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 10bit 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_INT, { .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 },
55 static const AVClass dnxhd_class = {
56 .class_name = "dnxhd",
57 .item_name = av_default_item_name,
59 .version = LIBAVUTIL_VERSION_INT,
62 static void dnxhd_8bit_get_pixels_8x4_sym(int16_t *av_restrict block,
63 const uint8_t *pixels,
67 for (i = 0; i < 4; i++) {
79 memcpy(block, block - 8, sizeof(*block) * 8);
80 memcpy(block + 8, block - 16, sizeof(*block) * 8);
81 memcpy(block + 16, block - 24, sizeof(*block) * 8);
82 memcpy(block + 24, block - 32, sizeof(*block) * 8);
85 static av_always_inline
86 void dnxhd_10bit_get_pixels_8x4_sym(int16_t *av_restrict block,
87 const uint8_t *pixels,
91 const uint16_t* pixels16 = (const uint16_t*)pixels;
94 for (i = 0; i < 4; i++) {
95 block[0] = pixels16[0]; block[1] = pixels16[1];
96 block[2] = pixels16[2]; block[3] = pixels16[3];
97 block[4] = pixels16[4]; block[5] = pixels16[5];
98 block[6] = pixels16[6]; block[7] = pixels16[7];
99 pixels16 += line_size;
102 memcpy(block, block - 8, sizeof(*block) * 8);
103 memcpy(block + 8, block - 16, sizeof(*block) * 8);
104 memcpy(block + 16, block - 24, sizeof(*block) * 8);
105 memcpy(block + 24, block - 32, sizeof(*block) * 8);
108 static int dnxhd_10bit_dct_quantize(MpegEncContext *ctx, int16_t *block,
109 int n, int qscale, int *overflow)
111 const uint8_t *scantable= ctx->intra_scantable.scantable;
112 const int *qmat = n<4 ? ctx->q_intra_matrix[qscale] : ctx->q_chroma_intra_matrix[qscale];
113 int last_non_zero = 0;
116 ctx->fdsp.fdct(block);
118 // Divide by 4 with rounding, to compensate scaling of DCT coefficients
119 block[0] = (block[0] + 2) >> 2;
121 for (i = 1; i < 64; ++i) {
122 int j = scantable[i];
123 int sign = FF_SIGNBIT(block[j]);
124 int level = (block[j] ^ sign) - sign;
125 level = level * qmat[j] >> DNX10BIT_QMAT_SHIFT;
126 block[j] = (level ^ sign) - sign;
131 return last_non_zero;
134 static av_cold int dnxhd_init_vlc(DNXHDEncContext *ctx)
136 int i, j, level, run;
137 int max_level = 1 << (ctx->cid_table->bit_depth + 2);
139 FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->vlc_codes,
140 max_level, 4 * sizeof(*ctx->vlc_codes), fail);
141 FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->vlc_bits,
142 max_level, 4 * sizeof(*ctx->vlc_bits), fail);
143 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_codes,
145 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_bits,
148 ctx->vlc_codes += max_level * 2;
149 ctx->vlc_bits += max_level * 2;
150 for (level = -max_level; level < max_level; level++) {
151 for (run = 0; run < 2; run++) {
152 int index = (level << 1) | run;
153 int sign, offset = 0, alevel = level;
155 MASK_ABS(sign, alevel);
157 offset = (alevel - 1) >> 6;
158 alevel -= offset << 6;
160 for (j = 0; j < 257; j++) {
161 if (ctx->cid_table->ac_level[j] >> 1 == alevel &&
162 (!offset || (ctx->cid_table->ac_flags[j] & 1) && offset) &&
163 (!run || (ctx->cid_table->ac_flags[j] & 2) && run)) {
164 av_assert1(!ctx->vlc_codes[index]);
166 ctx->vlc_codes[index] =
167 (ctx->cid_table->ac_codes[j] << 1) | (sign & 1);
168 ctx->vlc_bits[index] = ctx->cid_table->ac_bits[j] + 1;
170 ctx->vlc_codes[index] = ctx->cid_table->ac_codes[j];
171 ctx->vlc_bits[index] = ctx->cid_table->ac_bits[j];
176 av_assert0(!alevel || j < 257);
178 ctx->vlc_codes[index] =
179 (ctx->vlc_codes[index] << ctx->cid_table->index_bits) | offset;
180 ctx->vlc_bits[index] += ctx->cid_table->index_bits;
184 for (i = 0; i < 62; i++) {
185 int run = ctx->cid_table->run[i];
186 av_assert0(run < 63);
187 ctx->run_codes[run] = ctx->cid_table->run_codes[i];
188 ctx->run_bits[run] = ctx->cid_table->run_bits[i];
192 return AVERROR(ENOMEM);
195 static av_cold int dnxhd_init_qmat(DNXHDEncContext *ctx, int lbias, int cbias)
197 // init first elem to 1 to avoid div by 0 in convert_matrix
198 uint16_t weight_matrix[64] = { 1, }; // convert_matrix needs uint16_t*
200 const uint8_t *luma_weight_table = ctx->cid_table->luma_weight;
201 const uint8_t *chroma_weight_table = ctx->cid_table->chroma_weight;
203 FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l,
204 (ctx->m.avctx->qmax + 1), 64 * sizeof(int), fail);
205 FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c,
206 (ctx->m.avctx->qmax + 1), 64 * sizeof(int), fail);
207 FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l16,
208 (ctx->m.avctx->qmax + 1), 64 * 2 * sizeof(uint16_t),
210 FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c16,
211 (ctx->m.avctx->qmax + 1), 64 * 2 * sizeof(uint16_t),
214 if (ctx->cid_table->bit_depth == 8) {
215 for (i = 1; i < 64; i++) {
216 int j = ctx->m.idsp.idct_permutation[ff_zigzag_direct[i]];
217 weight_matrix[j] = ctx->cid_table->luma_weight[i];
219 ff_convert_matrix(&ctx->m, ctx->qmatrix_l, ctx->qmatrix_l16,
220 weight_matrix, ctx->intra_quant_bias, 1,
221 ctx->m.avctx->qmax, 1);
222 for (i = 1; i < 64; i++) {
223 int j = ctx->m.idsp.idct_permutation[ff_zigzag_direct[i]];
224 weight_matrix[j] = ctx->cid_table->chroma_weight[i];
226 ff_convert_matrix(&ctx->m, ctx->qmatrix_c, ctx->qmatrix_c16,
227 weight_matrix, ctx->intra_quant_bias, 1,
228 ctx->m.avctx->qmax, 1);
230 for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
231 for (i = 0; i < 64; i++) {
232 ctx->qmatrix_l[qscale][i] <<= 2;
233 ctx->qmatrix_c[qscale][i] <<= 2;
234 ctx->qmatrix_l16[qscale][0][i] <<= 2;
235 ctx->qmatrix_l16[qscale][1][i] <<= 2;
236 ctx->qmatrix_c16[qscale][0][i] <<= 2;
237 ctx->qmatrix_c16[qscale][1][i] <<= 2;
242 for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
243 for (i = 1; i < 64; i++) {
244 int j = ctx->m.idsp.idct_permutation[ff_zigzag_direct[i]];
246 /* The quantization formula from the VC-3 standard is:
247 * quantized = sign(block[i]) * floor(abs(block[i]/s) * p /
248 * (qscale * weight_table[i]))
249 * Where p is 32 for 8-bit samples and 8 for 10-bit ones.
250 * The s factor compensates scaling of DCT coefficients done by
251 * the DCT routines, and therefore is not present in standard.
252 * It's 8 for 8-bit samples and 4 for 10-bit ones.
253 * We want values of ctx->qtmatrix_l and ctx->qtmatrix_r to be:
254 * ((1 << DNX10BIT_QMAT_SHIFT) * (p / s)) /
255 * (qscale * weight_table[i])
256 * For 10-bit samples, p / s == 2 */
257 ctx->qmatrix_l[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) /
258 (qscale * luma_weight_table[i]);
259 ctx->qmatrix_c[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) /
260 (qscale * chroma_weight_table[i]);
265 ctx->m.q_chroma_intra_matrix16 = ctx->qmatrix_c16;
266 ctx->m.q_chroma_intra_matrix = ctx->qmatrix_c;
267 ctx->m.q_intra_matrix16 = ctx->qmatrix_l16;
268 ctx->m.q_intra_matrix = ctx->qmatrix_l;
272 return AVERROR(ENOMEM);
275 static av_cold int dnxhd_init_rc(DNXHDEncContext *ctx)
277 FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->mb_rc, (ctx->m.avctx->qmax + 1), 8160 * sizeof(RCEntry), fail);
278 if (ctx->m.avctx->mb_decision != FF_MB_DECISION_RD)
279 FF_ALLOCZ_ARRAY_OR_GOTO(ctx->m.avctx, ctx->mb_cmp,
280 ctx->m.mb_num, sizeof(RCCMPEntry), fail);
282 ctx->frame_bits = (ctx->cid_table->coding_unit_size -
283 640 - 4 - ctx->min_padding) * 8;
285 ctx->lambda = 2 << LAMBDA_FRAC_BITS; // qscale 2
288 return AVERROR(ENOMEM);
291 static av_cold int dnxhd_encode_init(AVCodecContext *avctx)
293 DNXHDEncContext *ctx = avctx->priv_data;
294 int i, index, bit_depth, ret;
296 switch (avctx->pix_fmt) {
297 case AV_PIX_FMT_YUV422P:
300 case AV_PIX_FMT_YUV422P10:
304 av_log(avctx, AV_LOG_ERROR,
305 "pixel format is incompatible with DNxHD\n");
306 return AVERROR(EINVAL);
309 ctx->cid = ff_dnxhd_find_cid(avctx, bit_depth);
311 av_log(avctx, AV_LOG_ERROR,
312 "video parameters incompatible with DNxHD. Valid DNxHD profiles:\n");
313 ff_dnxhd_print_profiles(avctx, AV_LOG_ERROR);
314 return AVERROR(EINVAL);
316 av_log(avctx, AV_LOG_DEBUG, "cid %d\n", ctx->cid);
318 index = ff_dnxhd_get_cid_table(ctx->cid);
319 av_assert0(index >= 0);
321 ctx->cid_table = &ff_dnxhd_cid_table[index];
323 ctx->m.avctx = avctx;
327 avctx->bits_per_raw_sample = ctx->cid_table->bit_depth;
329 ff_blockdsp_init(&ctx->bdsp, avctx);
330 ff_fdctdsp_init(&ctx->m.fdsp, avctx);
331 ff_mpv_idct_init(&ctx->m);
332 ff_mpegvideoencdsp_init(&ctx->m.mpvencdsp, avctx);
333 ff_pixblockdsp_init(&ctx->m.pdsp, avctx);
334 ff_dct_encode_init(&ctx->m);
336 if (!ctx->m.dct_quantize)
337 ctx->m.dct_quantize = ff_dct_quantize_c;
339 if (ctx->cid_table->bit_depth == 10) {
340 ctx->m.dct_quantize = dnxhd_10bit_dct_quantize;
341 ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym;
342 ctx->block_width_l2 = 4;
344 ctx->get_pixels_8x4_sym = dnxhd_8bit_get_pixels_8x4_sym;
345 ctx->block_width_l2 = 3;
349 ff_dnxhdenc_init_x86(ctx);
351 ctx->m.mb_height = (avctx->height + 15) / 16;
352 ctx->m.mb_width = (avctx->width + 15) / 16;
354 if (avctx->flags & AV_CODEC_FLAG_INTERLACED_DCT) {
356 ctx->m.mb_height /= 2;
359 ctx->m.mb_num = ctx->m.mb_height * ctx->m.mb_width;
361 #if FF_API_QUANT_BIAS
362 FF_DISABLE_DEPRECATION_WARNINGS
363 if (avctx->intra_quant_bias != FF_DEFAULT_QUANT_BIAS)
364 ctx->intra_quant_bias = avctx->intra_quant_bias;
365 FF_ENABLE_DEPRECATION_WARNINGS
367 // XXX tune lbias/cbias
368 if ((ret = dnxhd_init_qmat(ctx, ctx->intra_quant_bias, 0)) < 0)
371 /* Avid Nitris hardware decoder requires a minimum amount of padding
372 * in the coding unit payload */
373 if (ctx->nitris_compat)
374 ctx->min_padding = 1600;
376 if ((ret = dnxhd_init_vlc(ctx)) < 0)
378 if ((ret = dnxhd_init_rc(ctx)) < 0)
381 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_size,
382 ctx->m.mb_height * sizeof(uint32_t), fail);
383 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_offs,
384 ctx->m.mb_height * sizeof(uint32_t), fail);
385 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_bits,
386 ctx->m.mb_num * sizeof(uint16_t), fail);
387 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_qscale,
388 ctx->m.mb_num * sizeof(uint8_t), fail);
390 #if FF_API_CODED_FRAME
391 FF_DISABLE_DEPRECATION_WARNINGS
392 avctx->coded_frame->key_frame = 1;
393 avctx->coded_frame->pict_type = AV_PICTURE_TYPE_I;
394 FF_ENABLE_DEPRECATION_WARNINGS
397 if (avctx->thread_count > MAX_THREADS) {
398 av_log(avctx, AV_LOG_ERROR, "too many threads\n");
399 return AVERROR(EINVAL);
402 if (avctx->qmax <= 1) {
403 av_log(avctx, AV_LOG_ERROR, "qmax must be at least 2\n");
404 return AVERROR(EINVAL);
407 ctx->thread[0] = ctx;
408 for (i = 1; i < avctx->thread_count; i++) {
409 ctx->thread[i] = av_malloc(sizeof(DNXHDEncContext));
410 memcpy(ctx->thread[i], ctx, sizeof(DNXHDEncContext));
414 fail: // for FF_ALLOCZ_OR_GOTO
415 return AVERROR(ENOMEM);
418 static int dnxhd_write_header(AVCodecContext *avctx, uint8_t *buf)
420 DNXHDEncContext *ctx = avctx->priv_data;
421 static const uint8_t header_prefix[5] = { 0x00, 0x00, 0x02, 0x80, 0x01 };
425 memcpy(buf, header_prefix, 5);
426 buf[5] = ctx->interlaced ? ctx->cur_field + 2 : 0x01;
427 buf[6] = 0x80; // crc flag off
428 buf[7] = 0xa0; // reserved
429 AV_WB16(buf + 0x18, avctx->height >> ctx->interlaced); // ALPF
430 AV_WB16(buf + 0x1a, avctx->width); // SPL
431 AV_WB16(buf + 0x1d, avctx->height >> ctx->interlaced); // NAL
433 buf[0x21] = ctx->cid_table->bit_depth == 10 ? 0x58 : 0x38;
434 buf[0x22] = 0x88 + (ctx->interlaced << 2);
435 AV_WB32(buf + 0x28, ctx->cid); // CID
436 buf[0x2c] = ctx->interlaced ? 0 : 0x80;
438 buf[0x5f] = 0x01; // UDL
440 buf[0x167] = 0x02; // reserved
441 AV_WB16(buf + 0x16a, ctx->m.mb_height * 4 + 4); // MSIPS
442 buf[0x16d] = ctx->m.mb_height; // Ns
443 buf[0x16f] = 0x10; // reserved
445 ctx->msip = buf + 0x170;
449 static av_always_inline void dnxhd_encode_dc(DNXHDEncContext *ctx, int diff)
453 nbits = av_log2_16bit(-2 * diff);
456 nbits = av_log2_16bit(2 * diff);
458 put_bits(&ctx->m.pb, ctx->cid_table->dc_bits[nbits] + nbits,
459 (ctx->cid_table->dc_codes[nbits] << nbits) +
460 av_mod_uintp2(diff, nbits));
463 static av_always_inline
464 void dnxhd_encode_block(DNXHDEncContext *ctx, int16_t *block,
465 int last_index, int n)
467 int last_non_zero = 0;
470 dnxhd_encode_dc(ctx, block[0] - ctx->m.last_dc[n]);
471 ctx->m.last_dc[n] = block[0];
473 for (i = 1; i <= last_index; i++) {
474 j = ctx->m.intra_scantable.permutated[i];
477 int run_level = i - last_non_zero - 1;
478 int rlevel = (slevel << 1) | !!run_level;
479 put_bits(&ctx->m.pb, ctx->vlc_bits[rlevel], ctx->vlc_codes[rlevel]);
481 put_bits(&ctx->m.pb, ctx->run_bits[run_level],
482 ctx->run_codes[run_level]);
486 put_bits(&ctx->m.pb, ctx->vlc_bits[0], ctx->vlc_codes[0]); // EOB
489 static av_always_inline
490 void dnxhd_unquantize_c(DNXHDEncContext *ctx, int16_t *block, int n,
491 int qscale, int last_index)
493 const uint8_t *weight_matrix;
497 weight_matrix = (n & 2) ? ctx->cid_table->chroma_weight
498 : ctx->cid_table->luma_weight;
500 for (i = 1; i <= last_index; i++) {
501 int j = ctx->m.intra_scantable.permutated[i];
505 level = (1 - 2 * level) * qscale * weight_matrix[i];
506 if (ctx->cid_table->bit_depth == 10) {
507 if (weight_matrix[i] != 8)
511 if (weight_matrix[i] != 32)
517 level = (2 * level + 1) * qscale * weight_matrix[i];
518 if (ctx->cid_table->bit_depth == 10) {
519 if (weight_matrix[i] != 8)
523 if (weight_matrix[i] != 32)
533 static av_always_inline int dnxhd_ssd_block(int16_t *qblock, int16_t *block)
537 for (i = 0; i < 64; i++)
538 score += (block[i] - qblock[i]) * (block[i] - qblock[i]);
542 static av_always_inline
543 int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, int16_t *block, int last_index)
545 int last_non_zero = 0;
548 for (i = 1; i <= last_index; i++) {
549 j = ctx->m.intra_scantable.permutated[i];
552 int run_level = i - last_non_zero - 1;
553 bits += ctx->vlc_bits[(level << 1) |
554 !!run_level] + ctx->run_bits[run_level];
561 static av_always_inline
562 void dnxhd_get_blocks(DNXHDEncContext *ctx, int mb_x, int mb_y)
564 const int bs = ctx->block_width_l2;
565 const int bw = 1 << bs;
566 const uint8_t *ptr_y = ctx->thread[0]->src[0] +
567 ((mb_y << 4) * ctx->m.linesize) + (mb_x << bs + 1);
568 const uint8_t *ptr_u = ctx->thread[0]->src[1] +
569 ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs);
570 const uint8_t *ptr_v = ctx->thread[0]->src[2] +
571 ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs);
572 PixblockDSPContext *pdsp = &ctx->m.pdsp;
574 pdsp->get_pixels(ctx->blocks[0], ptr_y, ctx->m.linesize);
575 pdsp->get_pixels(ctx->blocks[1], ptr_y + bw, ctx->m.linesize);
576 pdsp->get_pixels(ctx->blocks[2], ptr_u, ctx->m.uvlinesize);
577 pdsp->get_pixels(ctx->blocks[3], ptr_v, ctx->m.uvlinesize);
579 if (mb_y + 1 == ctx->m.mb_height && ctx->m.avctx->height == 1080) {
580 if (ctx->interlaced) {
581 ctx->get_pixels_8x4_sym(ctx->blocks[4],
582 ptr_y + ctx->dct_y_offset,
584 ctx->get_pixels_8x4_sym(ctx->blocks[5],
585 ptr_y + ctx->dct_y_offset + bw,
587 ctx->get_pixels_8x4_sym(ctx->blocks[6],
588 ptr_u + ctx->dct_uv_offset,
590 ctx->get_pixels_8x4_sym(ctx->blocks[7],
591 ptr_v + ctx->dct_uv_offset,
594 ctx->bdsp.clear_block(ctx->blocks[4]);
595 ctx->bdsp.clear_block(ctx->blocks[5]);
596 ctx->bdsp.clear_block(ctx->blocks[6]);
597 ctx->bdsp.clear_block(ctx->blocks[7]);
600 pdsp->get_pixels(ctx->blocks[4],
601 ptr_y + ctx->dct_y_offset, ctx->m.linesize);
602 pdsp->get_pixels(ctx->blocks[5],
603 ptr_y + ctx->dct_y_offset + bw, ctx->m.linesize);
604 pdsp->get_pixels(ctx->blocks[6],
605 ptr_u + ctx->dct_uv_offset, ctx->m.uvlinesize);
606 pdsp->get_pixels(ctx->blocks[7],
607 ptr_v + ctx->dct_uv_offset, ctx->m.uvlinesize);
611 static av_always_inline
612 int dnxhd_switch_matrix(DNXHDEncContext *ctx, int i)
614 const static uint8_t component[8]={0,0,1,2,0,0,1,2};
618 static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg,
619 int jobnr, int threadnr)
621 DNXHDEncContext *ctx = avctx->priv_data;
622 int mb_y = jobnr, mb_x;
623 int qscale = ctx->qscale;
624 LOCAL_ALIGNED_16(int16_t, block, [64]);
625 ctx = ctx->thread[threadnr];
629 ctx->m.last_dc[2] = 1 << (ctx->cid_table->bit_depth + 2);
631 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
632 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
638 dnxhd_get_blocks(ctx, mb_x, mb_y);
640 for (i = 0; i < 8; i++) {
641 int16_t *src_block = ctx->blocks[i];
642 int overflow, nbits, diff, last_index;
643 int n = dnxhd_switch_matrix(ctx, i);
645 memcpy(block, src_block, 64 * sizeof(*block));
646 last_index = ctx->m.dct_quantize(&ctx->m, block, 4 & (2*i),
648 ac_bits += dnxhd_calc_ac_bits(ctx, block, last_index);
650 diff = block[0] - ctx->m.last_dc[n];
652 nbits = av_log2_16bit(-2 * diff);
654 nbits = av_log2_16bit(2 * diff);
656 av_assert1(nbits < ctx->cid_table->bit_depth + 4);
657 dc_bits += ctx->cid_table->dc_bits[nbits] + nbits;
659 ctx->m.last_dc[n] = block[0];
661 if (avctx->mb_decision == FF_MB_DECISION_RD || !RC_VARIANCE) {
662 dnxhd_unquantize_c(ctx, block, i, qscale, last_index);
663 ctx->m.idsp.idct(block);
664 ssd += dnxhd_ssd_block(block, src_block);
667 ctx->mb_rc[qscale][mb].ssd = ssd;
668 ctx->mb_rc[qscale][mb].bits = ac_bits + dc_bits + 12 +
669 8 * ctx->vlc_bits[0];
674 static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg,
675 int jobnr, int threadnr)
677 DNXHDEncContext *ctx = avctx->priv_data;
678 int mb_y = jobnr, mb_x;
679 ctx = ctx->thread[threadnr];
680 init_put_bits(&ctx->m.pb, (uint8_t *)arg + 640 + ctx->slice_offs[jobnr],
681 ctx->slice_size[jobnr]);
685 ctx->m.last_dc[2] = 1 << (ctx->cid_table->bit_depth + 2);
686 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
687 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
688 int qscale = ctx->mb_qscale[mb];
691 put_bits(&ctx->m.pb, 12, qscale << 1);
693 dnxhd_get_blocks(ctx, mb_x, mb_y);
695 for (i = 0; i < 8; i++) {
696 int16_t *block = ctx->blocks[i];
697 int overflow, n = dnxhd_switch_matrix(ctx, i);
698 int last_index = ctx->m.dct_quantize(&ctx->m, block, 4 & (2*i),
701 dnxhd_encode_block(ctx, block, last_index, n);
702 // STOP_TIMER("encode_block");
705 if (put_bits_count(&ctx->m.pb) & 31)
706 put_bits(&ctx->m.pb, 32 - (put_bits_count(&ctx->m.pb) & 31), 0);
707 flush_put_bits(&ctx->m.pb);
711 static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx)
715 for (mb_y = 0; mb_y < ctx->m.mb_height; mb_y++) {
717 ctx->slice_offs[mb_y] = offset;
718 ctx->slice_size[mb_y] = 0;
719 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
720 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
721 ctx->slice_size[mb_y] += ctx->mb_bits[mb];
723 ctx->slice_size[mb_y] = (ctx->slice_size[mb_y] + 31) & ~31;
724 ctx->slice_size[mb_y] >>= 3;
725 thread_size = ctx->slice_size[mb_y];
726 offset += thread_size;
730 static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg,
731 int jobnr, int threadnr)
733 DNXHDEncContext *ctx = avctx->priv_data;
734 int mb_y = jobnr, mb_x, x, y;
735 int partial_last_row = (mb_y == ctx->m.mb_height - 1) &&
736 ((avctx->height >> ctx->interlaced) & 0xF);
738 ctx = ctx->thread[threadnr];
739 if (ctx->cid_table->bit_depth == 8) {
740 uint8_t *pix = ctx->thread[0]->src[0] + ((mb_y << 4) * ctx->m.linesize);
741 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x, pix += 16) {
742 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
746 if (!partial_last_row && mb_x * 16 <= avctx->width - 16) {
747 sum = ctx->m.mpvencdsp.pix_sum(pix, ctx->m.linesize);
748 varc = ctx->m.mpvencdsp.pix_norm1(pix, ctx->m.linesize);
750 int bw = FFMIN(avctx->width - 16 * mb_x, 16);
751 int bh = FFMIN((avctx->height >> ctx->interlaced) - 16 * mb_y, 16);
753 for (y = 0; y < bh; y++) {
754 for (x = 0; x < bw; x++) {
755 uint8_t val = pix[x + y * ctx->m.linesize];
761 varc = (varc - (((unsigned) sum * sum) >> 8) + 128) >> 8;
763 ctx->mb_cmp[mb].value = varc;
764 ctx->mb_cmp[mb].mb = mb;
767 int const linesize = ctx->m.linesize >> 1;
768 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x) {
769 uint16_t *pix = (uint16_t *)ctx->thread[0]->src[0] +
770 ((mb_y << 4) * linesize) + (mb_x << 4);
771 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
776 // Macroblocks are 16x16 pixels, unlike DCT blocks which are 8x8.
777 for (i = 0; i < 16; ++i) {
778 for (j = 0; j < 16; ++j) {
779 // Turn 16-bit pixels into 10-bit ones.
780 int const sample = (unsigned) pix[j] >> 6;
782 sqsum += sample * sample;
783 // 2^10 * 2^10 * 16 * 16 = 2^28, which is less than INT_MAX
787 mean = sum >> 8; // 16*16 == 2^8
789 ctx->mb_cmp[mb].value = sqmean - mean * mean;
790 ctx->mb_cmp[mb].mb = mb;
796 static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx)
798 int lambda, up_step, down_step;
799 int last_lower = INT_MAX, last_higher = 0;
802 for (q = 1; q < avctx->qmax; q++) {
804 avctx->execute2(avctx, dnxhd_calc_bits_thread,
805 NULL, NULL, ctx->m.mb_height);
807 up_step = down_step = 2 << LAMBDA_FRAC_BITS;
808 lambda = ctx->lambda;
813 if (lambda == last_higher) {
815 end = 1; // need to set final qscales/bits
817 for (y = 0; y < ctx->m.mb_height; y++) {
818 for (x = 0; x < ctx->m.mb_width; x++) {
819 unsigned min = UINT_MAX;
821 int mb = y * ctx->m.mb_width + x;
822 for (q = 1; q < avctx->qmax; q++) {
823 unsigned score = ctx->mb_rc[q][mb].bits * lambda +
824 ((unsigned) ctx->mb_rc[q][mb].ssd << LAMBDA_FRAC_BITS);
830 bits += ctx->mb_rc[qscale][mb].bits;
831 ctx->mb_qscale[mb] = qscale;
832 ctx->mb_bits[mb] = ctx->mb_rc[qscale][mb].bits;
834 bits = (bits + 31) & ~31; // padding
835 if (bits > ctx->frame_bits)
838 // ff_dlog(ctx->m.avctx,
839 // "lambda %d, up %u, down %u, bits %d, frame %d\n",
840 // lambda, last_higher, last_lower, bits, ctx->frame_bits);
842 if (bits > ctx->frame_bits)
843 return AVERROR(EINVAL);
846 if (bits < ctx->frame_bits) {
847 last_lower = FFMIN(lambda, last_lower);
848 if (last_higher != 0)
849 lambda = (lambda+last_higher)>>1;
852 down_step = FFMIN((int64_t)down_step*5, INT_MAX);
853 up_step = 1<<LAMBDA_FRAC_BITS;
854 lambda = FFMAX(1, lambda);
855 if (lambda == last_lower)
858 last_higher = FFMAX(lambda, last_higher);
859 if (last_lower != INT_MAX)
860 lambda = (lambda+last_lower)>>1;
861 else if ((int64_t)lambda + up_step > INT_MAX)
862 return AVERROR(EINVAL);
865 up_step = FFMIN((int64_t)up_step*5, INT_MAX);
866 down_step = 1<<LAMBDA_FRAC_BITS;
869 //ff_dlog(ctx->m.avctx, "out lambda %d\n", lambda);
870 ctx->lambda = lambda;
874 static int dnxhd_find_qscale(DNXHDEncContext *ctx)
880 int last_lower = INT_MAX;
884 qscale = ctx->qscale;
887 ctx->qscale = qscale;
888 // XXX avoid recalculating bits
889 ctx->m.avctx->execute2(ctx->m.avctx, dnxhd_calc_bits_thread,
890 NULL, NULL, ctx->m.mb_height);
891 for (y = 0; y < ctx->m.mb_height; y++) {
892 for (x = 0; x < ctx->m.mb_width; x++)
893 bits += ctx->mb_rc[qscale][y*ctx->m.mb_width+x].bits;
894 bits = (bits+31)&~31; // padding
895 if (bits > ctx->frame_bits)
898 // ff_dlog(ctx->m.avctx,
899 // "%d, qscale %d, bits %d, frame %d, higher %d, lower %d\n",
900 // ctx->m.avctx->frame_number, qscale, bits, ctx->frame_bits,
901 // last_higher, last_lower);
902 if (bits < ctx->frame_bits) {
905 if (last_higher == qscale - 1) {
906 qscale = last_higher;
909 last_lower = FFMIN(qscale, last_lower);
910 if (last_higher != 0)
911 qscale = (qscale + last_higher) >> 1;
913 qscale -= down_step++;
918 if (last_lower == qscale + 1)
920 last_higher = FFMAX(qscale, last_higher);
921 if (last_lower != INT_MAX)
922 qscale = (qscale + last_lower) >> 1;
926 if (qscale >= ctx->m.avctx->qmax)
927 return AVERROR(EINVAL);
930 //ff_dlog(ctx->m.avctx, "out qscale %d\n", qscale);
931 ctx->qscale = qscale;
935 #define BUCKET_BITS 8
936 #define RADIX_PASSES 4
937 #define NBUCKETS (1 << BUCKET_BITS)
939 static inline int get_bucket(int value, int shift)
942 value &= NBUCKETS - 1;
943 return NBUCKETS - 1 - value;
946 static void radix_count(const RCCMPEntry *data, int size,
947 int buckets[RADIX_PASSES][NBUCKETS])
950 memset(buckets, 0, sizeof(buckets[0][0]) * RADIX_PASSES * NBUCKETS);
951 for (i = 0; i < size; i++) {
952 int v = data[i].value;
953 for (j = 0; j < RADIX_PASSES; j++) {
954 buckets[j][get_bucket(v, 0)]++;
959 for (j = 0; j < RADIX_PASSES; j++) {
961 for (i = NBUCKETS - 1; i >= 0; i--)
962 buckets[j][i] = offset -= buckets[j][i];
963 av_assert1(!buckets[j][0]);
967 static void radix_sort_pass(RCCMPEntry *dst, const RCCMPEntry *data,
968 int size, int buckets[NBUCKETS], int pass)
970 int shift = pass * BUCKET_BITS;
972 for (i = 0; i < size; i++) {
973 int v = get_bucket(data[i].value, shift);
974 int pos = buckets[v]++;
979 static void radix_sort(RCCMPEntry *data, int size)
981 int buckets[RADIX_PASSES][NBUCKETS];
982 RCCMPEntry *tmp = av_malloc_array(size, sizeof(*tmp));
983 radix_count(data, size, buckets);
984 radix_sort_pass(tmp, data, size, buckets[0], 0);
985 radix_sort_pass(data, tmp, size, buckets[1], 1);
986 if (buckets[2][NBUCKETS - 1] || buckets[3][NBUCKETS - 1]) {
987 radix_sort_pass(tmp, data, size, buckets[2], 2);
988 radix_sort_pass(data, tmp, size, buckets[3], 3);
993 static int dnxhd_encode_fast(AVCodecContext *avctx, DNXHDEncContext *ctx)
997 if ((ret = dnxhd_find_qscale(ctx)) < 0)
999 for (y = 0; y < ctx->m.mb_height; y++) {
1000 for (x = 0; x < ctx->m.mb_width; x++) {
1001 int mb = y * ctx->m.mb_width + x;
1003 ctx->mb_qscale[mb] = ctx->qscale;
1004 ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale][mb].bits;
1005 max_bits += ctx->mb_rc[ctx->qscale][mb].bits;
1007 delta_bits = ctx->mb_rc[ctx->qscale][mb].bits -
1008 ctx->mb_rc[ctx->qscale + 1][mb].bits;
1009 ctx->mb_cmp[mb].mb = mb;
1010 ctx->mb_cmp[mb].value =
1011 delta_bits ? ((ctx->mb_rc[ctx->qscale][mb].ssd -
1012 ctx->mb_rc[ctx->qscale + 1][mb].ssd) * 100) /
1014 : INT_MIN; // avoid increasing qscale
1017 max_bits += 31; // worst padding
1021 avctx->execute2(avctx, dnxhd_mb_var_thread,
1022 NULL, NULL, ctx->m.mb_height);
1023 radix_sort(ctx->mb_cmp, ctx->m.mb_num);
1024 for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) {
1025 int mb = ctx->mb_cmp[x].mb;
1026 max_bits -= ctx->mb_rc[ctx->qscale][mb].bits -
1027 ctx->mb_rc[ctx->qscale + 1][mb].bits;
1028 ctx->mb_qscale[mb] = ctx->qscale + 1;
1029 ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale + 1][mb].bits;
1035 static void dnxhd_load_picture(DNXHDEncContext *ctx, const AVFrame *frame)
1039 for (i = 0; i < ctx->m.avctx->thread_count; i++) {
1040 ctx->thread[i]->m.linesize = frame->linesize[0] << ctx->interlaced;
1041 ctx->thread[i]->m.uvlinesize = frame->linesize[1] << ctx->interlaced;
1042 ctx->thread[i]->dct_y_offset = ctx->m.linesize *8;
1043 ctx->thread[i]->dct_uv_offset = ctx->m.uvlinesize*8;
1046 #if FF_API_CODED_FRAME
1047 FF_DISABLE_DEPRECATION_WARNINGS
1048 ctx->m.avctx->coded_frame->interlaced_frame = frame->interlaced_frame;
1049 FF_ENABLE_DEPRECATION_WARNINGS
1051 ctx->cur_field = frame->interlaced_frame && !frame->top_field_first;
1054 static int dnxhd_encode_picture(AVCodecContext *avctx, AVPacket *pkt,
1055 const AVFrame *frame, int *got_packet)
1057 DNXHDEncContext *ctx = avctx->priv_data;
1058 int first_field = 1;
1062 if ((ret = ff_alloc_packet2(avctx, pkt, ctx->cid_table->frame_size, 0)) < 0)
1066 dnxhd_load_picture(ctx, frame);
1069 for (i = 0; i < 3; i++) {
1070 ctx->src[i] = frame->data[i];
1071 if (ctx->interlaced && ctx->cur_field)
1072 ctx->src[i] += frame->linesize[i];
1075 dnxhd_write_header(avctx, buf);
1077 if (avctx->mb_decision == FF_MB_DECISION_RD)
1078 ret = dnxhd_encode_rdo(avctx, ctx);
1080 ret = dnxhd_encode_fast(avctx, ctx);
1082 av_log(avctx, AV_LOG_ERROR,
1083 "picture could not fit ratecontrol constraints, increase qmax\n");
1087 dnxhd_setup_threads_slices(ctx);
1090 for (i = 0; i < ctx->m.mb_height; i++) {
1091 AV_WB32(ctx->msip + i * 4, offset);
1092 offset += ctx->slice_size[i];
1093 av_assert1(!(ctx->slice_size[i] & 3));
1096 avctx->execute2(avctx, dnxhd_encode_thread, buf, NULL, ctx->m.mb_height);
1098 av_assert1(640 + offset + 4 <= ctx->cid_table->coding_unit_size);
1099 memset(buf + 640 + offset, 0,
1100 ctx->cid_table->coding_unit_size - 4 - offset - 640);
1102 AV_WB32(buf + ctx->cid_table->coding_unit_size - 4, 0x600DC0DE); // EOF
1104 if (ctx->interlaced && first_field) {
1106 ctx->cur_field ^= 1;
1107 buf += ctx->cid_table->coding_unit_size;
1108 goto encode_coding_unit;
1111 #if FF_API_CODED_FRAME
1112 FF_DISABLE_DEPRECATION_WARNINGS
1113 avctx->coded_frame->quality = ctx->qscale * FF_QP2LAMBDA;
1114 FF_ENABLE_DEPRECATION_WARNINGS
1117 ff_side_data_set_encoder_stats(pkt, ctx->qscale * FF_QP2LAMBDA, NULL, 0, AV_PICTURE_TYPE_I);
1119 pkt->flags |= AV_PKT_FLAG_KEY;
1124 static av_cold int dnxhd_encode_end(AVCodecContext *avctx)
1126 DNXHDEncContext *ctx = avctx->priv_data;
1127 int max_level = 1 << (ctx->cid_table->bit_depth + 2);
1130 av_free(ctx->vlc_codes - max_level * 2);
1131 av_free(ctx->vlc_bits - max_level * 2);
1132 av_freep(&ctx->run_codes);
1133 av_freep(&ctx->run_bits);
1135 av_freep(&ctx->mb_bits);
1136 av_freep(&ctx->mb_qscale);
1137 av_freep(&ctx->mb_rc);
1138 av_freep(&ctx->mb_cmp);
1139 av_freep(&ctx->slice_size);
1140 av_freep(&ctx->slice_offs);
1142 av_freep(&ctx->qmatrix_c);
1143 av_freep(&ctx->qmatrix_l);
1144 av_freep(&ctx->qmatrix_c16);
1145 av_freep(&ctx->qmatrix_l16);
1147 for (i = 1; i < avctx->thread_count; i++)
1148 av_freep(&ctx->thread[i]);
1153 static const AVCodecDefault dnxhd_defaults[] = {
1154 { "qmax", "1024" }, /* Maximum quantization scale factor allowed for VC-3 */
1158 AVCodec ff_dnxhd_encoder = {
1160 .long_name = NULL_IF_CONFIG_SMALL("VC3/DNxHD"),
1161 .type = AVMEDIA_TYPE_VIDEO,
1162 .id = AV_CODEC_ID_DNXHD,
1163 .priv_data_size = sizeof(DNXHDEncContext),
1164 .init = dnxhd_encode_init,
1165 .encode2 = dnxhd_encode_picture,
1166 .close = dnxhd_encode_end,
1167 .capabilities = AV_CODEC_CAP_SLICE_THREADS,
1168 .pix_fmts = (const enum AVPixelFormat[]) {
1170 AV_PIX_FMT_YUV422P10,
1173 .priv_class = &dnxhd_class,
1174 .defaults = dnxhd_defaults,