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_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 FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->orig_vlc_codes,
211 max_level, 4 * sizeof(*ctx->orig_vlc_codes), fail);
212 FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->orig_vlc_bits,
213 max_level, 4 * sizeof(*ctx->orig_vlc_bits), fail);
214 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_codes,
216 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_bits,
219 ctx->vlc_codes = ctx->orig_vlc_codes + max_level * 2;
220 ctx->vlc_bits = ctx->orig_vlc_bits + max_level * 2;
221 for (level = -max_level; level < max_level; level++) {
222 for (run = 0; run < 2; run++) {
223 int index = level * (1 << 1) | run;
224 int sign, offset = 0, alevel = level;
226 MASK_ABS(sign, alevel);
228 offset = (alevel - 1) >> 6;
229 alevel -= offset << 6;
231 for (j = 0; j < 257; j++) {
232 if (ctx->cid_table->ac_info[2*j+0] >> 1 == alevel &&
233 (!offset || (ctx->cid_table->ac_info[2*j+1] & 1) && offset) &&
234 (!run || (ctx->cid_table->ac_info[2*j+1] & 2) && run)) {
235 av_assert1(!ctx->vlc_codes[index]);
237 ctx->vlc_codes[index] =
238 (ctx->cid_table->ac_codes[j] << 1) | (sign & 1);
239 ctx->vlc_bits[index] = ctx->cid_table->ac_bits[j] + 1;
241 ctx->vlc_codes[index] = ctx->cid_table->ac_codes[j];
242 ctx->vlc_bits[index] = ctx->cid_table->ac_bits[j];
247 av_assert0(!alevel || j < 257);
249 ctx->vlc_codes[index] =
250 (ctx->vlc_codes[index] << ctx->cid_table->index_bits) | offset;
251 ctx->vlc_bits[index] += ctx->cid_table->index_bits;
255 for (i = 0; i < 62; i++) {
256 int run = ctx->cid_table->run[i];
257 av_assert0(run < 63);
258 ctx->run_codes[run] = ctx->cid_table->run_codes[i];
259 ctx->run_bits[run] = ctx->cid_table->run_bits[i];
263 return AVERROR(ENOMEM);
266 static av_cold int dnxhd_init_qmat(DNXHDEncContext *ctx, int lbias, int cbias)
268 // init first elem to 1 to avoid div by 0 in convert_matrix
269 uint16_t weight_matrix[64] = { 1, }; // convert_matrix needs uint16_t*
271 const uint8_t *luma_weight_table = ctx->cid_table->luma_weight;
272 const uint8_t *chroma_weight_table = ctx->cid_table->chroma_weight;
274 FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l,
275 (ctx->m.avctx->qmax + 1), 64 * sizeof(int), fail);
276 FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c,
277 (ctx->m.avctx->qmax + 1), 64 * sizeof(int), fail);
278 FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l16,
279 (ctx->m.avctx->qmax + 1), 64 * 2 * sizeof(uint16_t),
281 FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c16,
282 (ctx->m.avctx->qmax + 1), 64 * 2 * sizeof(uint16_t),
285 if (ctx->bit_depth == 8) {
286 for (i = 1; i < 64; i++) {
287 int j = ctx->m.idsp.idct_permutation[ff_zigzag_direct[i]];
288 weight_matrix[j] = ctx->cid_table->luma_weight[i];
290 ff_convert_matrix(&ctx->m, ctx->qmatrix_l, ctx->qmatrix_l16,
291 weight_matrix, ctx->intra_quant_bias, 1,
292 ctx->m.avctx->qmax, 1);
293 for (i = 1; i < 64; i++) {
294 int j = ctx->m.idsp.idct_permutation[ff_zigzag_direct[i]];
295 weight_matrix[j] = ctx->cid_table->chroma_weight[i];
297 ff_convert_matrix(&ctx->m, ctx->qmatrix_c, ctx->qmatrix_c16,
298 weight_matrix, ctx->intra_quant_bias, 1,
299 ctx->m.avctx->qmax, 1);
301 for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
302 for (i = 0; i < 64; i++) {
303 ctx->qmatrix_l[qscale][i] <<= 2;
304 ctx->qmatrix_c[qscale][i] <<= 2;
305 ctx->qmatrix_l16[qscale][0][i] <<= 2;
306 ctx->qmatrix_l16[qscale][1][i] <<= 2;
307 ctx->qmatrix_c16[qscale][0][i] <<= 2;
308 ctx->qmatrix_c16[qscale][1][i] <<= 2;
313 for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
314 for (i = 1; i < 64; i++) {
315 int j = ff_zigzag_direct[i];
317 /* The quantization formula from the VC-3 standard is:
318 * quantized = sign(block[i]) * floor(abs(block[i]/s) * p /
319 * (qscale * weight_table[i]))
320 * Where p is 32 for 8-bit samples and 8 for 10-bit ones.
321 * The s factor compensates scaling of DCT coefficients done by
322 * the DCT routines, and therefore is not present in standard.
323 * It's 8 for 8-bit samples and 4 for 10-bit ones.
324 * We want values of ctx->qtmatrix_l and ctx->qtmatrix_r to be:
325 * ((1 << DNX10BIT_QMAT_SHIFT) * (p / s)) /
326 * (qscale * weight_table[i])
327 * For 10-bit samples, p / s == 2 */
328 ctx->qmatrix_l[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) /
329 (qscale * luma_weight_table[i]);
330 ctx->qmatrix_c[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) /
331 (qscale * chroma_weight_table[i]);
336 ctx->m.q_chroma_intra_matrix16 = ctx->qmatrix_c16;
337 ctx->m.q_chroma_intra_matrix = ctx->qmatrix_c;
338 ctx->m.q_intra_matrix16 = ctx->qmatrix_l16;
339 ctx->m.q_intra_matrix = ctx->qmatrix_l;
343 return AVERROR(ENOMEM);
346 static av_cold int dnxhd_init_rc(DNXHDEncContext *ctx)
348 FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->mb_rc, (ctx->m.avctx->qmax + 1),
349 ctx->m.mb_num * sizeof(RCEntry), fail);
350 if (ctx->m.avctx->mb_decision != FF_MB_DECISION_RD) {
351 FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->mb_cmp,
352 ctx->m.mb_num, sizeof(RCCMPEntry), fail);
353 FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->mb_cmp_tmp,
354 ctx->m.mb_num, sizeof(RCCMPEntry), fail);
356 ctx->frame_bits = (ctx->coding_unit_size -
357 ctx->data_offset - 4 - ctx->min_padding) * 8;
359 ctx->lambda = 2 << LAMBDA_FRAC_BITS; // qscale 2
362 return AVERROR(ENOMEM);
365 static av_cold int dnxhd_encode_init(AVCodecContext *avctx)
367 DNXHDEncContext *ctx = avctx->priv_data;
370 switch (avctx->pix_fmt) {
371 case AV_PIX_FMT_YUV422P:
374 case AV_PIX_FMT_YUV422P10:
375 case AV_PIX_FMT_YUV444P10:
376 case AV_PIX_FMT_GBRP10:
380 av_log(avctx, AV_LOG_ERROR,
381 "pixel format is incompatible with DNxHD\n");
382 return AVERROR(EINVAL);
385 if ((ctx->profile == FF_PROFILE_DNXHR_444 && (avctx->pix_fmt != AV_PIX_FMT_YUV444P10 &&
386 avctx->pix_fmt != AV_PIX_FMT_GBRP10)) ||
387 (ctx->profile != FF_PROFILE_DNXHR_444 && (avctx->pix_fmt == AV_PIX_FMT_YUV444P10 ||
388 avctx->pix_fmt == AV_PIX_FMT_GBRP10))) {
389 av_log(avctx, AV_LOG_ERROR,
390 "pixel format is incompatible with DNxHD profile\n");
391 return AVERROR(EINVAL);
394 if (ctx->profile == FF_PROFILE_DNXHR_HQX && avctx->pix_fmt != AV_PIX_FMT_YUV422P10) {
395 av_log(avctx, AV_LOG_ERROR,
396 "pixel format is incompatible with DNxHR HQX profile\n");
397 return AVERROR(EINVAL);
400 if ((ctx->profile == FF_PROFILE_DNXHR_LB ||
401 ctx->profile == FF_PROFILE_DNXHR_SQ ||
402 ctx->profile == FF_PROFILE_DNXHR_HQ) && avctx->pix_fmt != AV_PIX_FMT_YUV422P) {
403 av_log(avctx, AV_LOG_ERROR,
404 "pixel format is incompatible with DNxHR LB/SQ/HQ profile\n");
405 return AVERROR(EINVAL);
408 ctx->is_444 = ctx->profile == FF_PROFILE_DNXHR_444;
409 avctx->profile = ctx->profile;
410 ctx->cid = ff_dnxhd_find_cid(avctx, ctx->bit_depth);
412 av_log(avctx, AV_LOG_ERROR,
413 "video parameters incompatible with DNxHD. Valid DNxHD profiles:\n");
414 ff_dnxhd_print_profiles(avctx, AV_LOG_ERROR);
415 return AVERROR(EINVAL);
417 av_log(avctx, AV_LOG_DEBUG, "cid %d\n", ctx->cid);
419 if (ctx->cid >= 1270 && ctx->cid <= 1274)
420 avctx->codec_tag = MKTAG('A','V','d','h');
422 if (avctx->width < 256 || avctx->height < 120) {
423 av_log(avctx, AV_LOG_ERROR,
424 "Input dimensions too small, input must be at least 256x120\n");
425 return AVERROR(EINVAL);
428 index = ff_dnxhd_get_cid_table(ctx->cid);
429 av_assert0(index >= 0);
431 ctx->cid_table = &ff_dnxhd_cid_table[index];
433 ctx->m.avctx = avctx;
437 avctx->bits_per_raw_sample = ctx->bit_depth;
439 ff_blockdsp_init(&ctx->bdsp, avctx);
440 ff_fdctdsp_init(&ctx->m.fdsp, avctx);
441 ff_mpv_idct_init(&ctx->m);
442 ff_mpegvideoencdsp_init(&ctx->m.mpvencdsp, avctx);
443 ff_pixblockdsp_init(&ctx->m.pdsp, avctx);
444 ff_dct_encode_init(&ctx->m);
446 if (ctx->profile != FF_PROFILE_DNXHD)
447 ff_videodsp_init(&ctx->m.vdsp, ctx->bit_depth);
449 if (!ctx->m.dct_quantize)
450 ctx->m.dct_quantize = ff_dct_quantize_c;
452 if (ctx->is_444 || ctx->profile == FF_PROFILE_DNXHR_HQX) {
453 ctx->m.dct_quantize = dnxhd_10bit_dct_quantize_444;
454 ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym;
455 ctx->block_width_l2 = 4;
456 } else if (ctx->bit_depth == 10) {
457 ctx->m.dct_quantize = dnxhd_10bit_dct_quantize;
458 ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym;
459 ctx->block_width_l2 = 4;
461 ctx->get_pixels_8x4_sym = dnxhd_8bit_get_pixels_8x4_sym;
462 ctx->block_width_l2 = 3;
466 ff_dnxhdenc_init_x86(ctx);
468 ctx->m.mb_height = (avctx->height + 15) / 16;
469 ctx->m.mb_width = (avctx->width + 15) / 16;
471 if (avctx->flags & AV_CODEC_FLAG_INTERLACED_DCT) {
473 ctx->m.mb_height /= 2;
476 if (ctx->interlaced && ctx->profile != FF_PROFILE_DNXHD) {
477 av_log(avctx, AV_LOG_ERROR,
478 "Interlaced encoding is not supported for DNxHR profiles.\n");
479 return AVERROR(EINVAL);
482 ctx->m.mb_num = ctx->m.mb_height * ctx->m.mb_width;
484 if (ctx->cid_table->frame_size == DNXHD_VARIABLE) {
485 ctx->frame_size = avpriv_dnxhd_get_hr_frame_size(ctx->cid,
486 avctx->width, avctx->height);
487 av_assert0(ctx->frame_size >= 0);
488 ctx->coding_unit_size = ctx->frame_size;
490 ctx->frame_size = ctx->cid_table->frame_size;
491 ctx->coding_unit_size = ctx->cid_table->coding_unit_size;
494 if (ctx->m.mb_height > 68)
495 ctx->data_offset = 0x170 + (ctx->m.mb_height << 2);
497 ctx->data_offset = 0x280;
499 // XXX tune lbias/cbias
500 if ((ret = dnxhd_init_qmat(ctx, ctx->intra_quant_bias, 0)) < 0)
503 /* Avid Nitris hardware decoder requires a minimum amount of padding
504 * in the coding unit payload */
505 if (ctx->nitris_compat)
506 ctx->min_padding = 1600;
508 if ((ret = dnxhd_init_vlc(ctx)) < 0)
510 if ((ret = dnxhd_init_rc(ctx)) < 0)
513 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_size,
514 ctx->m.mb_height * sizeof(uint32_t), fail);
515 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_offs,
516 ctx->m.mb_height * sizeof(uint32_t), fail);
517 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_bits,
518 ctx->m.mb_num * sizeof(uint16_t), fail);
519 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_qscale,
520 ctx->m.mb_num * sizeof(uint8_t), fail);
522 #if FF_API_CODED_FRAME
523 FF_DISABLE_DEPRECATION_WARNINGS
524 avctx->coded_frame->key_frame = 1;
525 avctx->coded_frame->pict_type = AV_PICTURE_TYPE_I;
526 FF_ENABLE_DEPRECATION_WARNINGS
529 if (avctx->active_thread_type == FF_THREAD_SLICE) {
530 if (avctx->thread_count > MAX_THREADS) {
531 av_log(avctx, AV_LOG_ERROR, "too many threads\n");
532 return AVERROR(EINVAL);
536 if (avctx->qmax <= 1) {
537 av_log(avctx, AV_LOG_ERROR, "qmax must be at least 2\n");
538 return AVERROR(EINVAL);
541 ctx->thread[0] = ctx;
542 if (avctx->active_thread_type == FF_THREAD_SLICE) {
543 for (i = 1; i < avctx->thread_count; i++) {
544 ctx->thread[i] = av_malloc(sizeof(DNXHDEncContext));
547 memcpy(ctx->thread[i], ctx, sizeof(DNXHDEncContext));
552 fail: // for FF_ALLOCZ_OR_GOTO
553 return AVERROR(ENOMEM);
556 static int dnxhd_write_header(AVCodecContext *avctx, uint8_t *buf)
558 DNXHDEncContext *ctx = avctx->priv_data;
560 memset(buf, 0, ctx->data_offset);
563 AV_WB16(buf + 0x02, ctx->data_offset);
564 if (ctx->cid >= 1270 && ctx->cid <= 1274)
569 buf[5] = ctx->interlaced ? ctx->cur_field + 2 : 0x01;
570 buf[6] = 0x80; // crc flag off
571 buf[7] = 0xa0; // reserved
572 AV_WB16(buf + 0x18, avctx->height >> ctx->interlaced); // ALPF
573 AV_WB16(buf + 0x1a, avctx->width); // SPL
574 AV_WB16(buf + 0x1d, avctx->height >> ctx->interlaced); // NAL
576 buf[0x21] = ctx->bit_depth == 10 ? 0x58 : 0x38;
577 buf[0x22] = 0x88 + (ctx->interlaced << 2);
578 AV_WB32(buf + 0x28, ctx->cid); // CID
579 buf[0x2c] = (!ctx->interlaced << 7) | (ctx->is_444 << 6) | (avctx->pix_fmt == AV_PIX_FMT_YUV444P10);
581 buf[0x5f] = 0x01; // UDL
583 buf[0x167] = 0x02; // reserved
584 AV_WB16(buf + 0x16a, ctx->m.mb_height * 4 + 4); // MSIPS
585 AV_WB16(buf + 0x16c, ctx->m.mb_height); // Ns
586 buf[0x16f] = 0x10; // reserved
588 ctx->msip = buf + 0x170;
592 static av_always_inline void dnxhd_encode_dc(DNXHDEncContext *ctx, int diff)
596 nbits = av_log2_16bit(-2 * diff);
599 nbits = av_log2_16bit(2 * diff);
601 put_bits(&ctx->m.pb, ctx->cid_table->dc_bits[nbits] + nbits,
602 (ctx->cid_table->dc_codes[nbits] << nbits) +
603 av_mod_uintp2(diff, nbits));
606 static av_always_inline
607 void dnxhd_encode_block(DNXHDEncContext *ctx, int16_t *block,
608 int last_index, int n)
610 int last_non_zero = 0;
613 dnxhd_encode_dc(ctx, block[0] - ctx->m.last_dc[n]);
614 ctx->m.last_dc[n] = block[0];
616 for (i = 1; i <= last_index; i++) {
617 j = ctx->m.intra_scantable.permutated[i];
620 int run_level = i - last_non_zero - 1;
621 int rlevel = slevel * (1 << 1) | !!run_level;
622 put_bits(&ctx->m.pb, ctx->vlc_bits[rlevel], ctx->vlc_codes[rlevel]);
624 put_bits(&ctx->m.pb, ctx->run_bits[run_level],
625 ctx->run_codes[run_level]);
629 put_bits(&ctx->m.pb, ctx->vlc_bits[0], ctx->vlc_codes[0]); // EOB
632 static av_always_inline
633 void dnxhd_unquantize_c(DNXHDEncContext *ctx, int16_t *block, int n,
634 int qscale, int last_index)
636 const uint8_t *weight_matrix;
641 weight_matrix = ((n % 6) < 2) ? ctx->cid_table->luma_weight
642 : ctx->cid_table->chroma_weight;
644 weight_matrix = (n & 2) ? ctx->cid_table->chroma_weight
645 : ctx->cid_table->luma_weight;
648 for (i = 1; i <= last_index; i++) {
649 int j = ctx->m.intra_scantable.permutated[i];
653 level = (1 - 2 * level) * qscale * weight_matrix[i];
654 if (ctx->bit_depth == 10) {
655 if (weight_matrix[i] != 8)
659 if (weight_matrix[i] != 32)
665 level = (2 * level + 1) * qscale * weight_matrix[i];
666 if (ctx->bit_depth == 10) {
667 if (weight_matrix[i] != 8)
671 if (weight_matrix[i] != 32)
681 static av_always_inline int dnxhd_ssd_block(int16_t *qblock, int16_t *block)
685 for (i = 0; i < 64; i++)
686 score += (block[i] - qblock[i]) * (block[i] - qblock[i]);
690 static av_always_inline
691 int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, int16_t *block, int last_index)
693 int last_non_zero = 0;
696 for (i = 1; i <= last_index; i++) {
697 j = ctx->m.intra_scantable.permutated[i];
700 int run_level = i - last_non_zero - 1;
701 bits += ctx->vlc_bits[level * (1 << 1) |
702 !!run_level] + ctx->run_bits[run_level];
709 static av_always_inline
710 void dnxhd_get_blocks(DNXHDEncContext *ctx, int mb_x, int mb_y)
712 const int bs = ctx->block_width_l2;
713 const int bw = 1 << bs;
714 int dct_y_offset = ctx->dct_y_offset;
715 int dct_uv_offset = ctx->dct_uv_offset;
716 int linesize = ctx->m.linesize;
717 int uvlinesize = ctx->m.uvlinesize;
718 const uint8_t *ptr_y = ctx->thread[0]->src[0] +
719 ((mb_y << 4) * ctx->m.linesize) + (mb_x << bs + 1);
720 const uint8_t *ptr_u = ctx->thread[0]->src[1] +
721 ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs + ctx->is_444);
722 const uint8_t *ptr_v = ctx->thread[0]->src[2] +
723 ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs + ctx->is_444);
724 PixblockDSPContext *pdsp = &ctx->m.pdsp;
725 VideoDSPContext *vdsp = &ctx->m.vdsp;
727 if (ctx->bit_depth != 10 && vdsp->emulated_edge_mc && ((mb_x << 4) + 16 > ctx->m.avctx->width ||
728 (mb_y << 4) + 16 > ctx->m.avctx->height)) {
729 int y_w = ctx->m.avctx->width - (mb_x << 4);
730 int y_h = ctx->m.avctx->height - (mb_y << 4);
731 int uv_w = (y_w + 1) / 2;
736 vdsp->emulated_edge_mc(&ctx->edge_buf_y[0], ptr_y,
737 linesize, ctx->m.linesize,
740 vdsp->emulated_edge_mc(&ctx->edge_buf_uv[0][0], ptr_u,
741 uvlinesize, ctx->m.uvlinesize,
744 vdsp->emulated_edge_mc(&ctx->edge_buf_uv[1][0], ptr_v,
745 uvlinesize, ctx->m.uvlinesize,
749 dct_y_offset = bw * linesize;
750 dct_uv_offset = bw * uvlinesize;
751 ptr_y = &ctx->edge_buf_y[0];
752 ptr_u = &ctx->edge_buf_uv[0][0];
753 ptr_v = &ctx->edge_buf_uv[1][0];
754 } else if (ctx->bit_depth == 10 && vdsp->emulated_edge_mc && ((mb_x << 4) + 16 > ctx->m.avctx->width ||
755 (mb_y << 4) + 16 > ctx->m.avctx->height)) {
756 int y_w = ctx->m.avctx->width - (mb_x << 4);
757 int y_h = ctx->m.avctx->height - (mb_y << 4);
758 int uv_w = ctx->is_444 ? y_w : (y_w + 1) / 2;
761 uvlinesize = 16 + 16 * ctx->is_444;
763 vdsp->emulated_edge_mc(&ctx->edge_buf_y[0], ptr_y,
764 linesize, ctx->m.linesize,
767 vdsp->emulated_edge_mc(&ctx->edge_buf_uv[0][0], ptr_u,
768 uvlinesize, ctx->m.uvlinesize,
771 vdsp->emulated_edge_mc(&ctx->edge_buf_uv[1][0], ptr_v,
772 uvlinesize, ctx->m.uvlinesize,
776 dct_y_offset = bw * linesize / 2;
777 dct_uv_offset = bw * uvlinesize / 2;
778 ptr_y = &ctx->edge_buf_y[0];
779 ptr_u = &ctx->edge_buf_uv[0][0];
780 ptr_v = &ctx->edge_buf_uv[1][0];
784 pdsp->get_pixels(ctx->blocks[0], ptr_y, linesize);
785 pdsp->get_pixels(ctx->blocks[1], ptr_y + bw, linesize);
786 pdsp->get_pixels(ctx->blocks[2], ptr_u, uvlinesize);
787 pdsp->get_pixels(ctx->blocks[3], ptr_v, uvlinesize);
789 if (mb_y + 1 == ctx->m.mb_height && ctx->m.avctx->height == 1080) {
790 if (ctx->interlaced) {
791 ctx->get_pixels_8x4_sym(ctx->blocks[4],
792 ptr_y + dct_y_offset,
794 ctx->get_pixels_8x4_sym(ctx->blocks[5],
795 ptr_y + dct_y_offset + bw,
797 ctx->get_pixels_8x4_sym(ctx->blocks[6],
798 ptr_u + dct_uv_offset,
800 ctx->get_pixels_8x4_sym(ctx->blocks[7],
801 ptr_v + dct_uv_offset,
804 ctx->bdsp.clear_block(ctx->blocks[4]);
805 ctx->bdsp.clear_block(ctx->blocks[5]);
806 ctx->bdsp.clear_block(ctx->blocks[6]);
807 ctx->bdsp.clear_block(ctx->blocks[7]);
810 pdsp->get_pixels(ctx->blocks[4],
811 ptr_y + dct_y_offset, linesize);
812 pdsp->get_pixels(ctx->blocks[5],
813 ptr_y + dct_y_offset + bw, linesize);
814 pdsp->get_pixels(ctx->blocks[6],
815 ptr_u + dct_uv_offset, uvlinesize);
816 pdsp->get_pixels(ctx->blocks[7],
817 ptr_v + dct_uv_offset, uvlinesize);
820 pdsp->get_pixels(ctx->blocks[0], ptr_y, linesize);
821 pdsp->get_pixels(ctx->blocks[1], ptr_y + bw, linesize);
822 pdsp->get_pixels(ctx->blocks[6], ptr_y + dct_y_offset, linesize);
823 pdsp->get_pixels(ctx->blocks[7], ptr_y + dct_y_offset + bw, linesize);
825 pdsp->get_pixels(ctx->blocks[2], ptr_u, uvlinesize);
826 pdsp->get_pixels(ctx->blocks[3], ptr_u + bw, uvlinesize);
827 pdsp->get_pixels(ctx->blocks[8], ptr_u + dct_uv_offset, uvlinesize);
828 pdsp->get_pixels(ctx->blocks[9], ptr_u + dct_uv_offset + bw, uvlinesize);
830 pdsp->get_pixels(ctx->blocks[4], ptr_v, uvlinesize);
831 pdsp->get_pixels(ctx->blocks[5], ptr_v + bw, uvlinesize);
832 pdsp->get_pixels(ctx->blocks[10], ptr_v + dct_uv_offset, uvlinesize);
833 pdsp->get_pixels(ctx->blocks[11], ptr_v + dct_uv_offset + bw, uvlinesize);
837 static av_always_inline
838 int dnxhd_switch_matrix(DNXHDEncContext *ctx, int i)
845 const static uint8_t component[8]={0,0,1,2,0,0,1,2};
851 static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg,
852 int jobnr, int threadnr)
854 DNXHDEncContext *ctx = avctx->priv_data;
855 int mb_y = jobnr, mb_x;
856 int qscale = ctx->qscale;
857 LOCAL_ALIGNED_16(int16_t, block, [64]);
858 ctx = ctx->thread[threadnr];
862 ctx->m.last_dc[2] = 1 << (ctx->bit_depth + 2);
864 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
865 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
871 dnxhd_get_blocks(ctx, mb_x, mb_y);
873 for (i = 0; i < 8 + 4 * ctx->is_444; i++) {
874 int16_t *src_block = ctx->blocks[i];
875 int overflow, nbits, diff, last_index;
876 int n = dnxhd_switch_matrix(ctx, i);
878 memcpy(block, src_block, 64 * sizeof(*block));
879 last_index = ctx->m.dct_quantize(&ctx->m, block,
880 ctx->is_444 ? 4 * (n > 0): 4 & (2*i),
882 ac_bits += dnxhd_calc_ac_bits(ctx, block, last_index);
884 diff = block[0] - ctx->m.last_dc[n];
886 nbits = av_log2_16bit(-2 * diff);
888 nbits = av_log2_16bit(2 * diff);
890 av_assert1(nbits < ctx->bit_depth + 4);
891 dc_bits += ctx->cid_table->dc_bits[nbits] + nbits;
893 ctx->m.last_dc[n] = block[0];
895 if (avctx->mb_decision == FF_MB_DECISION_RD || !RC_VARIANCE) {
896 dnxhd_unquantize_c(ctx, block, i, qscale, last_index);
897 ctx->m.idsp.idct(block);
898 ssd += dnxhd_ssd_block(block, src_block);
901 ctx->mb_rc[(qscale * ctx->m.mb_num) + mb].ssd = ssd;
902 ctx->mb_rc[(qscale * ctx->m.mb_num) + mb].bits = ac_bits + dc_bits + 12 +
903 (1 + ctx->is_444) * 8 * ctx->vlc_bits[0];
908 static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg,
909 int jobnr, int threadnr)
911 DNXHDEncContext *ctx = avctx->priv_data;
912 int mb_y = jobnr, mb_x;
913 ctx = ctx->thread[threadnr];
914 init_put_bits(&ctx->m.pb, (uint8_t *)arg + ctx->data_offset + ctx->slice_offs[jobnr],
915 ctx->slice_size[jobnr]);
919 ctx->m.last_dc[2] = 1 << (ctx->bit_depth + 2);
920 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
921 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
922 int qscale = ctx->mb_qscale[mb];
925 put_bits(&ctx->m.pb, 11, qscale);
926 put_bits(&ctx->m.pb, 1, avctx->pix_fmt == AV_PIX_FMT_YUV444P10);
928 dnxhd_get_blocks(ctx, mb_x, mb_y);
930 for (i = 0; i < 8 + 4 * ctx->is_444; i++) {
931 int16_t *block = ctx->blocks[i];
932 int overflow, n = dnxhd_switch_matrix(ctx, i);
933 int last_index = ctx->m.dct_quantize(&ctx->m, block,
934 ctx->is_444 ? (((i >> 1) % 3) < 1 ? 0 : 4): 4 & (2*i),
937 dnxhd_encode_block(ctx, block, last_index, n);
938 // STOP_TIMER("encode_block");
941 if (put_bits_count(&ctx->m.pb) & 31)
942 put_bits(&ctx->m.pb, 32 - (put_bits_count(&ctx->m.pb) & 31), 0);
943 flush_put_bits(&ctx->m.pb);
947 static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx)
951 for (mb_y = 0; mb_y < ctx->m.mb_height; mb_y++) {
953 ctx->slice_offs[mb_y] = offset;
954 ctx->slice_size[mb_y] = 0;
955 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
956 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
957 ctx->slice_size[mb_y] += ctx->mb_bits[mb];
959 ctx->slice_size[mb_y] = (ctx->slice_size[mb_y] + 31) & ~31;
960 ctx->slice_size[mb_y] >>= 3;
961 thread_size = ctx->slice_size[mb_y];
962 offset += thread_size;
966 static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg,
967 int jobnr, int threadnr)
969 DNXHDEncContext *ctx = avctx->priv_data;
970 int mb_y = jobnr, mb_x, x, y;
971 int partial_last_row = (mb_y == ctx->m.mb_height - 1) &&
972 ((avctx->height >> ctx->interlaced) & 0xF);
974 ctx = ctx->thread[threadnr];
975 if (ctx->bit_depth == 8) {
976 uint8_t *pix = ctx->thread[0]->src[0] + ((mb_y << 4) * ctx->m.linesize);
977 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x, pix += 16) {
978 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
982 if (!partial_last_row && mb_x * 16 <= avctx->width - 16 && (avctx->width % 16) == 0) {
983 sum = ctx->m.mpvencdsp.pix_sum(pix, ctx->m.linesize);
984 varc = ctx->m.mpvencdsp.pix_norm1(pix, ctx->m.linesize);
986 int bw = FFMIN(avctx->width - 16 * mb_x, 16);
987 int bh = FFMIN((avctx->height >> ctx->interlaced) - 16 * mb_y, 16);
989 for (y = 0; y < bh; y++) {
990 for (x = 0; x < bw; x++) {
991 uint8_t val = pix[x + y * ctx->m.linesize];
997 varc = (varc - (((unsigned) sum * sum) >> 8) + 128) >> 8;
999 ctx->mb_cmp[mb].value = varc;
1000 ctx->mb_cmp[mb].mb = mb;
1003 const int linesize = ctx->m.linesize >> 1;
1004 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x) {
1005 uint16_t *pix = (uint16_t *)ctx->thread[0]->src[0] +
1006 ((mb_y << 4) * linesize) + (mb_x << 4);
1007 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
1010 int bw = FFMIN(avctx->width - 16 * mb_x, 16);
1011 int bh = FFMIN((avctx->height >> ctx->interlaced) - 16 * mb_y, 16);
1014 // Macroblocks are 16x16 pixels, unlike DCT blocks which are 8x8.
1015 for (i = 0; i < bh; ++i) {
1016 for (j = 0; j < bw; ++j) {
1017 // Turn 16-bit pixels into 10-bit ones.
1018 const int sample = (unsigned) pix[j] >> 6;
1020 sqsum += sample * sample;
1021 // 2^10 * 2^10 * 16 * 16 = 2^28, which is less than INT_MAX
1025 mean = sum >> 8; // 16*16 == 2^8
1026 sqmean = sqsum >> 8;
1027 ctx->mb_cmp[mb].value = sqmean - mean * mean;
1028 ctx->mb_cmp[mb].mb = mb;
1034 static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx)
1036 int lambda, up_step, down_step;
1037 int last_lower = INT_MAX, last_higher = 0;
1040 for (q = 1; q < avctx->qmax; q++) {
1042 avctx->execute2(avctx, dnxhd_calc_bits_thread,
1043 NULL, NULL, ctx->m.mb_height);
1045 up_step = down_step = 2 << LAMBDA_FRAC_BITS;
1046 lambda = ctx->lambda;
1051 if (lambda == last_higher) {
1053 end = 1; // need to set final qscales/bits
1055 for (y = 0; y < ctx->m.mb_height; y++) {
1056 for (x = 0; x < ctx->m.mb_width; x++) {
1057 unsigned min = UINT_MAX;
1059 int mb = y * ctx->m.mb_width + x;
1061 for (q = 1; q < avctx->qmax; q++) {
1062 int i = (q*ctx->m.mb_num) + mb;
1063 unsigned score = ctx->mb_rc[i].bits * lambda +
1064 ((unsigned) ctx->mb_rc[i].ssd << LAMBDA_FRAC_BITS);
1071 bits += ctx->mb_rc[rc].bits;
1072 ctx->mb_qscale[mb] = qscale;
1073 ctx->mb_bits[mb] = ctx->mb_rc[rc].bits;
1075 bits = (bits + 31) & ~31; // padding
1076 if (bits > ctx->frame_bits)
1080 if (bits > ctx->frame_bits)
1081 return AVERROR(EINVAL);
1084 if (bits < ctx->frame_bits) {
1085 last_lower = FFMIN(lambda, last_lower);
1086 if (last_higher != 0)
1087 lambda = (lambda+last_higher)>>1;
1089 lambda -= down_step;
1090 down_step = FFMIN((int64_t)down_step*5, INT_MAX);
1091 up_step = 1<<LAMBDA_FRAC_BITS;
1092 lambda = FFMAX(1, lambda);
1093 if (lambda == last_lower)
1096 last_higher = FFMAX(lambda, last_higher);
1097 if (last_lower != INT_MAX)
1098 lambda = (lambda+last_lower)>>1;
1099 else if ((int64_t)lambda + up_step > INT_MAX)
1100 return AVERROR(EINVAL);
1103 up_step = FFMIN((int64_t)up_step*5, INT_MAX);
1104 down_step = 1<<LAMBDA_FRAC_BITS;
1107 ctx->lambda = lambda;
1111 static int dnxhd_find_qscale(DNXHDEncContext *ctx)
1116 int last_higher = 0;
1117 int last_lower = INT_MAX;
1121 qscale = ctx->qscale;
1124 ctx->qscale = qscale;
1125 // XXX avoid recalculating bits
1126 ctx->m.avctx->execute2(ctx->m.avctx, dnxhd_calc_bits_thread,
1127 NULL, NULL, ctx->m.mb_height);
1128 for (y = 0; y < ctx->m.mb_height; y++) {
1129 for (x = 0; x < ctx->m.mb_width; x++)
1130 bits += ctx->mb_rc[(qscale*ctx->m.mb_num) + (y*ctx->m.mb_width+x)].bits;
1131 bits = (bits+31)&~31; // padding
1132 if (bits > ctx->frame_bits)
1135 if (bits < ctx->frame_bits) {
1138 if (last_higher == qscale - 1) {
1139 qscale = last_higher;
1142 last_lower = FFMIN(qscale, last_lower);
1143 if (last_higher != 0)
1144 qscale = (qscale + last_higher) >> 1;
1146 qscale -= down_step++;
1151 if (last_lower == qscale + 1)
1153 last_higher = FFMAX(qscale, last_higher);
1154 if (last_lower != INT_MAX)
1155 qscale = (qscale + last_lower) >> 1;
1157 qscale += up_step++;
1159 if (qscale >= ctx->m.avctx->qmax)
1160 return AVERROR(EINVAL);
1163 ctx->qscale = qscale;
1167 #define BUCKET_BITS 8
1168 #define RADIX_PASSES 4
1169 #define NBUCKETS (1 << BUCKET_BITS)
1171 static inline int get_bucket(int value, int shift)
1174 value &= NBUCKETS - 1;
1175 return NBUCKETS - 1 - value;
1178 static void radix_count(const RCCMPEntry *data, int size,
1179 int buckets[RADIX_PASSES][NBUCKETS])
1182 memset(buckets, 0, sizeof(buckets[0][0]) * RADIX_PASSES * NBUCKETS);
1183 for (i = 0; i < size; i++) {
1184 int v = data[i].value;
1185 for (j = 0; j < RADIX_PASSES; j++) {
1186 buckets[j][get_bucket(v, 0)]++;
1191 for (j = 0; j < RADIX_PASSES; j++) {
1193 for (i = NBUCKETS - 1; i >= 0; i--)
1194 buckets[j][i] = offset -= buckets[j][i];
1195 av_assert1(!buckets[j][0]);
1199 static void radix_sort_pass(RCCMPEntry *dst, const RCCMPEntry *data,
1200 int size, int buckets[NBUCKETS], int pass)
1202 int shift = pass * BUCKET_BITS;
1204 for (i = 0; i < size; i++) {
1205 int v = get_bucket(data[i].value, shift);
1206 int pos = buckets[v]++;
1211 static void radix_sort(RCCMPEntry *data, RCCMPEntry *tmp, int size)
1213 int buckets[RADIX_PASSES][NBUCKETS];
1214 radix_count(data, size, buckets);
1215 radix_sort_pass(tmp, data, size, buckets[0], 0);
1216 radix_sort_pass(data, tmp, size, buckets[1], 1);
1217 if (buckets[2][NBUCKETS - 1] || buckets[3][NBUCKETS - 1]) {
1218 radix_sort_pass(tmp, data, size, buckets[2], 2);
1219 radix_sort_pass(data, tmp, size, buckets[3], 3);
1223 static int dnxhd_encode_fast(AVCodecContext *avctx, DNXHDEncContext *ctx)
1227 if ((ret = dnxhd_find_qscale(ctx)) < 0)
1229 for (y = 0; y < ctx->m.mb_height; y++) {
1230 for (x = 0; x < ctx->m.mb_width; x++) {
1231 int mb = y * ctx->m.mb_width + x;
1232 int rc = (ctx->qscale * ctx->m.mb_num ) + mb;
1234 ctx->mb_qscale[mb] = ctx->qscale;
1235 ctx->mb_bits[mb] = ctx->mb_rc[rc].bits;
1236 max_bits += ctx->mb_rc[rc].bits;
1238 delta_bits = ctx->mb_rc[rc].bits -
1239 ctx->mb_rc[rc + ctx->m.mb_num].bits;
1240 ctx->mb_cmp[mb].mb = mb;
1241 ctx->mb_cmp[mb].value =
1242 delta_bits ? ((ctx->mb_rc[rc].ssd -
1243 ctx->mb_rc[rc + ctx->m.mb_num].ssd) * 100) /
1245 : INT_MIN; // avoid increasing qscale
1248 max_bits += 31; // worst padding
1252 avctx->execute2(avctx, dnxhd_mb_var_thread,
1253 NULL, NULL, ctx->m.mb_height);
1254 radix_sort(ctx->mb_cmp, ctx->mb_cmp_tmp, ctx->m.mb_num);
1255 for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) {
1256 int mb = ctx->mb_cmp[x].mb;
1257 int rc = (ctx->qscale * ctx->m.mb_num ) + mb;
1258 max_bits -= ctx->mb_rc[rc].bits -
1259 ctx->mb_rc[rc + ctx->m.mb_num].bits;
1260 ctx->mb_qscale[mb] = ctx->qscale + 1;
1261 ctx->mb_bits[mb] = ctx->mb_rc[rc + ctx->m.mb_num].bits;
1267 static void dnxhd_load_picture(DNXHDEncContext *ctx, const AVFrame *frame)
1271 for (i = 0; i < ctx->m.avctx->thread_count; i++) {
1272 ctx->thread[i]->m.linesize = frame->linesize[0] << ctx->interlaced;
1273 ctx->thread[i]->m.uvlinesize = frame->linesize[1] << ctx->interlaced;
1274 ctx->thread[i]->dct_y_offset = ctx->m.linesize *8;
1275 ctx->thread[i]->dct_uv_offset = ctx->m.uvlinesize*8;
1278 #if FF_API_CODED_FRAME
1279 FF_DISABLE_DEPRECATION_WARNINGS
1280 ctx->m.avctx->coded_frame->interlaced_frame = frame->interlaced_frame;
1281 FF_ENABLE_DEPRECATION_WARNINGS
1283 ctx->cur_field = frame->interlaced_frame && !frame->top_field_first;
1286 static int dnxhd_encode_picture(AVCodecContext *avctx, AVPacket *pkt,
1287 const AVFrame *frame, int *got_packet)
1289 DNXHDEncContext *ctx = avctx->priv_data;
1290 int first_field = 1;
1294 if ((ret = ff_alloc_packet2(avctx, pkt, ctx->frame_size, 0)) < 0)
1298 dnxhd_load_picture(ctx, frame);
1301 for (i = 0; i < 3; i++) {
1302 ctx->src[i] = frame->data[i];
1303 if (ctx->interlaced && ctx->cur_field)
1304 ctx->src[i] += frame->linesize[i];
1307 dnxhd_write_header(avctx, buf);
1309 if (avctx->mb_decision == FF_MB_DECISION_RD)
1310 ret = dnxhd_encode_rdo(avctx, ctx);
1312 ret = dnxhd_encode_fast(avctx, ctx);
1314 av_log(avctx, AV_LOG_ERROR,
1315 "picture could not fit ratecontrol constraints, increase qmax\n");
1319 dnxhd_setup_threads_slices(ctx);
1322 for (i = 0; i < ctx->m.mb_height; i++) {
1323 AV_WB32(ctx->msip + i * 4, offset);
1324 offset += ctx->slice_size[i];
1325 av_assert1(!(ctx->slice_size[i] & 3));
1328 avctx->execute2(avctx, dnxhd_encode_thread, buf, NULL, ctx->m.mb_height);
1330 av_assert1(ctx->data_offset + offset + 4 <= ctx->coding_unit_size);
1331 memset(buf + ctx->data_offset + offset, 0,
1332 ctx->coding_unit_size - 4 - offset - ctx->data_offset);
1334 AV_WB32(buf + ctx->coding_unit_size - 4, 0x600DC0DE); // EOF
1336 if (ctx->interlaced && first_field) {
1338 ctx->cur_field ^= 1;
1339 buf += ctx->coding_unit_size;
1340 goto encode_coding_unit;
1343 #if FF_API_CODED_FRAME
1344 FF_DISABLE_DEPRECATION_WARNINGS
1345 avctx->coded_frame->quality = ctx->qscale * FF_QP2LAMBDA;
1346 FF_ENABLE_DEPRECATION_WARNINGS
1349 ff_side_data_set_encoder_stats(pkt, ctx->qscale * FF_QP2LAMBDA, NULL, 0, AV_PICTURE_TYPE_I);
1351 pkt->flags |= AV_PKT_FLAG_KEY;
1356 static av_cold int dnxhd_encode_end(AVCodecContext *avctx)
1358 DNXHDEncContext *ctx = avctx->priv_data;
1361 av_freep(&ctx->orig_vlc_codes);
1362 av_freep(&ctx->orig_vlc_bits);
1363 av_freep(&ctx->run_codes);
1364 av_freep(&ctx->run_bits);
1366 av_freep(&ctx->mb_bits);
1367 av_freep(&ctx->mb_qscale);
1368 av_freep(&ctx->mb_rc);
1369 av_freep(&ctx->mb_cmp);
1370 av_freep(&ctx->mb_cmp_tmp);
1371 av_freep(&ctx->slice_size);
1372 av_freep(&ctx->slice_offs);
1374 av_freep(&ctx->qmatrix_c);
1375 av_freep(&ctx->qmatrix_l);
1376 av_freep(&ctx->qmatrix_c16);
1377 av_freep(&ctx->qmatrix_l16);
1379 if (avctx->active_thread_type == FF_THREAD_SLICE) {
1380 for (i = 1; i < avctx->thread_count; i++)
1381 av_freep(&ctx->thread[i]);
1387 static const AVCodecDefault dnxhd_defaults[] = {
1388 { "qmax", "1024" }, /* Maximum quantization scale factor allowed for VC-3 */
1392 AVCodec ff_dnxhd_encoder = {
1394 .long_name = NULL_IF_CONFIG_SMALL("VC3/DNxHD"),
1395 .type = AVMEDIA_TYPE_VIDEO,
1396 .id = AV_CODEC_ID_DNXHD,
1397 .priv_data_size = sizeof(DNXHDEncContext),
1398 .init = dnxhd_encode_init,
1399 .encode2 = dnxhd_encode_picture,
1400 .close = dnxhd_encode_end,
1401 .capabilities = AV_CODEC_CAP_SLICE_THREADS | AV_CODEC_CAP_FRAME_THREADS | AV_CODEC_CAP_INTRA_ONLY,
1402 .caps_internal = FF_CODEC_CAP_INIT_CLEANUP,
1403 .pix_fmts = (const enum AVPixelFormat[]) {
1405 AV_PIX_FMT_YUV422P10,
1406 AV_PIX_FMT_YUV444P10,
1410 .priv_class = &dnxhd_class,
1411 .defaults = dnxhd_defaults,
1412 .profiles = NULL_IF_CONFIG_SMALL(ff_dnxhd_profiles),