHDEncContext));
551 fail: // for FF_ALLOCZ_OR_GOTO
552 return AVERROR(ENOMEM);
555 static int dnxhd_write_header(AVCodecContext *avctx, uint8_t *buf)
557 DNXHDEncContext *ctx = avctx->priv_data;
559 memset(buf, 0, ctx->data_offset);
562 AV_WB16(buf + 0x02, ctx->data_offset);
563 if (ctx->cid >= 1270 && ctx->cid <= 1274)
568 buf[5] = ctx->interlaced ? ctx->cur_field + 2 : 0x01;
569 buf[6] = 0x80; // crc flag off
570 buf[7] = 0xa0; // reserved
571 AV_WB16(buf + 0x18, avctx->height >> ctx->interlaced); // ALPF
572 AV_WB16(buf + 0x1a, avctx->width); // SPL
573 AV_WB16(buf + 0x1d, avctx->height >> ctx->interlaced); // NAL
575 buf[0x21] = ctx->bit_depth == 10 ? 0x58 : 0x38;
576 buf[0x22] = 0x88 + (ctx->interlaced << 2);
577 AV_WB32(buf + 0x28, ctx->cid); // CID
578 buf[0x2c] = (!ctx->interlaced << 7) | (ctx->is_444 << 6) | (avctx->pix_fmt == AV_PIX_FMT_YUV444P10);
580 buf[0x5f] = 0x01; // UDL
582 buf[0x167] = 0x02; // reserved
583 AV_WB16(buf + 0x16a, ctx->m.mb_height * 4 + 4); // MSIPS
584 AV_WB16(buf + 0x16c, ctx->m.mb_height); // Ns
585 buf[0x16f] = 0x10; // reserved
587 ctx->msip = buf + 0x170;
591 static av_always_inline void dnxhd_encode_dc(DNXHDEncContext *ctx, int diff)
595 nbits = av_log2_16bit(-2 * diff);
598 nbits = av_log2_16bit(2 * diff);
600 put_bits(&ctx->m.pb, ctx->cid_table->dc_bits[nbits] + nbits,
601 (ctx->cid_table->dc_codes[nbits] << nbits) +
602 av_mod_uintp2(diff, nbits));
605 static av_always_inline
606 void dnxhd_encode_block(DNXHDEncContext *ctx, int16_t *block,
607 int last_index, int n)
609 int last_non_zero = 0;
612 dnxhd_encode_dc(ctx, block[0] - ctx->m.last_dc[n]);
613 ctx->m.last_dc[n] = block[0];
615 for (i = 1; i <= last_index; i++) {
616 j = ctx->m.intra_scantable.permutated[i];
619 int run_level = i - last_non_zero - 1;
620 int rlevel = slevel * (1 << 1) | !!run_level;
621 put_bits(&ctx->m.pb, ctx->vlc_bits[rlevel], ctx->vlc_codes[rlevel]);
623 put_bits(&ctx->m.pb, ctx->run_bits[run_level],
624 ctx->run_codes[run_level]);
628 put_bits(&ctx->m.pb, ctx->vlc_bits[0], ctx->vlc_codes[0]); // EOB
631 static av_always_inline
632 void dnxhd_unquantize_c(DNXHDEncContext *ctx, int16_t *block, int n,
633 int qscale, int last_index)
635 const uint8_t *weight_matrix;
640 weight_matrix = ((n % 6) < 2) ? ctx->cid_table->luma_weight
641 : ctx->cid_table->chroma_weight;
643 weight_matrix = (n & 2) ? ctx->cid_table->chroma_weight
644 : ctx->cid_table->luma_weight;
647 for (i = 1; i <= last_index; i++) {
648 int j = ctx->m.intra_scantable.permutated[i];
652 level = (1 - 2 * level) * qscale * weight_matrix[i];
653 if (ctx->bit_depth == 10) {
654 if (weight_matrix[i] != 8)
658 if (weight_matrix[i] != 32)
664 level = (2 * level + 1) * qscale * weight_matrix[i];
665 if (ctx->bit_depth == 10) {
666 if (weight_matrix[i] != 8)
670 if (weight_matrix[i] != 32)
680 static av_always_inline int dnxhd_ssd_block(int16_t *qblock, int16_t *block)
684 for (i = 0; i < 64; i++)
685 score += (block[i] - qblock[i]) * (block[i] - qblock[i]);
689 static av_always_inline
690 int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, int16_t *block, int last_index)
692 int last_non_zero = 0;
695 for (i = 1; i <= last_index; i++) {
696 j = ctx->m.intra_scantable.permutated[i];
699 int run_level = i - last_non_zero - 1;
700 bits += ctx->vlc_bits[level * (1 << 1) |
701 !!run_level] + ctx->run_bits[run_level];
708 static av_always_inline
709 void dnxhd_get_blocks(DNXHDEncContext *ctx, int mb_x, int mb_y)
711 const int bs = ctx->block_width_l2;
712 const int bw = 1 << bs;
713 int dct_y_offset = ctx->dct_y_offset;
714 int dct_uv_offset = ctx->dct_uv_offset;
715 int linesize = ctx->m.linesize;
716 int uvlinesize = ctx->m.uvlinesize;
717 const uint8_t *ptr_y = ctx->thread[0]->src[0] +
718 ((mb_y << 4) * ctx->m.linesize) + (mb_x << bs + 1);
719 const uint8_t *ptr_u = ctx->thread[0]->src[1] +
720 ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs + ctx->is_444);
721 const uint8_t *ptr_v = ctx->thread[0]->src[2] +
722 ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs + ctx->is_444);
723 PixblockDSPContext *pdsp = &ctx->m.pdsp;
724 VideoDSPContext *vdsp = &ctx->m.vdsp;
726 if (ctx->bit_depth != 10 && vdsp->emulated_edge_mc && ((mb_x << 4) + 16 > ctx->m.avctx->width ||
727 (mb_y << 4) + 16 > ctx->m.avctx->height)) {
728 int y_w = ctx->m.avctx->width - (mb_x << 4);
729 int y_h = ctx->m.avctx->height - (mb_y << 4);
730 int uv_w = (y_w + 1) / 2;
735 vdsp->emulated_edge_mc(&ctx->edge_buf_y[0], ptr_y,
736 linesize, ctx->m.linesize,
739 vdsp->emulated_edge_mc(&ctx->edge_buf_uv[0][0], ptr_u,
740 uvlinesize, ctx->m.uvlinesize,
743 vdsp->emulated_edge_mc(&ctx->edge_buf_uv[1][0], ptr_v,
744 uvlinesize, ctx->m.uvlinesize,
748 dct_y_offset = bw * linesize;
749 dct_uv_offset = bw * uvlinesize;
750 ptr_y = &ctx->edge_buf_y[0];
751 ptr_u = &ctx->edge_buf_uv[0][0];
752 ptr_v = &ctx->edge_buf_uv[1][0];
753 } else if (ctx->bit_depth == 10 && vdsp->emulated_edge_mc && ((mb_x << 4) + 16 > ctx->m.avctx->width ||
754 (mb_y << 4) + 16 > ctx->m.avctx->height)) {
755 int y_w = ctx->m.avctx->width - (mb_x << 4);
756 int y_h = ctx->m.avctx->height - (mb_y << 4);
757 int uv_w = ctx->is_444 ? y_w : (y_w + 1) / 2;
760 uvlinesize = 16 + 16 * ctx->is_444;
762 vdsp->emulated_edge_mc(&ctx->edge_buf_y[0], ptr_y,
763 linesize, ctx->m.linesize,
766 vdsp->emulated_edge_mc(&ctx->edge_buf_uv[0][0], ptr_u,
767 uvlinesize, ctx->m.uvlinesize,
770 vdsp->emulated_edge_mc(&ctx->edge_buf_uv[1][0], ptr_v,
771 uvlinesize, ctx->m.uvlinesize,
775 dct_y_offset = bw * linesize / 2;
776 dct_uv_offset = bw * uvlinesize / 2;
777 ptr_y = &ctx->edge_buf_y[0];
778 ptr_u = &ctx->edge_buf_uv[0][0];
779 ptr_v = &ctx->edge_buf_uv[1][0];
783 pdsp->get_pixels(ctx->blocks[0], ptr_y, linesize);
784 pdsp->get_pixels(ctx->blocks[1], ptr_y + bw, linesize);
785 pdsp->get_pixels(ctx->blocks[2], ptr_u, uvlinesize);
786 pdsp->get_pixels(ctx->blocks[3], ptr_v, uvlinesize);
788 if (mb_y + 1 == ctx->m.mb_height && ctx->m.avctx->height == 1080) {
789 if (ctx->interlaced) {
790 ctx->get_pixels_8x4_sym(ctx->blocks[4],
791 ptr_y + dct_y_offset,
793 ctx->get_pixels_8x4_sym(ctx->blocks[5],
794 ptr_y + dct_y_offset + bw,
796 ctx->get_pixels_8x4_sym(ctx->blocks[6],
797 ptr_u + dct_uv_offset,
799 ctx->get_pixels_8x4_sym(ctx->blocks[7],
800 ptr_v + dct_uv_offset,
803 ctx->bdsp.clear_block(ctx->blocks[4]);
804 ctx->bdsp.clear_block(ctx->blocks[5]);
805 ctx->bdsp.clear_block(ctx->blocks[6]);
806 ctx->bdsp.clear_block(ctx->blocks[7]);
809 pdsp->get_pixels(ctx->blocks[4],
810 ptr_y + dct_y_offset, linesize);
811 pdsp->get_pixels(ctx->blocks[5],
812 ptr_y + dct_y_offset + bw, linesize);
813 pdsp->get_pixels(ctx->blocks[6],
814 ptr_u + dct_uv_offset, uvlinesize);
815 pdsp->get_pixels(ctx->blocks[7],
816 ptr_v + dct_uv_offset, uvlinesize);
819 pdsp->get_pixels(ctx->blocks[0], ptr_y, linesize);
820 pdsp->get_pixels(ctx->blocks[1], ptr_y + bw, linesize);
821 pdsp->get_pixels(ctx->blocks[6], ptr_y + dct_y_offset, linesize);
822 pdsp->get_pixels(ctx->blocks[7], ptr_y + dct_y_offset + bw, linesize);
824 pdsp->get_pixels(ctx->blocks[2], ptr_u, uvlinesize);
825 pdsp->get_pixels(ctx->blocks[3], ptr_u + bw, uvlinesize);
826 pdsp->get_pixels(ctx->blocks[8], ptr_u + dct_uv_offset, uvlinesize);
827 pdsp->get_pixels(ctx->blocks[9], ptr_u + dct_uv_offset + bw, uvlinesize);
829 pdsp->get_pixels(ctx->blocks[4], ptr_v, uvlinesize);
830 pdsp->get_pixels(ctx->blocks[5], ptr_v + bw, uvlinesize);
831 pdsp->get_pixels(ctx->blocks[10], ptr_v + dct_uv_offset, uvlinesize);
832 pdsp->get_pixels(ctx->blocks[11], ptr_v + dct_uv_offset + bw, uvlinesize);
836 static av_always_inline
837 int dnxhd_switch_matrix(DNXHDEncContext *ctx, int i)
844 const static uint8_t component[8]={0,0,1,2,0,0,1,2};
850 static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg,
851 int jobnr, int threadnr)
853 DNXHDEncContext *ctx = avctx->priv_data;
854 int mb_y = jobnr, mb_x;
855 int qscale = ctx->qscale;
856 LOCAL_ALIGNED_16(int16_t, block, [64]);
857 ctx = ctx->thread[threadnr];
861 ctx->m.last_dc[2] = 1 << (ctx->bit_depth + 2);
863 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
864 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
870 dnxhd_get_blocks(ctx, mb_x, mb_y);
872 for (i = 0; i < 8 + 4 * ctx->is_444; i++) {
873 int16_t *src_block = ctx->blocks[i];
874 int overflow, nbits, diff, last_index;
875 int n = dnxhd_switch_matrix(ctx, i);
877 memcpy(block, src_block, 64 * sizeof(*block));
878 last_index = ctx->m.dct_quantize(&ctx->m, block,
879 ctx->is_444 ? 4 * (n > 0): 4 & (2*i),
881 ac_bits += dnxhd_calc_ac_bits(ctx, block, last_index);
883 diff = block[0] - ctx->m.last_dc[n];
885 nbits = av_log2_16bit(-2 * diff);
887 nbits = av_log2_16bit(2 * diff);
889 av_assert1(nbits < ctx->bit_depth + 4);
890 dc_bits += ctx->cid_table->dc_bits[nbits] + nbits;
892 ctx->m.last_dc[n] = block[0];
894 if (avctx->mb_decision == FF_MB_DECISION_RD || !RC_VARIANCE) {
895 dnxhd_unquantize_c(ctx, block, i, qscale, last_index);
896 ctx->m.idsp.idct(block);
897 ssd += dnxhd_ssd_block(block, src_block);
900 ctx->mb_rc[(qscale * ctx->m.mb_num) + mb].ssd = ssd;
901 ctx->mb_rc[(qscale * ctx->m.mb_num) + mb].bits = ac_bits + dc_bits + 12 +
902 (1 + ctx->is_444) * 8 * ctx->vlc_bits[0];
907 static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg,
908 int jobnr, int threadnr)
910 DNXHDEncContext *ctx = avctx->priv_data;
911 int mb_y = jobnr, mb_x;
912 ctx = ctx->thread[threadnr];
913 init_put_bits(&ctx->m.pb, (uint8_t *)arg + ctx->data_offset + ctx->slice_offs[jobnr],
914 ctx->slice_size[jobnr]);
918 ctx->m.last_dc[2] = 1 << (ctx->bit_depth + 2);
919 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
920 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
921 int qscale = ctx->mb_qscale[mb];
924 put_bits(&ctx->m.pb, 11, qscale);
925 put_bits(&ctx->m.pb, 1, avctx->pix_fmt == AV_PIX_FMT_YUV444P10);
927 dnxhd_get_blocks(ctx, mb_x, mb_y);
929 for (i = 0; i < 8 + 4 * ctx->is_444; i++) {
930 int16_t *block = ctx->blocks[i];
931 int overflow, n = dnxhd_switch_matrix(ctx, i);
932 int last_index = ctx->m.dct_quantize(&ctx->m, block,
933 ctx->is_444 ? (((i >> 1) % 3) < 1 ? 0 : 4): 4 & (2*i),
936 dnxhd_encode_block(ctx, block, last_index, n);
939 if (put_bits_count(&ctx->m.pb) & 31)
940 put_bits(&ctx->m.pb, 32 - (put_bits_count(&ctx->m.pb) & 31), 0);
941 flush_put_bits(&ctx->m.pb);
945 static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx)
949 for (mb_y = 0; mb_y < ctx->m.mb_height; mb_y++) {
951 ctx->slice_offs[mb_y] = offset;
952 ctx->slice_size[mb_y] = 0;
953 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
954 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
955 ctx->slice_size[mb_y] += ctx->mb_bits[mb];
957 ctx->slice_size[mb_y] = (ctx->slice_size[mb_y] + 31) & ~31;
958 ctx->slice_size[mb_y] >>= 3;
959 thread_size = ctx->slice_size[mb_y];
960 offset += thread_size;
964 static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg,
965 int jobnr, int threadnr)
967 DNXHDEncContext *ctx = avctx->priv_data;
968 int mb_y = jobnr, mb_x, x, y;
969 int partial_last_row = (mb_y == ctx->m.mb_height - 1) &&
970 ((avctx->height >> ctx->interlaced) & 0xF);
972 ctx = ctx->thread[threadnr];
973 if (ctx->bit_depth == 8) {
974 uint8_t *pix = ctx->thread[0]->src[0] + ((mb_y << 4) * ctx->m.linesize);
975 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x, pix += 16) {
976 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
980 if (!partial_last_row && mb_x * 16 <= avctx->width - 16 && (avctx->width % 16) == 0) {
981 sum = ctx->m.mpvencdsp.pix_sum(pix, ctx->m.linesize);
982 varc = ctx->m.mpvencdsp.pix_norm1(pix, ctx->m.linesize);
984 int bw = FFMIN(avctx->width - 16 * mb_x, 16);
985 int bh = FFMIN((avctx->height >> ctx->interlaced) - 16 * mb_y, 16);
987 for (y = 0; y < bh; y++) {
988 for (x = 0; x < bw; x++) {
989 uint8_t val = pix[x + y * ctx->m.linesize];
995 varc = (varc - (((unsigned) sum * sum) >> 8) + 128) >> 8;
997 ctx->mb_cmp[mb].value = varc;
998 ctx->mb_cmp[mb].mb = mb;
1001 const int linesize = ctx->m.linesize >> 1;
1002 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x) {
1003 uint16_t *pix = (uint16_t *)ctx->thread[0]->src[0] +
1004 ((mb_y << 4) * linesize) + (mb_x << 4);
1005 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
1008 int bw = FFMIN(avctx->width - 16 * mb_x, 16);
1009 int bh = FFMIN((avctx->height >> ctx->interlaced) - 16 * mb_y, 16);
1012 // Macroblocks are 16x16 pixels, unlike DCT blocks which are 8x8.
1013 for (i = 0; i < bh; ++i) {
1014 for (j = 0; j < bw; ++j) {
1015 // Turn 16-bit pixels into 10-bit ones.
1016 const int sample = (unsigned) pix[j] >> 6;
1018 sqsum += sample * sample;
1019 // 2^10 * 2^10 * 16 * 16 = 2^28, which is less than INT_MAX
1023 mean = sum >> 8; // 16*16 == 2^8
1024 sqmean = sqsum >> 8;
1025 ctx->mb_cmp[mb].value = sqmean - mean * mean;
1026 ctx->mb_cmp[mb].mb = mb;
1032 static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx)
1034 int lambda, up_step, down_step;
1035 int last_lower = INT_MAX, last_higher = 0;
1038 for (q = 1; q < avctx->qmax; q++) {
1040 avctx->execute2(avctx, dnxhd_calc_bits_thread,
1041 NULL, NULL, ctx->m.mb_height);
1043 up_step = down_step = 2 << LAMBDA_FRAC_BITS;
1044 lambda = ctx->lambda;
1049 if (lambda == last_higher) {
1051 end = 1; // need to set final qscales/bits
1053 for (y = 0; y < ctx->m.mb_height; y++) {
1054 for (x = 0; x < ctx->m.mb_width; x++) {
1055 unsigned min = UINT_MAX;
1057 int mb = y * ctx->m.mb_width + x;
1059 for (q = 1; q < avctx->qmax; q++) {
1060 int i = (q*ctx->m.mb_num) + mb;
1061 unsigned score = ctx->mb_rc[i].bits * lambda +
1062 ((unsigned) ctx->mb_rc[i].ssd << LAMBDA_FRAC_BITS);
1069 bits += ctx->mb_rc[rc].bits;
1070 ctx->mb_qscale[mb] = qscale;
1071 ctx->mb_bits[mb] = ctx->mb_rc[rc].bits;
1073 bits = (bits + 31) & ~31; // padding
1074 if (bits > ctx->frame_bits)
1078 if (bits > ctx->frame_bits)
1079 return AVERROR(EINVAL);
1082 if (bits < ctx->frame_bits) {
1083 last_lower = FFMIN(lambda, last_lower);
1084 if (last_higher != 0)
1085 lambda = (lambda+last_higher)>>1;
1087 lambda -= down_step;
1088 down_step = FFMIN((int64_t)down_step*5, INT_MAX);
1089 up_step = 1<<LAMBDA_FRAC_BITS;
1090 lambda = FFMAX(1, lambda);
1091 if (lambda == last_lower)
1094 last_higher = FFMAX(lambda, last_higher);
1095 if (last_lower != INT_MAX)
1096 lambda = (lambda+last_lower)>>1;
1097 else if ((int64_t)lambda + up_step > INT_MAX)
1098 return AVERROR(EINVAL);
1101 up_step = FFMIN((int64_t)up_step*5, INT_MAX);
1102 down_step = 1<<LAMBDA_FRAC_BITS;
1105 ctx->lambda = lambda;
1109 static int dnxhd_find_qscale(DNXHDEncContext *ctx)
1114 int last_higher = 0;
1115 int last_lower = INT_MAX;
1119 qscale = ctx->qscale;
1122 ctx->qscale = qscale;
1123 // XXX avoid recalculating bits
1124 ctx->m.avctx->execute2(ctx->m.avctx, dnxhd_calc_bits_thread,
1125 NULL, NULL, ctx->m.mb_height);
1126 for (y = 0; y < ctx->m.mb_height; y++) {
1127 for (x = 0; x < ctx->m.mb_width; x++)
1128 bits += ctx->mb_rc[(qscale*ctx->m.mb_num) + (y*ctx->m.mb_width+x)].bits;
1129 bits = (bits+31)&~31; // padding
1130 if (bits > ctx->frame_bits)
1133 if (bits < ctx->frame_bits) {
1136 if (last_higher == qscale - 1) {
1137 qscale = last_higher;
1140 last_lower = FFMIN(qscale, last_lower);
1141 if (last_higher != 0)
1142 qscale = (qscale + last_higher) >> 1;
1144 qscale -= down_step++;
1149 if (last_lower == qscale + 1)
1151 last_higher = FFMAX(qscale, last_higher);
1152 if (last_lower != INT_MAX)
1153 qscale = (qscale + last_lower) >> 1;
1155 qscale += up_step++;
1157 if (qscale >= ctx->m.avctx->qmax)
1158 return AVERROR(EINVAL);
1161 ctx->qscale = qscale;
1165 #define BUCKET_BITS 8
1166 #define RADIX_PASSES 4
1167 #define NBUCKETS (1 << BUCKET_BITS)
1169 static inline int get_bucket(int value, int shift)
1172 value &= NBUCKETS - 1;
1173 return NBUCKETS - 1 - value;
1176 static void radix_count(const RCCMPEntry *data, int size,
1177 int buckets[RADIX_PASSES][NBUCKETS])
1180 memset(buckets, 0, sizeof(buckets[0][0]) * RADIX_PASSES * NBUCKETS);
1181 for (i = 0; i < size; i++) {
1182 int v = data[i].value;
1183 for (j = 0; j < RADIX_PASSES; j++) {
1184 buckets[j][get_bucket(v, 0)]++;
1189 for (j = 0; j < RADIX_PASSES; j++) {
1191 for (i = NBUCKETS - 1; i >= 0; i--)
1192 buckets[j][i] = offset -= buckets[j][i];
1193 av_assert1(!buckets[j][0]);
1197 static void radix_sort_pass(RCCMPEntry *dst, const RCCMPEntry *data,
1198 int size, int buckets[NBUCKETS], int pass)
1200 int shift = pass * BUCKET_BITS;
1202 for (i = 0; i < size; i++) {
1203 int v = get_bucket(data[i].value, shift);
1204 int pos = buckets[v]++;
1209 static void radix_sort(RCCMPEntry *data, RCCMPEntry *tmp, int size)
1211 int buckets[RADIX_PASSES][NBUCKETS];
1212 radix_count(data, size, buckets);
1213 radix_sort_pass(tmp, data, size, buckets[0], 0);
1214 radix_sort_pass(data, tmp, size, buckets[1], 1);
1215 if (buckets[2][NBUCKETS - 1] || buckets[3][NBUCKETS - 1]) {
1216 radix_sort_pass(tmp, data, size, buckets[2], 2);
1217 radix_sort_pass(data, tmp, size, buckets[3], 3);
1221 static int dnxhd_encode_fast(AVCodecContext *avctx, DNXHDEncContext *ctx)
1225 if ((ret = dnxhd_find_qscale(ctx)) < 0)
1227 for (y = 0; y < ctx->m.mb_height; y++) {
1228 for (x = 0; x < ctx->m.mb_width; x++) {
1229 int mb = y * ctx->m.mb_width + x;
1230 int rc = (ctx->qscale * ctx->m.mb_num ) + mb;
1232 ctx->mb_qscale[mb] = ctx->qscale;
1233 ctx->mb_bits[mb] = ctx->mb_rc[rc].bits;
1234 max_bits += ctx->mb_rc[rc].bits;
1236 delta_bits = ctx->mb_rc[rc].bits -
1237 ctx->mb_rc[rc + ctx->m.mb_num].bits;
1238 ctx->mb_cmp[mb].mb = mb;
1239 ctx->mb_cmp[mb].value =
1240 delta_bits ? ((ctx->mb_rc[rc].ssd -
1241 ctx->mb_rc[rc + ctx->m.mb_num].ssd) * 100) /
1243 : INT_MIN; // avoid increasing qscale
1246 max_bits += 31; // worst padding
1250 avctx->execute2(avctx, dnxhd_mb_var_thread,
1251 NULL, NULL, ctx->m.mb_height);
1252 radix_sort(ctx->mb_cmp, ctx->mb_cmp_tmp, ctx->m.mb_num);
1253 for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) {
1254 int mb = ctx->mb_cmp[x].mb;
1255 int rc = (ctx->qscale * ctx->m.mb_num ) + mb;
1256 max_bits -= ctx->mb_rc[rc].bits -
1257 ctx->mb_rc[rc + ctx->m.mb_num].bits;
1258 ctx->mb_qscale[mb] = ctx->qscale + 1;
1259 ctx->mb_bits[mb] = ctx->mb_rc[rc + ctx->m.mb_num].bits;
1265 static void dnxhd_load_picture(DNXHDEncContext *ctx, const AVFrame *frame)
1269 for (i = 0; i < ctx->m.avctx->thread_count; i++) {
1270 ctx->thread[i]->m.linesize = frame->linesize[0] << ctx->interlaced;
1271 ctx->thread[i]->m.uvlinesize = frame->linesize[1] << ctx->interlaced;
1272 ctx->thread[i]->dct_y_offset = ctx->m.linesize *8;
1273 ctx->thread[i]->dct_uv_offset = ctx->m.uvlinesize*8;
1276 #if FF_API_CODED_FRAME
1277 FF_DISABLE_DEPRECATION_WARNINGS
1278 ctx->m.avctx->coded_frame->interlaced_frame = frame->interlaced_frame;
1279 FF_ENABLE_DEPRECATION_WARNINGS
1281 ctx->cur_field = frame->interlaced_frame && !frame->top_field_first;
1284 static int dnxhd_encode_picture(AVCodecContext *avctx, AVPacket *pkt,
1285 const AVFrame *frame, int *got_packet)
1287 DNXHDEncContext *ctx = avctx->priv_data;
1288 int first_field = 1;
1292 if ((ret = ff_alloc_packet2(avctx, pkt, ctx->frame_size, 0)) < 0)
1296 dnxhd_load_picture(ctx, frame);
1299 for (i = 0; i < 3; i++) {
1300 ctx->src[i] = frame->data[i];
1301 if (ctx->interlaced && ctx->cur_field)
1302 ctx->src[i] += frame->linesize[i];
1305 dnxhd_write_header(avctx, buf);
1307 if (avctx->mb_decision == FF_MB_DECISION_RD)
1308 ret = dnxhd_encode_rdo(avctx, ctx);
1310 ret = dnxhd_encode_fast(avctx, ctx);
1312 av_log(avctx, AV_LOG_ERROR,
1313 "picture could not fit ratecontrol constraints, increase qmax\n");
1317 dnxhd_setup_threads_slices(ctx);
1320 for (i = 0; i < ctx->m.mb_height; i++) {
1321 AV_WB32(ctx->msip + i * 4, offset);
1322 offset += ctx->slice_size[i];
1323 av_assert1(!(ctx->slice_size[i] & 3));
1326 avctx->execute2(avctx, dnxhd_encode_thread, buf, NULL, ctx->m.mb_height);
1328 av_assert1(ctx->data_offset + offset + 4 <= ctx->coding_unit_size);
1329 memset(buf + ctx->data_offset + offset, 0,
1330 ctx->coding_unit_size - 4 - offset - ctx->data_offset);
1332 AV_WB32(buf + ctx->coding_unit_size - 4, 0x600DC0DE); // EOF
1334 if (ctx->interlaced && first_field) {
1336 ctx->cur_field ^= 1;
1337 buf += ctx->coding_unit_size;
1338 goto encode_coding_unit;
1341 #if FF_API_CODED_FRAME
1342 FF_DISABLE_DEPRECATION_WARNINGS
1343 avctx->coded_frame->quality = ctx->qscale * FF_QP2LAMBDA;
1344 FF_ENABLE_DEPRECATION_WARNINGS
1347 ff_side_data_set_encoder_stats(pkt, ctx->qscale * FF_QP2LAMBDA, NULL, 0, AV_PICTURE_TYPE_I);
1349 pkt->flags |= AV_PKT_FLAG_KEY;
1354 static av_cold int dnxhd_encode_end(AVCodecContext *avctx)
1356 DNXHDEncContext *ctx = avctx->priv_data;
1359 av_freep(&ctx->orig_vlc_codes);
1360 av_freep(&ctx->orig_vlc_bits);
1361 av_freep(&ctx->run_codes);
1362 av_freep(&ctx->run_bits);
1364 av_freep(&ctx->mb_bits);
1365 av_freep(&ctx->mb_qscale);
1366 av_freep(&ctx->mb_rc);
1367 av_freep(&ctx->mb_cmp);
1368 av_freep(&ctx->mb_cmp_tmp);
1369 av_freep(&ctx->slice_size);
1370 av_freep(&ctx->slice_offs);
1372 av_freep(&ctx->qmatrix_c);
1373 av_freep(&ctx->qmatrix_l);
1374 av_freep(&ctx->qmatrix_c16);
1375 av_freep(&ctx->qmatrix_l16);
1377 if (avctx->active_thread_type == FF_THREAD_SLICE) {
1378 for (i = 1; i < avctx->thread_count; i++)
1379 av_freep(&ctx->thread[i]);
1385 static const AVCodecDefault dnxhd_defaults[] = {
1386 { "qmax", "1024" }, /* Maximum quantization scale factor allowed for VC-3 */
1390 AVCodec ff_dnxhd_encoder = {
1392 .long_name = NULL_IF_CONFIG_SMALL("VC3/DNxHD"),
1393 .type = AVMEDIA_TYPE_VIDEO,
1394 .id = AV_CODEC_ID_DNXHD,
1395 .priv_data_size = sizeof(DNXHDEncContext),
1396 .init = dnxhd_encode_init,
1397 .encode2 = dnxhd_encode_picture,
1398 .close = dnxhd_encode_end,
1399 .capabilities = AV_CODEC_CAP_SLICE_THREADS | AV_CODEC_CAP_FRAME_THREADS,
1400 .caps_internal = FF_CODEC_CAP_INIT_CLEANUP,
1401 .pix_fmts = (const enum AVPixelFormat[]) {
1403 AV_PIX_FMT_YUV422P10,
1404 AV_PIX_FMT_YUV444P10,
1408 .priv_class = &dnxhd_class,
1409 .defaults = dnxhd_defaults,
1410 .profiles = NULL_IF_CONFIG_SMALL(ff_dnxhd_profiles),