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/mem_internal.h"
29 #include "libavutil/opt.h"
35 #include "mpegvideo.h"
36 #include "pixblockdsp.h"
37 #include "packet_internal.h"
41 // The largest value that will not lead to overflow for 10-bit samples.
42 #define DNX10BIT_QMAT_SHIFT 18
43 #define RC_VARIANCE 1 // use variance or ssd for fast rc
44 #define LAMBDA_FRAC_BITS 10
46 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
47 static const AVOption options[] = {
48 { "nitris_compat", "encode with Avid Nitris compatibility",
49 offsetof(DNXHDEncContext, nitris_compat), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, VE },
50 { "ibias", "intra quant bias",
51 offsetof(DNXHDEncContext, intra_quant_bias), AV_OPT_TYPE_INT,
52 { .i64 = 0 }, INT_MIN, INT_MAX, VE },
53 { "profile", NULL, offsetof(DNXHDEncContext, profile), AV_OPT_TYPE_INT,
54 { .i64 = FF_PROFILE_DNXHD },
55 FF_PROFILE_DNXHD, FF_PROFILE_DNXHR_444, VE, "profile" },
56 { "dnxhd", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_PROFILE_DNXHD },
57 0, 0, VE, "profile" },
58 { "dnxhr_444", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_PROFILE_DNXHR_444 },
59 0, 0, VE, "profile" },
60 { "dnxhr_hqx", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_PROFILE_DNXHR_HQX },
61 0, 0, VE, "profile" },
62 { "dnxhr_hq", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_PROFILE_DNXHR_HQ },
63 0, 0, VE, "profile" },
64 { "dnxhr_sq", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_PROFILE_DNXHR_SQ },
65 0, 0, VE, "profile" },
66 { "dnxhr_lb", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FF_PROFILE_DNXHR_LB },
67 0, 0, VE, "profile" },
71 static const AVClass dnxhd_class = {
72 .class_name = "dnxhd",
73 .item_name = av_default_item_name,
75 .version = LIBAVUTIL_VERSION_INT,
78 static void dnxhd_8bit_get_pixels_8x4_sym(int16_t *av_restrict block,
79 const uint8_t *pixels,
83 for (i = 0; i < 4; i++) {
95 memcpy(block, block - 8, sizeof(*block) * 8);
96 memcpy(block + 8, block - 16, sizeof(*block) * 8);
97 memcpy(block + 16, block - 24, sizeof(*block) * 8);
98 memcpy(block + 24, block - 32, sizeof(*block) * 8);
101 static av_always_inline
102 void dnxhd_10bit_get_pixels_8x4_sym(int16_t *av_restrict block,
103 const uint8_t *pixels,
106 memcpy(block + 0 * 8, pixels + 0 * line_size, 8 * sizeof(*block));
107 memcpy(block + 7 * 8, pixels + 0 * line_size, 8 * sizeof(*block));
108 memcpy(block + 1 * 8, pixels + 1 * line_size, 8 * sizeof(*block));
109 memcpy(block + 6 * 8, pixels + 1 * line_size, 8 * sizeof(*block));
110 memcpy(block + 2 * 8, pixels + 2 * line_size, 8 * sizeof(*block));
111 memcpy(block + 5 * 8, pixels + 2 * line_size, 8 * sizeof(*block));
112 memcpy(block + 3 * 8, pixels + 3 * line_size, 8 * sizeof(*block));
113 memcpy(block + 4 * 8, pixels + 3 * line_size, 8 * sizeof(*block));
116 static int dnxhd_10bit_dct_quantize_444(MpegEncContext *ctx, int16_t *block,
117 int n, int qscale, int *overflow)
119 int i, j, level, last_non_zero, start_i;
121 const uint8_t *scantable= ctx->intra_scantable.scantable;
124 unsigned int threshold1, threshold2;
126 ctx->fdsp.fdct(block);
128 block[0] = (block[0] + 2) >> 2;
131 qmat = n < 4 ? ctx->q_intra_matrix[qscale] : ctx->q_chroma_intra_matrix[qscale];
132 bias= ctx->intra_quant_bias * (1 << (16 - 8));
133 threshold1 = (1 << 16) - bias - 1;
134 threshold2 = (threshold1 << 1);
136 for (i = 63; i >= start_i; i--) {
138 level = block[j] * qmat[j];
140 if (((unsigned)(level + threshold1)) > threshold2) {
148 for (i = start_i; i <= last_non_zero; i++) {
150 level = block[j] * qmat[j];
152 if (((unsigned)(level + threshold1)) > threshold2) {
154 level = (bias + level) >> 16;
157 level = (bias - level) >> 16;
165 *overflow = ctx->max_qcoeff < max; //overflow might have happened
167 /* we need this permutation so that we correct the IDCT, we only permute the !=0 elements */
168 if (ctx->idsp.perm_type != FF_IDCT_PERM_NONE)
169 ff_block_permute(block, ctx->idsp.idct_permutation,
170 scantable, last_non_zero);
172 return last_non_zero;
175 static int dnxhd_10bit_dct_quantize(MpegEncContext *ctx, int16_t *block,
176 int n, int qscale, int *overflow)
178 const uint8_t *scantable= ctx->intra_scantable.scantable;
179 const int *qmat = n<4 ? ctx->q_intra_matrix[qscale] : ctx->q_chroma_intra_matrix[qscale];
180 int last_non_zero = 0;
183 ctx->fdsp.fdct(block);
185 // Divide by 4 with rounding, to compensate scaling of DCT coefficients
186 block[0] = (block[0] + 2) >> 2;
188 for (i = 1; i < 64; ++i) {
189 int j = scantable[i];
190 int sign = FF_SIGNBIT(block[j]);
191 int level = (block[j] ^ sign) - sign;
192 level = level * qmat[j] >> DNX10BIT_QMAT_SHIFT;
193 block[j] = (level ^ sign) - sign;
198 /* we need this permutation so that we correct the IDCT, we only permute the !=0 elements */
199 if (ctx->idsp.perm_type != FF_IDCT_PERM_NONE)
200 ff_block_permute(block, ctx->idsp.idct_permutation,
201 scantable, last_non_zero);
203 return last_non_zero;
206 static av_cold int dnxhd_init_vlc(DNXHDEncContext *ctx)
208 int i, j, level, run;
209 int max_level = 1 << (ctx->bit_depth + 2);
211 if (!FF_ALLOCZ_TYPED_ARRAY(ctx->orig_vlc_codes, max_level * 4) ||
212 !FF_ALLOCZ_TYPED_ARRAY(ctx->orig_vlc_bits, max_level * 4) ||
213 !(ctx->run_codes = av_mallocz(63 * 2)) ||
214 !(ctx->run_bits = av_mallocz(63)))
215 return AVERROR(ENOMEM);
216 ctx->vlc_codes = ctx->orig_vlc_codes + max_level * 2;
217 ctx->vlc_bits = ctx->orig_vlc_bits + max_level * 2;
218 for (level = -max_level; level < max_level; level++) {
219 for (run = 0; run < 2; run++) {
220 int index = level * (1 << 1) | run;
221 int sign, offset = 0, alevel = level;
223 MASK_ABS(sign, alevel);
225 offset = (alevel - 1) >> 6;
226 alevel -= offset << 6;
228 for (j = 0; j < 257; j++) {
229 if (ctx->cid_table->ac_info[2*j+0] >> 1 == alevel &&
230 (!offset || (ctx->cid_table->ac_info[2*j+1] & 1) && offset) &&
231 (!run || (ctx->cid_table->ac_info[2*j+1] & 2) && run)) {
232 av_assert1(!ctx->vlc_codes[index]);
234 ctx->vlc_codes[index] =
235 (ctx->cid_table->ac_codes[j] << 1) | (sign & 1);
236 ctx->vlc_bits[index] = ctx->cid_table->ac_bits[j] + 1;
238 ctx->vlc_codes[index] = ctx->cid_table->ac_codes[j];
239 ctx->vlc_bits[index] = ctx->cid_table->ac_bits[j];
244 av_assert0(!alevel || j < 257);
246 ctx->vlc_codes[index] =
247 (ctx->vlc_codes[index] << ctx->cid_table->index_bits) | offset;
248 ctx->vlc_bits[index] += ctx->cid_table->index_bits;
252 for (i = 0; i < 62; i++) {
253 int run = ctx->cid_table->run[i];
254 av_assert0(run < 63);
255 ctx->run_codes[run] = ctx->cid_table->run_codes[i];
256 ctx->run_bits[run] = ctx->cid_table->run_bits[i];
261 static av_cold int dnxhd_init_qmat(DNXHDEncContext *ctx, int lbias, int cbias)
263 // init first elem to 1 to avoid div by 0 in convert_matrix
264 uint16_t weight_matrix[64] = { 1, }; // convert_matrix needs uint16_t*
266 const uint8_t *luma_weight_table = ctx->cid_table->luma_weight;
267 const uint8_t *chroma_weight_table = ctx->cid_table->chroma_weight;
269 if (!FF_ALLOCZ_TYPED_ARRAY(ctx->qmatrix_l, ctx->m.avctx->qmax + 1) ||
270 !FF_ALLOCZ_TYPED_ARRAY(ctx->qmatrix_c, ctx->m.avctx->qmax + 1) ||
271 !FF_ALLOCZ_TYPED_ARRAY(ctx->qmatrix_l16, ctx->m.avctx->qmax + 1) ||
272 !FF_ALLOCZ_TYPED_ARRAY(ctx->qmatrix_c16, ctx->m.avctx->qmax + 1))
273 return AVERROR(ENOMEM);
275 if (ctx->bit_depth == 8) {
276 for (i = 1; i < 64; i++) {
277 int j = ctx->m.idsp.idct_permutation[ff_zigzag_direct[i]];
278 weight_matrix[j] = ctx->cid_table->luma_weight[i];
280 ff_convert_matrix(&ctx->m, ctx->qmatrix_l, ctx->qmatrix_l16,
281 weight_matrix, ctx->intra_quant_bias, 1,
282 ctx->m.avctx->qmax, 1);
283 for (i = 1; i < 64; i++) {
284 int j = ctx->m.idsp.idct_permutation[ff_zigzag_direct[i]];
285 weight_matrix[j] = ctx->cid_table->chroma_weight[i];
287 ff_convert_matrix(&ctx->m, ctx->qmatrix_c, ctx->qmatrix_c16,
288 weight_matrix, ctx->intra_quant_bias, 1,
289 ctx->m.avctx->qmax, 1);
291 for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
292 for (i = 0; i < 64; i++) {
293 ctx->qmatrix_l[qscale][i] <<= 2;
294 ctx->qmatrix_c[qscale][i] <<= 2;
295 ctx->qmatrix_l16[qscale][0][i] <<= 2;
296 ctx->qmatrix_l16[qscale][1][i] <<= 2;
297 ctx->qmatrix_c16[qscale][0][i] <<= 2;
298 ctx->qmatrix_c16[qscale][1][i] <<= 2;
303 for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
304 for (i = 1; i < 64; i++) {
305 int j = ff_zigzag_direct[i];
307 /* The quantization formula from the VC-3 standard is:
308 * quantized = sign(block[i]) * floor(abs(block[i]/s) * p /
309 * (qscale * weight_table[i]))
310 * Where p is 32 for 8-bit samples and 8 for 10-bit ones.
311 * The s factor compensates scaling of DCT coefficients done by
312 * the DCT routines, and therefore is not present in standard.
313 * It's 8 for 8-bit samples and 4 for 10-bit ones.
314 * We want values of ctx->qtmatrix_l and ctx->qtmatrix_r to be:
315 * ((1 << DNX10BIT_QMAT_SHIFT) * (p / s)) /
316 * (qscale * weight_table[i])
317 * For 10-bit samples, p / s == 2 */
318 ctx->qmatrix_l[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) /
319 (qscale * luma_weight_table[i]);
320 ctx->qmatrix_c[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) /
321 (qscale * chroma_weight_table[i]);
326 ctx->m.q_chroma_intra_matrix16 = ctx->qmatrix_c16;
327 ctx->m.q_chroma_intra_matrix = ctx->qmatrix_c;
328 ctx->m.q_intra_matrix16 = ctx->qmatrix_l16;
329 ctx->m.q_intra_matrix = ctx->qmatrix_l;
334 static av_cold int dnxhd_init_rc(DNXHDEncContext *ctx)
336 if (!FF_ALLOCZ_TYPED_ARRAY(ctx->mb_rc, (ctx->m.avctx->qmax + 1) * ctx->m.mb_num))
337 return AVERROR(ENOMEM);
339 if (ctx->m.avctx->mb_decision != FF_MB_DECISION_RD) {
340 if (!FF_ALLOCZ_TYPED_ARRAY(ctx->mb_cmp, ctx->m.mb_num) ||
341 !FF_ALLOCZ_TYPED_ARRAY(ctx->mb_cmp_tmp, ctx->m.mb_num))
342 return AVERROR(ENOMEM);
344 ctx->frame_bits = (ctx->coding_unit_size -
345 ctx->data_offset - 4 - ctx->min_padding) * 8;
347 ctx->lambda = 2 << LAMBDA_FRAC_BITS; // qscale 2
351 static av_cold int dnxhd_encode_init(AVCodecContext *avctx)
353 DNXHDEncContext *ctx = avctx->priv_data;
356 switch (avctx->pix_fmt) {
357 case AV_PIX_FMT_YUV422P:
360 case AV_PIX_FMT_YUV422P10:
361 case AV_PIX_FMT_YUV444P10:
362 case AV_PIX_FMT_GBRP10:
366 av_log(avctx, AV_LOG_ERROR,
367 "pixel format is incompatible with DNxHD\n");
368 return AVERROR(EINVAL);
371 if ((ctx->profile == FF_PROFILE_DNXHR_444 && (avctx->pix_fmt != AV_PIX_FMT_YUV444P10 &&
372 avctx->pix_fmt != AV_PIX_FMT_GBRP10)) ||
373 (ctx->profile != FF_PROFILE_DNXHR_444 && (avctx->pix_fmt == AV_PIX_FMT_YUV444P10 ||
374 avctx->pix_fmt == AV_PIX_FMT_GBRP10))) {
375 av_log(avctx, AV_LOG_ERROR,
376 "pixel format is incompatible with DNxHD profile\n");
377 return AVERROR(EINVAL);
380 if (ctx->profile == FF_PROFILE_DNXHR_HQX && avctx->pix_fmt != AV_PIX_FMT_YUV422P10) {
381 av_log(avctx, AV_LOG_ERROR,
382 "pixel format is incompatible with DNxHR HQX profile\n");
383 return AVERROR(EINVAL);
386 if ((ctx->profile == FF_PROFILE_DNXHR_LB ||
387 ctx->profile == FF_PROFILE_DNXHR_SQ ||
388 ctx->profile == FF_PROFILE_DNXHR_HQ) && avctx->pix_fmt != AV_PIX_FMT_YUV422P) {
389 av_log(avctx, AV_LOG_ERROR,
390 "pixel format is incompatible with DNxHR LB/SQ/HQ profile\n");
391 return AVERROR(EINVAL);
394 ctx->is_444 = ctx->profile == FF_PROFILE_DNXHR_444;
395 avctx->profile = ctx->profile;
396 ctx->cid = ff_dnxhd_find_cid(avctx, ctx->bit_depth);
398 av_log(avctx, AV_LOG_ERROR,
399 "video parameters incompatible with DNxHD. Valid DNxHD profiles:\n");
400 ff_dnxhd_print_profiles(avctx, AV_LOG_ERROR);
401 return AVERROR(EINVAL);
403 av_log(avctx, AV_LOG_DEBUG, "cid %d\n", ctx->cid);
405 if (ctx->cid >= 1270 && ctx->cid <= 1274)
406 avctx->codec_tag = MKTAG('A','V','d','h');
408 if (avctx->width < 256 || avctx->height < 120) {
409 av_log(avctx, AV_LOG_ERROR,
410 "Input dimensions too small, input must be at least 256x120\n");
411 return AVERROR(EINVAL);
414 index = ff_dnxhd_get_cid_table(ctx->cid);
415 av_assert0(index >= 0);
417 ctx->cid_table = &ff_dnxhd_cid_table[index];
419 ctx->m.avctx = avctx;
423 avctx->bits_per_raw_sample = ctx->bit_depth;
425 ff_blockdsp_init(&ctx->bdsp, avctx);
426 ff_fdctdsp_init(&ctx->m.fdsp, avctx);
427 ff_mpv_idct_init(&ctx->m);
428 ff_mpegvideoencdsp_init(&ctx->m.mpvencdsp, avctx);
429 ff_pixblockdsp_init(&ctx->m.pdsp, avctx);
430 ff_dct_encode_init(&ctx->m);
432 if (ctx->profile != FF_PROFILE_DNXHD)
433 ff_videodsp_init(&ctx->m.vdsp, ctx->bit_depth);
435 if (!ctx->m.dct_quantize)
436 ctx->m.dct_quantize = ff_dct_quantize_c;
438 if (ctx->is_444 || ctx->profile == FF_PROFILE_DNXHR_HQX) {
439 ctx->m.dct_quantize = dnxhd_10bit_dct_quantize_444;
440 ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym;
441 ctx->block_width_l2 = 4;
442 } else if (ctx->bit_depth == 10) {
443 ctx->m.dct_quantize = dnxhd_10bit_dct_quantize;
444 ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym;
445 ctx->block_width_l2 = 4;
447 ctx->get_pixels_8x4_sym = dnxhd_8bit_get_pixels_8x4_sym;
448 ctx->block_width_l2 = 3;
452 ff_dnxhdenc_init_x86(ctx);
454 ctx->m.mb_height = (avctx->height + 15) / 16;
455 ctx->m.mb_width = (avctx->width + 15) / 16;
457 if (avctx->flags & AV_CODEC_FLAG_INTERLACED_DCT) {
459 ctx->m.mb_height /= 2;
462 if (ctx->interlaced && ctx->profile != FF_PROFILE_DNXHD) {
463 av_log(avctx, AV_LOG_ERROR,
464 "Interlaced encoding is not supported for DNxHR profiles.\n");
465 return AVERROR(EINVAL);
468 ctx->m.mb_num = ctx->m.mb_height * ctx->m.mb_width;
470 if (ctx->cid_table->frame_size == DNXHD_VARIABLE) {
471 ctx->frame_size = avpriv_dnxhd_get_hr_frame_size(ctx->cid,
472 avctx->width, avctx->height);
473 av_assert0(ctx->frame_size >= 0);
474 ctx->coding_unit_size = ctx->frame_size;
476 ctx->frame_size = ctx->cid_table->frame_size;
477 ctx->coding_unit_size = ctx->cid_table->coding_unit_size;
480 if (ctx->m.mb_height > 68)
481 ctx->data_offset = 0x170 + (ctx->m.mb_height << 2);
483 ctx->data_offset = 0x280;
485 // XXX tune lbias/cbias
486 if ((ret = dnxhd_init_qmat(ctx, ctx->intra_quant_bias, 0)) < 0)
489 /* Avid Nitris hardware decoder requires a minimum amount of padding
490 * in the coding unit payload */
491 if (ctx->nitris_compat)
492 ctx->min_padding = 1600;
494 if ((ret = dnxhd_init_vlc(ctx)) < 0)
496 if ((ret = dnxhd_init_rc(ctx)) < 0)
499 if (!FF_ALLOCZ_TYPED_ARRAY(ctx->slice_size, ctx->m.mb_height) ||
500 !FF_ALLOCZ_TYPED_ARRAY(ctx->slice_offs, ctx->m.mb_height) ||
501 !FF_ALLOCZ_TYPED_ARRAY(ctx->mb_bits, ctx->m.mb_num) ||
502 !FF_ALLOCZ_TYPED_ARRAY(ctx->mb_qscale, ctx->m.mb_num))
503 return AVERROR(ENOMEM);
504 #if FF_API_CODED_FRAME
505 FF_DISABLE_DEPRECATION_WARNINGS
506 avctx->coded_frame->key_frame = 1;
507 avctx->coded_frame->pict_type = AV_PICTURE_TYPE_I;
508 FF_ENABLE_DEPRECATION_WARNINGS
511 if (avctx->active_thread_type == FF_THREAD_SLICE) {
512 if (avctx->thread_count > MAX_THREADS) {
513 av_log(avctx, AV_LOG_ERROR, "too many threads\n");
514 return AVERROR(EINVAL);
518 if (avctx->qmax <= 1) {
519 av_log(avctx, AV_LOG_ERROR, "qmax must be at least 2\n");
520 return AVERROR(EINVAL);
523 ctx->thread[0] = ctx;
524 if (avctx->active_thread_type == FF_THREAD_SLICE) {
525 for (i = 1; i < avctx->thread_count; i++) {
526 ctx->thread[i] = av_malloc(sizeof(DNXHDEncContext));
528 return AVERROR(ENOMEM);
529 memcpy(ctx->thread[i], ctx, sizeof(DNXHDEncContext));
536 static int dnxhd_write_header(AVCodecContext *avctx, uint8_t *buf)
538 DNXHDEncContext *ctx = avctx->priv_data;
540 memset(buf, 0, ctx->data_offset);
543 AV_WB16(buf + 0x02, ctx->data_offset);
544 if (ctx->cid >= 1270 && ctx->cid <= 1274)
549 buf[5] = ctx->interlaced ? ctx->cur_field + 2 : 0x01;
550 buf[6] = 0x80; // crc flag off
551 buf[7] = 0xa0; // reserved
552 AV_WB16(buf + 0x18, avctx->height >> ctx->interlaced); // ALPF
553 AV_WB16(buf + 0x1a, avctx->width); // SPL
554 AV_WB16(buf + 0x1d, avctx->height >> ctx->interlaced); // NAL
556 buf[0x21] = ctx->bit_depth == 10 ? 0x58 : 0x38;
557 buf[0x22] = 0x88 + (ctx->interlaced << 2);
558 AV_WB32(buf + 0x28, ctx->cid); // CID
559 buf[0x2c] = (!ctx->interlaced << 7) | (ctx->is_444 << 6) | (avctx->pix_fmt == AV_PIX_FMT_YUV444P10);
561 buf[0x5f] = 0x01; // UDL
563 buf[0x167] = 0x02; // reserved
564 AV_WB16(buf + 0x16a, ctx->m.mb_height * 4 + 4); // MSIPS
565 AV_WB16(buf + 0x16c, ctx->m.mb_height); // Ns
566 buf[0x16f] = 0x10; // reserved
568 ctx->msip = buf + 0x170;
572 static av_always_inline void dnxhd_encode_dc(DNXHDEncContext *ctx, int diff)
576 nbits = av_log2_16bit(-2 * diff);
579 nbits = av_log2_16bit(2 * diff);
581 put_bits(&ctx->m.pb, ctx->cid_table->dc_bits[nbits] + nbits,
582 (ctx->cid_table->dc_codes[nbits] << nbits) +
583 av_mod_uintp2(diff, nbits));
586 static av_always_inline
587 void dnxhd_encode_block(DNXHDEncContext *ctx, int16_t *block,
588 int last_index, int n)
590 int last_non_zero = 0;
593 dnxhd_encode_dc(ctx, block[0] - ctx->m.last_dc[n]);
594 ctx->m.last_dc[n] = block[0];
596 for (i = 1; i <= last_index; i++) {
597 j = ctx->m.intra_scantable.permutated[i];
600 int run_level = i - last_non_zero - 1;
601 int rlevel = slevel * (1 << 1) | !!run_level;
602 put_bits(&ctx->m.pb, ctx->vlc_bits[rlevel], ctx->vlc_codes[rlevel]);
604 put_bits(&ctx->m.pb, ctx->run_bits[run_level],
605 ctx->run_codes[run_level]);
609 put_bits(&ctx->m.pb, ctx->vlc_bits[0], ctx->vlc_codes[0]); // EOB
612 static av_always_inline
613 void dnxhd_unquantize_c(DNXHDEncContext *ctx, int16_t *block, int n,
614 int qscale, int last_index)
616 const uint8_t *weight_matrix;
621 weight_matrix = ((n % 6) < 2) ? ctx->cid_table->luma_weight
622 : ctx->cid_table->chroma_weight;
624 weight_matrix = (n & 2) ? ctx->cid_table->chroma_weight
625 : ctx->cid_table->luma_weight;
628 for (i = 1; i <= last_index; i++) {
629 int j = ctx->m.intra_scantable.permutated[i];
633 level = (1 - 2 * level) * qscale * weight_matrix[i];
634 if (ctx->bit_depth == 10) {
635 if (weight_matrix[i] != 8)
639 if (weight_matrix[i] != 32)
645 level = (2 * level + 1) * qscale * weight_matrix[i];
646 if (ctx->bit_depth == 10) {
647 if (weight_matrix[i] != 8)
651 if (weight_matrix[i] != 32)
661 static av_always_inline int dnxhd_ssd_block(int16_t *qblock, int16_t *block)
665 for (i = 0; i < 64; i++)
666 score += (block[i] - qblock[i]) * (block[i] - qblock[i]);
670 static av_always_inline
671 int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, int16_t *block, int last_index)
673 int last_non_zero = 0;
676 for (i = 1; i <= last_index; i++) {
677 j = ctx->m.intra_scantable.permutated[i];
680 int run_level = i - last_non_zero - 1;
681 bits += ctx->vlc_bits[level * (1 << 1) |
682 !!run_level] + ctx->run_bits[run_level];
689 static av_always_inline
690 void dnxhd_get_blocks(DNXHDEncContext *ctx, int mb_x, int mb_y)
692 const int bs = ctx->block_width_l2;
693 const int bw = 1 << bs;
694 int dct_y_offset = ctx->dct_y_offset;
695 int dct_uv_offset = ctx->dct_uv_offset;
696 int linesize = ctx->m.linesize;
697 int uvlinesize = ctx->m.uvlinesize;
698 const uint8_t *ptr_y = ctx->thread[0]->src[0] +
699 ((mb_y << 4) * ctx->m.linesize) + (mb_x << bs + 1);
700 const uint8_t *ptr_u = ctx->thread[0]->src[1] +
701 ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs + ctx->is_444);
702 const uint8_t *ptr_v = ctx->thread[0]->src[2] +
703 ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs + ctx->is_444);
704 PixblockDSPContext *pdsp = &ctx->m.pdsp;
705 VideoDSPContext *vdsp = &ctx->m.vdsp;
707 if (ctx->bit_depth != 10 && vdsp->emulated_edge_mc && ((mb_x << 4) + 16 > ctx->m.avctx->width ||
708 (mb_y << 4) + 16 > ctx->m.avctx->height)) {
709 int y_w = ctx->m.avctx->width - (mb_x << 4);
710 int y_h = ctx->m.avctx->height - (mb_y << 4);
711 int uv_w = (y_w + 1) / 2;
716 vdsp->emulated_edge_mc(&ctx->edge_buf_y[0], ptr_y,
717 linesize, ctx->m.linesize,
720 vdsp->emulated_edge_mc(&ctx->edge_buf_uv[0][0], ptr_u,
721 uvlinesize, ctx->m.uvlinesize,
724 vdsp->emulated_edge_mc(&ctx->edge_buf_uv[1][0], ptr_v,
725 uvlinesize, ctx->m.uvlinesize,
729 dct_y_offset = bw * linesize;
730 dct_uv_offset = bw * uvlinesize;
731 ptr_y = &ctx->edge_buf_y[0];
732 ptr_u = &ctx->edge_buf_uv[0][0];
733 ptr_v = &ctx->edge_buf_uv[1][0];
734 } else if (ctx->bit_depth == 10 && vdsp->emulated_edge_mc && ((mb_x << 4) + 16 > ctx->m.avctx->width ||
735 (mb_y << 4) + 16 > ctx->m.avctx->height)) {
736 int y_w = ctx->m.avctx->width - (mb_x << 4);
737 int y_h = ctx->m.avctx->height - (mb_y << 4);
738 int uv_w = ctx->is_444 ? y_w : (y_w + 1) / 2;
741 uvlinesize = 16 + 16 * ctx->is_444;
743 vdsp->emulated_edge_mc(&ctx->edge_buf_y[0], ptr_y,
744 linesize, ctx->m.linesize,
747 vdsp->emulated_edge_mc(&ctx->edge_buf_uv[0][0], ptr_u,
748 uvlinesize, ctx->m.uvlinesize,
751 vdsp->emulated_edge_mc(&ctx->edge_buf_uv[1][0], ptr_v,
752 uvlinesize, ctx->m.uvlinesize,
756 dct_y_offset = bw * linesize / 2;
757 dct_uv_offset = bw * uvlinesize / 2;
758 ptr_y = &ctx->edge_buf_y[0];
759 ptr_u = &ctx->edge_buf_uv[0][0];
760 ptr_v = &ctx->edge_buf_uv[1][0];
764 pdsp->get_pixels(ctx->blocks[0], ptr_y, linesize);
765 pdsp->get_pixels(ctx->blocks[1], ptr_y + bw, linesize);
766 pdsp->get_pixels(ctx->blocks[2], ptr_u, uvlinesize);
767 pdsp->get_pixels(ctx->blocks[3], ptr_v, uvlinesize);
769 if (mb_y + 1 == ctx->m.mb_height && ctx->m.avctx->height == 1080) {
770 if (ctx->interlaced) {
771 ctx->get_pixels_8x4_sym(ctx->blocks[4],
772 ptr_y + dct_y_offset,
774 ctx->get_pixels_8x4_sym(ctx->blocks[5],
775 ptr_y + dct_y_offset + bw,
777 ctx->get_pixels_8x4_sym(ctx->blocks[6],
778 ptr_u + dct_uv_offset,
780 ctx->get_pixels_8x4_sym(ctx->blocks[7],
781 ptr_v + dct_uv_offset,
784 ctx->bdsp.clear_block(ctx->blocks[4]);
785 ctx->bdsp.clear_block(ctx->blocks[5]);
786 ctx->bdsp.clear_block(ctx->blocks[6]);
787 ctx->bdsp.clear_block(ctx->blocks[7]);
790 pdsp->get_pixels(ctx->blocks[4],
791 ptr_y + dct_y_offset, linesize);
792 pdsp->get_pixels(ctx->blocks[5],
793 ptr_y + dct_y_offset + bw, linesize);
794 pdsp->get_pixels(ctx->blocks[6],
795 ptr_u + dct_uv_offset, uvlinesize);
796 pdsp->get_pixels(ctx->blocks[7],
797 ptr_v + dct_uv_offset, uvlinesize);
800 pdsp->get_pixels(ctx->blocks[0], ptr_y, linesize);
801 pdsp->get_pixels(ctx->blocks[1], ptr_y + bw, linesize);
802 pdsp->get_pixels(ctx->blocks[6], ptr_y + dct_y_offset, linesize);
803 pdsp->get_pixels(ctx->blocks[7], ptr_y + dct_y_offset + bw, linesize);
805 pdsp->get_pixels(ctx->blocks[2], ptr_u, uvlinesize);
806 pdsp->get_pixels(ctx->blocks[3], ptr_u + bw, uvlinesize);
807 pdsp->get_pixels(ctx->blocks[8], ptr_u + dct_uv_offset, uvlinesize);
808 pdsp->get_pixels(ctx->blocks[9], ptr_u + dct_uv_offset + bw, uvlinesize);
810 pdsp->get_pixels(ctx->blocks[4], ptr_v, uvlinesize);
811 pdsp->get_pixels(ctx->blocks[5], ptr_v + bw, uvlinesize);
812 pdsp->get_pixels(ctx->blocks[10], ptr_v + dct_uv_offset, uvlinesize);
813 pdsp->get_pixels(ctx->blocks[11], ptr_v + dct_uv_offset + bw, uvlinesize);
817 static av_always_inline
818 int dnxhd_switch_matrix(DNXHDEncContext *ctx, int i)
825 const static uint8_t component[8]={0,0,1,2,0,0,1,2};
831 static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg,
832 int jobnr, int threadnr)
834 DNXHDEncContext *ctx = avctx->priv_data;
835 int mb_y = jobnr, mb_x;
836 int qscale = ctx->qscale;
837 LOCAL_ALIGNED_16(int16_t, block, [64]);
838 ctx = ctx->thread[threadnr];
842 ctx->m.last_dc[2] = 1 << (ctx->bit_depth + 2);
844 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
845 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
851 dnxhd_get_blocks(ctx, mb_x, mb_y);
853 for (i = 0; i < 8 + 4 * ctx->is_444; i++) {
854 int16_t *src_block = ctx->blocks[i];
855 int overflow, nbits, diff, last_index;
856 int n = dnxhd_switch_matrix(ctx, i);
858 memcpy(block, src_block, 64 * sizeof(*block));
859 last_index = ctx->m.dct_quantize(&ctx->m, block,
860 ctx->is_444 ? 4 * (n > 0): 4 & (2*i),
862 ac_bits += dnxhd_calc_ac_bits(ctx, block, last_index);
864 diff = block[0] - ctx->m.last_dc[n];
866 nbits = av_log2_16bit(-2 * diff);
868 nbits = av_log2_16bit(2 * diff);
870 av_assert1(nbits < ctx->bit_depth + 4);
871 dc_bits += ctx->cid_table->dc_bits[nbits] + nbits;
873 ctx->m.last_dc[n] = block[0];
875 if (avctx->mb_decision == FF_MB_DECISION_RD || !RC_VARIANCE) {
876 dnxhd_unquantize_c(ctx, block, i, qscale, last_index);
877 ctx->m.idsp.idct(block);
878 ssd += dnxhd_ssd_block(block, src_block);
881 ctx->mb_rc[(qscale * ctx->m.mb_num) + mb].ssd = ssd;
882 ctx->mb_rc[(qscale * ctx->m.mb_num) + mb].bits = ac_bits + dc_bits + 12 +
883 (1 + ctx->is_444) * 8 * ctx->vlc_bits[0];
888 static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg,
889 int jobnr, int threadnr)
891 DNXHDEncContext *ctx = avctx->priv_data;
892 int mb_y = jobnr, mb_x;
893 ctx = ctx->thread[threadnr];
894 init_put_bits(&ctx->m.pb, (uint8_t *)arg + ctx->data_offset + ctx->slice_offs[jobnr],
895 ctx->slice_size[jobnr]);
899 ctx->m.last_dc[2] = 1 << (ctx->bit_depth + 2);
900 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
901 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
902 int qscale = ctx->mb_qscale[mb];
905 put_bits(&ctx->m.pb, 11, qscale);
906 put_bits(&ctx->m.pb, 1, avctx->pix_fmt == AV_PIX_FMT_YUV444P10);
908 dnxhd_get_blocks(ctx, mb_x, mb_y);
910 for (i = 0; i < 8 + 4 * ctx->is_444; i++) {
911 int16_t *block = ctx->blocks[i];
912 int overflow, n = dnxhd_switch_matrix(ctx, i);
913 int last_index = ctx->m.dct_quantize(&ctx->m, block,
914 ctx->is_444 ? (((i >> 1) % 3) < 1 ? 0 : 4): 4 & (2*i),
917 dnxhd_encode_block(ctx, block, last_index, n);
920 if (put_bits_count(&ctx->m.pb) & 31)
921 put_bits(&ctx->m.pb, 32 - (put_bits_count(&ctx->m.pb) & 31), 0);
922 flush_put_bits(&ctx->m.pb);
926 static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx)
930 for (mb_y = 0; mb_y < ctx->m.mb_height; mb_y++) {
932 ctx->slice_offs[mb_y] = offset;
933 ctx->slice_size[mb_y] = 0;
934 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
935 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
936 ctx->slice_size[mb_y] += ctx->mb_bits[mb];
938 ctx->slice_size[mb_y] = (ctx->slice_size[mb_y] + 31) & ~31;
939 ctx->slice_size[mb_y] >>= 3;
940 thread_size = ctx->slice_size[mb_y];
941 offset += thread_size;
945 static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg,
946 int jobnr, int threadnr)
948 DNXHDEncContext *ctx = avctx->priv_data;
949 int mb_y = jobnr, mb_x, x, y;
950 int partial_last_row = (mb_y == ctx->m.mb_height - 1) &&
951 ((avctx->height >> ctx->interlaced) & 0xF);
953 ctx = ctx->thread[threadnr];
954 if (ctx->bit_depth == 8) {
955 uint8_t *pix = ctx->thread[0]->src[0] + ((mb_y << 4) * ctx->m.linesize);
956 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x, pix += 16) {
957 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
961 if (!partial_last_row && mb_x * 16 <= avctx->width - 16 && (avctx->width % 16) == 0) {
962 sum = ctx->m.mpvencdsp.pix_sum(pix, ctx->m.linesize);
963 varc = ctx->m.mpvencdsp.pix_norm1(pix, ctx->m.linesize);
965 int bw = FFMIN(avctx->width - 16 * mb_x, 16);
966 int bh = FFMIN((avctx->height >> ctx->interlaced) - 16 * mb_y, 16);
968 for (y = 0; y < bh; y++) {
969 for (x = 0; x < bw; x++) {
970 uint8_t val = pix[x + y * ctx->m.linesize];
976 varc = (varc - (((unsigned) sum * sum) >> 8) + 128) >> 8;
978 ctx->mb_cmp[mb].value = varc;
979 ctx->mb_cmp[mb].mb = mb;
982 const int linesize = ctx->m.linesize >> 1;
983 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x) {
984 uint16_t *pix = (uint16_t *)ctx->thread[0]->src[0] +
985 ((mb_y << 4) * linesize) + (mb_x << 4);
986 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
989 int bw = FFMIN(avctx->width - 16 * mb_x, 16);
990 int bh = FFMIN((avctx->height >> ctx->interlaced) - 16 * mb_y, 16);
993 // Macroblocks are 16x16 pixels, unlike DCT blocks which are 8x8.
994 for (i = 0; i < bh; ++i) {
995 for (j = 0; j < bw; ++j) {
996 // Turn 16-bit pixels into 10-bit ones.
997 const int sample = (unsigned) pix[j] >> 6;
999 sqsum += sample * sample;
1000 // 2^10 * 2^10 * 16 * 16 = 2^28, which is less than INT_MAX
1004 mean = sum >> 8; // 16*16 == 2^8
1005 sqmean = sqsum >> 8;
1006 ctx->mb_cmp[mb].value = sqmean - mean * mean;
1007 ctx->mb_cmp[mb].mb = mb;
1013 static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx)
1015 int lambda, up_step, down_step;
1016 int last_lower = INT_MAX, last_higher = 0;
1019 for (q = 1; q < avctx->qmax; q++) {
1021 avctx->execute2(avctx, dnxhd_calc_bits_thread,
1022 NULL, NULL, ctx->m.mb_height);
1024 up_step = down_step = 2 << LAMBDA_FRAC_BITS;
1025 lambda = ctx->lambda;
1030 if (lambda == last_higher) {
1032 end = 1; // need to set final qscales/bits
1034 for (y = 0; y < ctx->m.mb_height; y++) {
1035 for (x = 0; x < ctx->m.mb_width; x++) {
1036 unsigned min = UINT_MAX;
1038 int mb = y * ctx->m.mb_width + x;
1040 for (q = 1; q < avctx->qmax; q++) {
1041 int i = (q*ctx->m.mb_num) + mb;
1042 unsigned score = ctx->mb_rc[i].bits * lambda +
1043 ((unsigned) ctx->mb_rc[i].ssd << LAMBDA_FRAC_BITS);
1050 bits += ctx->mb_rc[rc].bits;
1051 ctx->mb_qscale[mb] = qscale;
1052 ctx->mb_bits[mb] = ctx->mb_rc[rc].bits;
1054 bits = (bits + 31) & ~31; // padding
1055 if (bits > ctx->frame_bits)
1059 if (bits > ctx->frame_bits)
1060 return AVERROR(EINVAL);
1063 if (bits < ctx->frame_bits) {
1064 last_lower = FFMIN(lambda, last_lower);
1065 if (last_higher != 0)
1066 lambda = (lambda+last_higher)>>1;
1068 lambda -= down_step;
1069 down_step = FFMIN((int64_t)down_step*5, INT_MAX);
1070 up_step = 1<<LAMBDA_FRAC_BITS;
1071 lambda = FFMAX(1, lambda);
1072 if (lambda == last_lower)
1075 last_higher = FFMAX(lambda, last_higher);
1076 if (last_lower != INT_MAX)
1077 lambda = (lambda+last_lower)>>1;
1078 else if ((int64_t)lambda + up_step > INT_MAX)
1079 return AVERROR(EINVAL);
1082 up_step = FFMIN((int64_t)up_step*5, INT_MAX);
1083 down_step = 1<<LAMBDA_FRAC_BITS;
1086 ctx->lambda = lambda;
1090 static int dnxhd_find_qscale(DNXHDEncContext *ctx)
1095 int last_higher = 0;
1096 int last_lower = INT_MAX;
1100 qscale = ctx->qscale;
1103 ctx->qscale = qscale;
1104 // XXX avoid recalculating bits
1105 ctx->m.avctx->execute2(ctx->m.avctx, dnxhd_calc_bits_thread,
1106 NULL, NULL, ctx->m.mb_height);
1107 for (y = 0; y < ctx->m.mb_height; y++) {
1108 for (x = 0; x < ctx->m.mb_width; x++)
1109 bits += ctx->mb_rc[(qscale*ctx->m.mb_num) + (y*ctx->m.mb_width+x)].bits;
1110 bits = (bits+31)&~31; // padding
1111 if (bits > ctx->frame_bits)
1114 if (bits < ctx->frame_bits) {
1117 if (last_higher == qscale - 1) {
1118 qscale = last_higher;
1121 last_lower = FFMIN(qscale, last_lower);
1122 if (last_higher != 0)
1123 qscale = (qscale + last_higher) >> 1;
1125 qscale -= down_step++;
1130 if (last_lower == qscale + 1)
1132 last_higher = FFMAX(qscale, last_higher);
1133 if (last_lower != INT_MAX)
1134 qscale = (qscale + last_lower) >> 1;
1136 qscale += up_step++;
1138 if (qscale >= ctx->m.avctx->qmax)
1139 return AVERROR(EINVAL);
1142 ctx->qscale = qscale;
1146 #define BUCKET_BITS 8
1147 #define RADIX_PASSES 4
1148 #define NBUCKETS (1 << BUCKET_BITS)
1150 static inline int get_bucket(int value, int shift)
1153 value &= NBUCKETS - 1;
1154 return NBUCKETS - 1 - value;
1157 static void radix_count(const RCCMPEntry *data, int size,
1158 int buckets[RADIX_PASSES][NBUCKETS])
1161 memset(buckets, 0, sizeof(buckets[0][0]) * RADIX_PASSES * NBUCKETS);
1162 for (i = 0; i < size; i++) {
1163 int v = data[i].value;
1164 for (j = 0; j < RADIX_PASSES; j++) {
1165 buckets[j][get_bucket(v, 0)]++;
1170 for (j = 0; j < RADIX_PASSES; j++) {
1172 for (i = NBUCKETS - 1; i >= 0; i--)
1173 buckets[j][i] = offset -= buckets[j][i];
1174 av_assert1(!buckets[j][0]);
1178 static void radix_sort_pass(RCCMPEntry *dst, const RCCMPEntry *data,
1179 int size, int buckets[NBUCKETS], int pass)
1181 int shift = pass * BUCKET_BITS;
1183 for (i = 0; i < size; i++) {
1184 int v = get_bucket(data[i].value, shift);
1185 int pos = buckets[v]++;
1190 static void radix_sort(RCCMPEntry *data, RCCMPEntry *tmp, int size)
1192 int buckets[RADIX_PASSES][NBUCKETS];
1193 radix_count(data, size, buckets);
1194 radix_sort_pass(tmp, data, size, buckets[0], 0);
1195 radix_sort_pass(data, tmp, size, buckets[1], 1);
1196 if (buckets[2][NBUCKETS - 1] || buckets[3][NBUCKETS - 1]) {
1197 radix_sort_pass(tmp, data, size, buckets[2], 2);
1198 radix_sort_pass(data, tmp, size, buckets[3], 3);
1202 static int dnxhd_encode_fast(AVCodecContext *avctx, DNXHDEncContext *ctx)
1206 if ((ret = dnxhd_find_qscale(ctx)) < 0)
1208 for (y = 0; y < ctx->m.mb_height; y++) {
1209 for (x = 0; x < ctx->m.mb_width; x++) {
1210 int mb = y * ctx->m.mb_width + x;
1211 int rc = (ctx->qscale * ctx->m.mb_num ) + mb;
1213 ctx->mb_qscale[mb] = ctx->qscale;
1214 ctx->mb_bits[mb] = ctx->mb_rc[rc].bits;
1215 max_bits += ctx->mb_rc[rc].bits;
1217 delta_bits = ctx->mb_rc[rc].bits -
1218 ctx->mb_rc[rc + ctx->m.mb_num].bits;
1219 ctx->mb_cmp[mb].mb = mb;
1220 ctx->mb_cmp[mb].value =
1221 delta_bits ? ((ctx->mb_rc[rc].ssd -
1222 ctx->mb_rc[rc + ctx->m.mb_num].ssd) * 100) /
1224 : INT_MIN; // avoid increasing qscale
1227 max_bits += 31; // worst padding
1231 avctx->execute2(avctx, dnxhd_mb_var_thread,
1232 NULL, NULL, ctx->m.mb_height);
1233 radix_sort(ctx->mb_cmp, ctx->mb_cmp_tmp, ctx->m.mb_num);
1234 for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) {
1235 int mb = ctx->mb_cmp[x].mb;
1236 int rc = (ctx->qscale * ctx->m.mb_num ) + mb;
1237 max_bits -= ctx->mb_rc[rc].bits -
1238 ctx->mb_rc[rc + ctx->m.mb_num].bits;
1239 ctx->mb_qscale[mb] = ctx->qscale + 1;
1240 ctx->mb_bits[mb] = ctx->mb_rc[rc + ctx->m.mb_num].bits;
1246 static void dnxhd_load_picture(DNXHDEncContext *ctx, const AVFrame *frame)
1250 for (i = 0; i < ctx->m.avctx->thread_count; i++) {
1251 ctx->thread[i]->m.linesize = frame->linesize[0] << ctx->interlaced;
1252 ctx->thread[i]->m.uvlinesize = frame->linesize[1] << ctx->interlaced;
1253 ctx->thread[i]->dct_y_offset = ctx->m.linesize *8;
1254 ctx->thread[i]->dct_uv_offset = ctx->m.uvlinesize*8;
1257 #if FF_API_CODED_FRAME
1258 FF_DISABLE_DEPRECATION_WARNINGS
1259 ctx->m.avctx->coded_frame->interlaced_frame = frame->interlaced_frame;
1260 FF_ENABLE_DEPRECATION_WARNINGS
1262 ctx->cur_field = frame->interlaced_frame && !frame->top_field_first;
1265 static int dnxhd_encode_picture(AVCodecContext *avctx, AVPacket *pkt,
1266 const AVFrame *frame, int *got_packet)
1268 DNXHDEncContext *ctx = avctx->priv_data;
1269 int first_field = 1;
1273 if ((ret = ff_alloc_packet2(avctx, pkt, ctx->frame_size, 0)) < 0)
1277 dnxhd_load_picture(ctx, frame);
1280 for (i = 0; i < 3; i++) {
1281 ctx->src[i] = frame->data[i];
1282 if (ctx->interlaced && ctx->cur_field)
1283 ctx->src[i] += frame->linesize[i];
1286 dnxhd_write_header(avctx, buf);
1288 if (avctx->mb_decision == FF_MB_DECISION_RD)
1289 ret = dnxhd_encode_rdo(avctx, ctx);
1291 ret = dnxhd_encode_fast(avctx, ctx);
1293 av_log(avctx, AV_LOG_ERROR,
1294 "picture could not fit ratecontrol constraints, increase qmax\n");
1298 dnxhd_setup_threads_slices(ctx);
1301 for (i = 0; i < ctx->m.mb_height; i++) {
1302 AV_WB32(ctx->msip + i * 4, offset);
1303 offset += ctx->slice_size[i];
1304 av_assert1(!(ctx->slice_size[i] & 3));
1307 avctx->execute2(avctx, dnxhd_encode_thread, buf, NULL, ctx->m.mb_height);
1309 av_assert1(ctx->data_offset + offset + 4 <= ctx->coding_unit_size);
1310 memset(buf + ctx->data_offset + offset, 0,
1311 ctx->coding_unit_size - 4 - offset - ctx->data_offset);
1313 AV_WB32(buf + ctx->coding_unit_size - 4, 0x600DC0DE); // EOF
1315 if (ctx->interlaced && first_field) {
1317 ctx->cur_field ^= 1;
1318 buf += ctx->coding_unit_size;
1319 goto encode_coding_unit;
1322 #if FF_API_CODED_FRAME
1323 FF_DISABLE_DEPRECATION_WARNINGS
1324 avctx->coded_frame->quality = ctx->qscale * FF_QP2LAMBDA;
1325 FF_ENABLE_DEPRECATION_WARNINGS
1328 ff_side_data_set_encoder_stats(pkt, ctx->qscale * FF_QP2LAMBDA, NULL, 0, AV_PICTURE_TYPE_I);
1330 pkt->flags |= AV_PKT_FLAG_KEY;
1335 static av_cold int dnxhd_encode_end(AVCodecContext *avctx)
1337 DNXHDEncContext *ctx = avctx->priv_data;
1340 av_freep(&ctx->orig_vlc_codes);
1341 av_freep(&ctx->orig_vlc_bits);
1342 av_freep(&ctx->run_codes);
1343 av_freep(&ctx->run_bits);
1345 av_freep(&ctx->mb_bits);
1346 av_freep(&ctx->mb_qscale);
1347 av_freep(&ctx->mb_rc);
1348 av_freep(&ctx->mb_cmp);
1349 av_freep(&ctx->mb_cmp_tmp);
1350 av_freep(&ctx->slice_size);
1351 av_freep(&ctx->slice_offs);
1353 av_freep(&ctx->qmatrix_c);
1354 av_freep(&ctx->qmatrix_l);
1355 av_freep(&ctx->qmatrix_c16);
1356 av_freep(&ctx->qmatrix_l16);
1358 if (avctx->active_thread_type == FF_THREAD_SLICE) {
1359 for (i = 1; i < avctx->thread_count; i++)
1360 av_freep(&ctx->thread[i]);
1366 static const AVCodecDefault dnxhd_defaults[] = {
1367 { "qmax", "1024" }, /* Maximum quantization scale factor allowed for VC-3 */
1371 AVCodec ff_dnxhd_encoder = {
1373 .long_name = NULL_IF_CONFIG_SMALL("VC3/DNxHD"),
1374 .type = AVMEDIA_TYPE_VIDEO,
1375 .id = AV_CODEC_ID_DNXHD,
1376 .priv_data_size = sizeof(DNXHDEncContext),
1377 .init = dnxhd_encode_init,
1378 .encode2 = dnxhd_encode_picture,
1379 .close = dnxhd_encode_end,
1380 .capabilities = AV_CODEC_CAP_SLICE_THREADS | AV_CODEC_CAP_FRAME_THREADS,
1381 .caps_internal = FF_CODEC_CAP_INIT_CLEANUP,
1382 .pix_fmts = (const enum AVPixelFormat[]) {
1384 AV_PIX_FMT_YUV422P10,
1385 AV_PIX_FMT_YUV444P10,
1389 .priv_class = &dnxhd_class,
1390 .defaults = dnxhd_defaults,
1391 .profiles = NULL_IF_CONFIG_SMALL(ff_dnxhd_profiles),