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"
39 // The largest value that will not lead to overflow for 10bit samples.
40 #define DNX10BIT_QMAT_SHIFT 18
41 #define RC_VARIANCE 1 // use variance or ssd for fast rc
42 #define LAMBDA_FRAC_BITS 10
44 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
45 static const AVOption options[] = {
46 { "nitris_compat", "encode with Avid Nitris compatibility",
47 offsetof(DNXHDEncContext, nitris_compat), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 1, VE },
51 static const AVClass dnxhd_class = {
52 .class_name = "dnxhd",
53 .item_name = av_default_item_name,
55 .version = LIBAVUTIL_VERSION_INT,
58 static void dnxhd_8bit_get_pixels_8x4_sym(int16_t *av_restrict block,
59 const uint8_t *pixels,
63 for (i = 0; i < 4; i++) {
75 memcpy(block, block - 8, sizeof(*block) * 8);
76 memcpy(block + 8, block - 16, sizeof(*block) * 8);
77 memcpy(block + 16, block - 24, sizeof(*block) * 8);
78 memcpy(block + 24, block - 32, sizeof(*block) * 8);
81 static av_always_inline
82 void dnxhd_10bit_get_pixels_8x4_sym(int16_t *av_restrict block,
83 const uint8_t *pixels,
87 const uint16_t* pixels16 = (const uint16_t*)pixels;
90 for (i = 0; i < 4; i++) {
91 block[0] = pixels16[0]; block[1] = pixels16[1];
92 block[2] = pixels16[2]; block[3] = pixels16[3];
93 block[4] = pixels16[4]; block[5] = pixels16[5];
94 block[6] = pixels16[6]; block[7] = pixels16[7];
95 pixels16 += line_size;
98 memcpy(block, block - 8, sizeof(*block) * 8);
99 memcpy(block + 8, block - 16, sizeof(*block) * 8);
100 memcpy(block + 16, block - 24, sizeof(*block) * 8);
101 memcpy(block + 24, block - 32, sizeof(*block) * 8);
104 static int dnxhd_10bit_dct_quantize(MpegEncContext *ctx, int16_t *block,
105 int n, int qscale, int *overflow)
107 const uint8_t *scantable= ctx->intra_scantable.scantable;
108 const int *qmat = n<4 ? ctx->q_intra_matrix[qscale] : ctx->q_chroma_intra_matrix[qscale];
109 int last_non_zero = 0;
112 ctx->dsp.fdct(block);
114 // Divide by 4 with rounding, to compensate scaling of DCT coefficients
115 block[0] = (block[0] + 2) >> 2;
117 for (i = 1; i < 64; ++i) {
118 int j = scantable[i];
119 int sign = block[j] >> 31;
120 int level = (block[j] ^ sign) - sign;
121 level = level * qmat[j] >> DNX10BIT_QMAT_SHIFT;
122 block[j] = (level ^ sign) - sign;
127 return last_non_zero;
130 static av_cold int dnxhd_init_vlc(DNXHDEncContext *ctx)
132 int i, j, level, run;
133 int max_level = 1 << (ctx->cid_table->bit_depth + 2);
135 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->vlc_codes,
136 max_level * 4 * sizeof(*ctx->vlc_codes), fail);
137 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->vlc_bits,
138 max_level * 4 * sizeof(*ctx->vlc_bits), fail);
139 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_codes,
141 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_bits,
144 ctx->vlc_codes += max_level * 2;
145 ctx->vlc_bits += max_level * 2;
146 for (level = -max_level; level < max_level; level++) {
147 for (run = 0; run < 2; run++) {
148 int index = (level << 1) | run;
149 int sign, offset = 0, alevel = level;
151 MASK_ABS(sign, alevel);
153 offset = (alevel - 1) >> 6;
154 alevel -= offset << 6;
156 for (j = 0; j < 257; j++) {
157 if (ctx->cid_table->ac_level[j] >> 1 == alevel &&
158 (!offset || (ctx->cid_table->ac_flags[j] & 1) && offset) &&
159 (!run || (ctx->cid_table->ac_flags[j] & 2) && run)) {
160 av_assert1(!ctx->vlc_codes[index]);
162 ctx->vlc_codes[index] =
163 (ctx->cid_table->ac_codes[j] << 1) | (sign & 1);
164 ctx->vlc_bits[index] = ctx->cid_table->ac_bits[j] + 1;
166 ctx->vlc_codes[index] = ctx->cid_table->ac_codes[j];
167 ctx->vlc_bits[index] = ctx->cid_table->ac_bits[j];
172 av_assert0(!alevel || j < 257);
174 ctx->vlc_codes[index] =
175 (ctx->vlc_codes[index] << ctx->cid_table->index_bits) | offset;
176 ctx->vlc_bits[index] += ctx->cid_table->index_bits;
180 for (i = 0; i < 62; i++) {
181 int run = ctx->cid_table->run[i];
182 av_assert0(run < 63);
183 ctx->run_codes[run] = ctx->cid_table->run_codes[i];
184 ctx->run_bits[run] = ctx->cid_table->run_bits[i];
188 return AVERROR(ENOMEM);
191 static av_cold int dnxhd_init_qmat(DNXHDEncContext *ctx, int lbias, int cbias)
193 // init first elem to 1 to avoid div by 0 in convert_matrix
194 uint16_t weight_matrix[64] = { 1, }; // convert_matrix needs uint16_t*
196 const uint8_t *luma_weight_table = ctx->cid_table->luma_weight;
197 const uint8_t *chroma_weight_table = ctx->cid_table->chroma_weight;
199 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l,
200 (ctx->m.avctx->qmax + 1) * 64 * sizeof(int), fail);
201 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c,
202 (ctx->m.avctx->qmax + 1) * 64 * sizeof(int), fail);
203 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l16,
204 (ctx->m.avctx->qmax + 1) * 64 * 2 * sizeof(uint16_t),
206 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c16,
207 (ctx->m.avctx->qmax + 1) * 64 * 2 * sizeof(uint16_t),
210 if (ctx->cid_table->bit_depth == 8) {
211 for (i = 1; i < 64; i++) {
212 int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]];
213 weight_matrix[j] = ctx->cid_table->luma_weight[i];
215 ff_convert_matrix(&ctx->m, ctx->qmatrix_l, ctx->qmatrix_l16,
216 weight_matrix, ctx->m.intra_quant_bias, 1,
217 ctx->m.avctx->qmax, 1);
218 for (i = 1; i < 64; i++) {
219 int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]];
220 weight_matrix[j] = ctx->cid_table->chroma_weight[i];
222 ff_convert_matrix(&ctx->m, ctx->qmatrix_c, ctx->qmatrix_c16,
223 weight_matrix, ctx->m.intra_quant_bias, 1,
224 ctx->m.avctx->qmax, 1);
226 for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
227 for (i = 0; i < 64; i++) {
228 ctx->qmatrix_l[qscale][i] <<= 2;
229 ctx->qmatrix_c[qscale][i] <<= 2;
230 ctx->qmatrix_l16[qscale][0][i] <<= 2;
231 ctx->qmatrix_l16[qscale][1][i] <<= 2;
232 ctx->qmatrix_c16[qscale][0][i] <<= 2;
233 ctx->qmatrix_c16[qscale][1][i] <<= 2;
238 for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
239 for (i = 1; i < 64; i++) {
240 int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]];
242 /* The quantization formula from the VC-3 standard is:
243 * quantized = sign(block[i]) * floor(abs(block[i]/s) * p /
244 * (qscale * weight_table[i]))
245 * Where p is 32 for 8-bit samples and 8 for 10-bit ones.
246 * The s factor compensates scaling of DCT coefficients done by
247 * the DCT routines, and therefore is not present in standard.
248 * It's 8 for 8-bit samples and 4 for 10-bit ones.
249 * We want values of ctx->qtmatrix_l and ctx->qtmatrix_r to be:
250 * ((1 << DNX10BIT_QMAT_SHIFT) * (p / s)) /
251 * (qscale * weight_table[i])
252 * For 10-bit samples, p / s == 2 */
253 ctx->qmatrix_l[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) /
254 (qscale * luma_weight_table[i]);
255 ctx->qmatrix_c[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) /
256 (qscale * chroma_weight_table[i]);
261 ctx->m.q_chroma_intra_matrix16 = ctx->qmatrix_c16;
262 ctx->m.q_chroma_intra_matrix = ctx->qmatrix_c;
263 ctx->m.q_intra_matrix16 = ctx->qmatrix_l16;
264 ctx->m.q_intra_matrix = ctx->qmatrix_l;
268 return AVERROR(ENOMEM);
271 static av_cold int dnxhd_init_rc(DNXHDEncContext *ctx)
273 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_rc, 8160 * (ctx->m.avctx->qmax + 1) * sizeof(RCEntry), fail);
274 if (ctx->m.avctx->mb_decision != FF_MB_DECISION_RD)
275 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_cmp,
276 ctx->m.mb_num * sizeof(RCCMPEntry), fail);
278 ctx->frame_bits = (ctx->cid_table->coding_unit_size -
279 640 - 4 - ctx->min_padding) * 8;
281 ctx->lambda = 2 << LAMBDA_FRAC_BITS; // qscale 2
284 return AVERROR(ENOMEM);
287 static av_cold int dnxhd_encode_init(AVCodecContext *avctx)
289 DNXHDEncContext *ctx = avctx->priv_data;
290 int i, index, bit_depth, ret;
292 switch (avctx->pix_fmt) {
293 case AV_PIX_FMT_YUV422P:
296 case AV_PIX_FMT_YUV422P10:
300 av_log(avctx, AV_LOG_ERROR,
301 "pixel format is incompatible with DNxHD\n");
302 return AVERROR(EINVAL);
305 ctx->cid = ff_dnxhd_find_cid(avctx, bit_depth);
307 av_log(avctx, AV_LOG_ERROR,
308 "video parameters incompatible with DNxHD. Valid DNxHD profiles:\n");
309 ff_dnxhd_print_profiles(avctx, AV_LOG_ERROR);
310 return AVERROR(EINVAL);
312 av_log(avctx, AV_LOG_DEBUG, "cid %d\n", ctx->cid);
314 index = ff_dnxhd_get_cid_table(ctx->cid);
315 av_assert0(index >= 0);
316 ctx->cid_table = &ff_dnxhd_cid_table[index];
318 ctx->m.avctx = avctx;
322 avctx->bits_per_raw_sample = ctx->cid_table->bit_depth;
324 ff_blockdsp_init(&ctx->bdsp, avctx);
325 ff_dct_common_init(&ctx->m);
326 ff_dct_encode_init(&ctx->m);
328 if (!ctx->m.dct_quantize)
329 ctx->m.dct_quantize = ff_dct_quantize_c;
331 if (ctx->cid_table->bit_depth == 10) {
332 ctx->m.dct_quantize = dnxhd_10bit_dct_quantize;
333 ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym;
334 ctx->block_width_l2 = 4;
336 ctx->get_pixels_8x4_sym = dnxhd_8bit_get_pixels_8x4_sym;
337 ctx->block_width_l2 = 3;
341 ff_dnxhdenc_init_x86(ctx);
343 ctx->m.mb_height = (avctx->height + 15) / 16;
344 ctx->m.mb_width = (avctx->width + 15) / 16;
346 if (avctx->flags & CODEC_FLAG_INTERLACED_DCT) {
348 ctx->m.mb_height /= 2;
351 ctx->m.mb_num = ctx->m.mb_height * ctx->m.mb_width;
353 if (avctx->intra_quant_bias != FF_DEFAULT_QUANT_BIAS)
354 ctx->m.intra_quant_bias = avctx->intra_quant_bias;
355 // XXX tune lbias/cbias
356 if ((ret = dnxhd_init_qmat(ctx, ctx->m.intra_quant_bias, 0)) < 0)
359 /* Avid Nitris hardware decoder requires a minimum amount of padding
360 * in the coding unit payload */
361 if (ctx->nitris_compat)
362 ctx->min_padding = 1600;
364 if ((ret = dnxhd_init_vlc(ctx)) < 0)
366 if ((ret = dnxhd_init_rc(ctx)) < 0)
369 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_size,
370 ctx->m.mb_height * sizeof(uint32_t), fail);
371 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_offs,
372 ctx->m.mb_height * sizeof(uint32_t), fail);
373 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_bits,
374 ctx->m.mb_num * sizeof(uint16_t), fail);
375 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_qscale,
376 ctx->m.mb_num * sizeof(uint8_t), fail);
378 avctx->coded_frame = av_frame_alloc();
379 if (!avctx->coded_frame)
380 return AVERROR(ENOMEM);
382 avctx->coded_frame->key_frame = 1;
383 avctx->coded_frame->pict_type = AV_PICTURE_TYPE_I;
385 if (avctx->thread_count > MAX_THREADS) {
386 av_log(avctx, AV_LOG_ERROR, "too many threads\n");
387 return AVERROR(EINVAL);
390 if (avctx->qmax <= 1) {
391 av_log(avctx, AV_LOG_ERROR, "qmax must be at least 2\n");
392 return AVERROR(EINVAL);
395 ctx->thread[0] = ctx;
396 for (i = 1; i < avctx->thread_count; i++) {
397 ctx->thread[i] = av_malloc(sizeof(DNXHDEncContext));
398 memcpy(ctx->thread[i], ctx, sizeof(DNXHDEncContext));
402 fail: // for FF_ALLOCZ_OR_GOTO
403 return AVERROR(ENOMEM);
406 static int dnxhd_write_header(AVCodecContext *avctx, uint8_t *buf)
408 DNXHDEncContext *ctx = avctx->priv_data;
409 static const uint8_t header_prefix[5] = { 0x00, 0x00, 0x02, 0x80, 0x01 };
413 memcpy(buf, header_prefix, 5);
414 buf[5] = ctx->interlaced ? ctx->cur_field + 2 : 0x01;
415 buf[6] = 0x80; // crc flag off
416 buf[7] = 0xa0; // reserved
417 AV_WB16(buf + 0x18, avctx->height >> ctx->interlaced); // ALPF
418 AV_WB16(buf + 0x1a, avctx->width); // SPL
419 AV_WB16(buf + 0x1d, avctx->height >> ctx->interlaced); // NAL
421 buf[0x21] = ctx->cid_table->bit_depth == 10 ? 0x58 : 0x38;
422 buf[0x22] = 0x88 + (ctx->interlaced << 2);
423 AV_WB32(buf + 0x28, ctx->cid); // CID
424 buf[0x2c] = ctx->interlaced ? 0 : 0x80;
426 buf[0x5f] = 0x01; // UDL
428 buf[0x167] = 0x02; // reserved
429 AV_WB16(buf + 0x16a, ctx->m.mb_height * 4 + 4); // MSIPS
430 buf[0x16d] = ctx->m.mb_height; // Ns
431 buf[0x16f] = 0x10; // reserved
433 ctx->msip = buf + 0x170;
437 static av_always_inline void dnxhd_encode_dc(DNXHDEncContext *ctx, int diff)
441 nbits = av_log2_16bit(-2 * diff);
444 nbits = av_log2_16bit(2 * diff);
446 put_bits(&ctx->m.pb, ctx->cid_table->dc_bits[nbits] + nbits,
447 (ctx->cid_table->dc_codes[nbits] << nbits) +
448 (diff & ((1 << nbits) - 1)));
451 static av_always_inline
452 void dnxhd_encode_block(DNXHDEncContext *ctx, int16_t *block,
453 int last_index, int n)
455 int last_non_zero = 0;
458 dnxhd_encode_dc(ctx, block[0] - ctx->m.last_dc[n]);
459 ctx->m.last_dc[n] = block[0];
461 for (i = 1; i <= last_index; i++) {
462 j = ctx->m.intra_scantable.permutated[i];
465 int run_level = i - last_non_zero - 1;
466 int rlevel = (slevel << 1) | !!run_level;
467 put_bits(&ctx->m.pb, ctx->vlc_bits[rlevel], ctx->vlc_codes[rlevel]);
469 put_bits(&ctx->m.pb, ctx->run_bits[run_level],
470 ctx->run_codes[run_level]);
474 put_bits(&ctx->m.pb, ctx->vlc_bits[0], ctx->vlc_codes[0]); // EOB
477 static av_always_inline
478 void dnxhd_unquantize_c(DNXHDEncContext *ctx, int16_t *block, int n,
479 int qscale, int last_index)
481 const uint8_t *weight_matrix;
485 weight_matrix = (n & 2) ? ctx->cid_table->chroma_weight
486 : ctx->cid_table->luma_weight;
488 for (i = 1; i <= last_index; i++) {
489 int j = ctx->m.intra_scantable.permutated[i];
493 level = (1 - 2 * level) * qscale * weight_matrix[i];
494 if (ctx->cid_table->bit_depth == 10) {
495 if (weight_matrix[i] != 8)
499 if (weight_matrix[i] != 32)
505 level = (2 * level + 1) * qscale * weight_matrix[i];
506 if (ctx->cid_table->bit_depth == 10) {
507 if (weight_matrix[i] != 8)
511 if (weight_matrix[i] != 32)
521 static av_always_inline int dnxhd_ssd_block(int16_t *qblock, int16_t *block)
525 for (i = 0; i < 64; i++)
526 score += (block[i] - qblock[i]) * (block[i] - qblock[i]);
530 static av_always_inline
531 int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, int16_t *block, int last_index)
533 int last_non_zero = 0;
536 for (i = 1; i <= last_index; i++) {
537 j = ctx->m.intra_scantable.permutated[i];
540 int run_level = i - last_non_zero - 1;
541 bits += ctx->vlc_bits[(level << 1) |
542 !!run_level] + ctx->run_bits[run_level];
549 static av_always_inline
550 void dnxhd_get_blocks(DNXHDEncContext *ctx, int mb_x, int mb_y)
552 const int bs = ctx->block_width_l2;
553 const int bw = 1 << bs;
554 const uint8_t *ptr_y = ctx->thread[0]->src[0] +
555 ((mb_y << 4) * ctx->m.linesize) + (mb_x << bs + 1);
556 const uint8_t *ptr_u = ctx->thread[0]->src[1] +
557 ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs);
558 const uint8_t *ptr_v = ctx->thread[0]->src[2] +
559 ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs);
560 DSPContext *dsp = &ctx->m.dsp;
562 dsp->get_pixels(ctx->blocks[0], ptr_y, ctx->m.linesize);
563 dsp->get_pixels(ctx->blocks[1], ptr_y + bw, ctx->m.linesize);
564 dsp->get_pixels(ctx->blocks[2], ptr_u, ctx->m.uvlinesize);
565 dsp->get_pixels(ctx->blocks[3], ptr_v, ctx->m.uvlinesize);
567 if (mb_y + 1 == ctx->m.mb_height && ctx->m.avctx->height == 1080) {
568 if (ctx->interlaced) {
569 ctx->get_pixels_8x4_sym(ctx->blocks[4],
570 ptr_y + ctx->dct_y_offset,
572 ctx->get_pixels_8x4_sym(ctx->blocks[5],
573 ptr_y + ctx->dct_y_offset + bw,
575 ctx->get_pixels_8x4_sym(ctx->blocks[6],
576 ptr_u + ctx->dct_uv_offset,
578 ctx->get_pixels_8x4_sym(ctx->blocks[7],
579 ptr_v + ctx->dct_uv_offset,
582 ctx->bdsp.clear_block(ctx->blocks[4]);
583 ctx->bdsp.clear_block(ctx->blocks[5]);
584 ctx->bdsp.clear_block(ctx->blocks[6]);
585 ctx->bdsp.clear_block(ctx->blocks[7]);
588 dsp->get_pixels(ctx->blocks[4],
589 ptr_y + ctx->dct_y_offset, ctx->m.linesize);
590 dsp->get_pixels(ctx->blocks[5],
591 ptr_y + ctx->dct_y_offset + bw, ctx->m.linesize);
592 dsp->get_pixels(ctx->blocks[6],
593 ptr_u + ctx->dct_uv_offset, ctx->m.uvlinesize);
594 dsp->get_pixels(ctx->blocks[7],
595 ptr_v + ctx->dct_uv_offset, ctx->m.uvlinesize);
599 static av_always_inline
600 int dnxhd_switch_matrix(DNXHDEncContext *ctx, int i)
602 const static uint8_t component[8]={0,0,1,2,0,0,1,2};
606 static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg,
607 int jobnr, int threadnr)
609 DNXHDEncContext *ctx = avctx->priv_data;
610 int mb_y = jobnr, mb_x;
611 int qscale = ctx->qscale;
612 LOCAL_ALIGNED_16(int16_t, block, [64]);
613 ctx = ctx->thread[threadnr];
617 ctx->m.last_dc[2] = 1 << (ctx->cid_table->bit_depth + 2);
619 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
620 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
626 dnxhd_get_blocks(ctx, mb_x, mb_y);
628 for (i = 0; i < 8; i++) {
629 int16_t *src_block = ctx->blocks[i];
630 int overflow, nbits, diff, last_index;
631 int n = dnxhd_switch_matrix(ctx, i);
633 memcpy(block, src_block, 64 * sizeof(*block));
634 last_index = ctx->m.dct_quantize(&ctx->m, block, 4 & (2*i),
636 ac_bits += dnxhd_calc_ac_bits(ctx, block, last_index);
638 diff = block[0] - ctx->m.last_dc[n];
640 nbits = av_log2_16bit(-2 * diff);
642 nbits = av_log2_16bit(2 * diff);
644 av_assert1(nbits < ctx->cid_table->bit_depth + 4);
645 dc_bits += ctx->cid_table->dc_bits[nbits] + nbits;
647 ctx->m.last_dc[n] = block[0];
649 if (avctx->mb_decision == FF_MB_DECISION_RD || !RC_VARIANCE) {
650 dnxhd_unquantize_c(ctx, block, i, qscale, last_index);
651 ctx->m.dsp.idct(block);
652 ssd += dnxhd_ssd_block(block, src_block);
655 ctx->mb_rc[qscale][mb].ssd = ssd;
656 ctx->mb_rc[qscale][mb].bits = ac_bits + dc_bits + 12 +
657 8 * ctx->vlc_bits[0];
662 static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg,
663 int jobnr, int threadnr)
665 DNXHDEncContext *ctx = avctx->priv_data;
666 int mb_y = jobnr, mb_x;
667 ctx = ctx->thread[threadnr];
668 init_put_bits(&ctx->m.pb, (uint8_t *)arg + 640 + ctx->slice_offs[jobnr],
669 ctx->slice_size[jobnr]);
673 ctx->m.last_dc[2] = 1 << (ctx->cid_table->bit_depth + 2);
674 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
675 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
676 int qscale = ctx->mb_qscale[mb];
679 put_bits(&ctx->m.pb, 12, qscale << 1);
681 dnxhd_get_blocks(ctx, mb_x, mb_y);
683 for (i = 0; i < 8; i++) {
684 int16_t *block = ctx->blocks[i];
685 int overflow, n = dnxhd_switch_matrix(ctx, i);
686 int last_index = ctx->m.dct_quantize(&ctx->m, block, 4 & (2*i),
689 dnxhd_encode_block(ctx, block, last_index, n);
690 // STOP_TIMER("encode_block");
693 if (put_bits_count(&ctx->m.pb) & 31)
694 put_bits(&ctx->m.pb, 32 - (put_bits_count(&ctx->m.pb) & 31), 0);
695 flush_put_bits(&ctx->m.pb);
699 static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx)
703 for (mb_y = 0; mb_y < ctx->m.mb_height; mb_y++) {
705 ctx->slice_offs[mb_y] = offset;
706 ctx->slice_size[mb_y] = 0;
707 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
708 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
709 ctx->slice_size[mb_y] += ctx->mb_bits[mb];
711 ctx->slice_size[mb_y] = (ctx->slice_size[mb_y] + 31) & ~31;
712 ctx->slice_size[mb_y] >>= 3;
713 thread_size = ctx->slice_size[mb_y];
714 offset += thread_size;
718 static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg,
719 int jobnr, int threadnr)
721 DNXHDEncContext *ctx = avctx->priv_data;
722 int mb_y = jobnr, mb_x, x, y;
723 int partial_last_row = (mb_y == ctx->m.mb_height - 1) &&
724 ((avctx->height >> ctx->interlaced) & 0xF);
726 ctx = ctx->thread[threadnr];
727 if (ctx->cid_table->bit_depth == 8) {
728 uint8_t *pix = ctx->thread[0]->src[0] + ((mb_y << 4) * ctx->m.linesize);
729 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x, pix += 16) {
730 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
734 if (!partial_last_row && mb_x * 16 <= avctx->width - 16) {
735 sum = ctx->m.dsp.pix_sum(pix, ctx->m.linesize);
736 varc = ctx->m.dsp.pix_norm1(pix, ctx->m.linesize);
738 int bw = FFMIN(avctx->width - 16 * mb_x, 16);
739 int bh = FFMIN((avctx->height >> ctx->interlaced) - 16 * mb_y, 16);
741 for (y = 0; y < bh; y++) {
742 for (x = 0; x < bw; x++) {
743 uint8_t val = pix[x + y * ctx->m.linesize];
749 varc = (varc - (((unsigned) sum * sum) >> 8) + 128) >> 8;
751 ctx->mb_cmp[mb].value = varc;
752 ctx->mb_cmp[mb].mb = mb;
755 int const linesize = ctx->m.linesize >> 1;
756 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x) {
757 uint16_t *pix = (uint16_t *)ctx->thread[0]->src[0] +
758 ((mb_y << 4) * linesize) + (mb_x << 4);
759 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
764 // Macroblocks are 16x16 pixels, unlike DCT blocks which are 8x8.
765 for (i = 0; i < 16; ++i) {
766 for (j = 0; j < 16; ++j) {
767 // Turn 16-bit pixels into 10-bit ones.
768 int const sample = (unsigned) pix[j] >> 6;
770 sqsum += sample * sample;
771 // 2^10 * 2^10 * 16 * 16 = 2^28, which is less than INT_MAX
775 mean = sum >> 8; // 16*16 == 2^8
777 ctx->mb_cmp[mb].value = sqmean - mean * mean;
778 ctx->mb_cmp[mb].mb = mb;
784 static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx)
786 int lambda, up_step, down_step;
787 int last_lower = INT_MAX, last_higher = 0;
790 for (q = 1; q < avctx->qmax; q++) {
792 avctx->execute2(avctx, dnxhd_calc_bits_thread,
793 NULL, NULL, ctx->m.mb_height);
795 up_step = down_step = 2 << LAMBDA_FRAC_BITS;
796 lambda = ctx->lambda;
801 if (lambda == last_higher) {
803 end = 1; // need to set final qscales/bits
805 for (y = 0; y < ctx->m.mb_height; y++) {
806 for (x = 0; x < ctx->m.mb_width; x++) {
807 unsigned min = UINT_MAX;
809 int mb = y * ctx->m.mb_width + x;
810 for (q = 1; q < avctx->qmax; q++) {
811 unsigned score = ctx->mb_rc[q][mb].bits * lambda +
812 ((unsigned) ctx->mb_rc[q][mb].ssd << LAMBDA_FRAC_BITS);
818 bits += ctx->mb_rc[qscale][mb].bits;
819 ctx->mb_qscale[mb] = qscale;
820 ctx->mb_bits[mb] = ctx->mb_rc[qscale][mb].bits;
822 bits = (bits + 31) & ~31; // padding
823 if (bits > ctx->frame_bits)
826 // av_dlog(ctx->m.avctx,
827 // "lambda %d, up %u, down %u, bits %d, frame %d\n",
828 // lambda, last_higher, last_lower, bits, ctx->frame_bits);
830 if (bits > ctx->frame_bits)
831 return AVERROR(EINVAL);
834 if (bits < ctx->frame_bits) {
835 last_lower = FFMIN(lambda, last_lower);
836 if (last_higher != 0)
837 lambda = (lambda+last_higher)>>1;
840 down_step = FFMIN((int64_t)down_step*5, INT_MAX);
841 up_step = 1<<LAMBDA_FRAC_BITS;
842 lambda = FFMAX(1, lambda);
843 if (lambda == last_lower)
846 last_higher = FFMAX(lambda, last_higher);
847 if (last_lower != INT_MAX)
848 lambda = (lambda+last_lower)>>1;
849 else if ((int64_t)lambda + up_step > INT_MAX)
850 return AVERROR(EINVAL);
853 up_step = FFMIN((int64_t)up_step*5, INT_MAX);
854 down_step = 1<<LAMBDA_FRAC_BITS;
857 //av_dlog(ctx->m.avctx, "out lambda %d\n", lambda);
858 ctx->lambda = lambda;
862 static int dnxhd_find_qscale(DNXHDEncContext *ctx)
868 int last_lower = INT_MAX;
872 qscale = ctx->qscale;
875 ctx->qscale = qscale;
876 // XXX avoid recalculating bits
877 ctx->m.avctx->execute2(ctx->m.avctx, dnxhd_calc_bits_thread,
878 NULL, NULL, ctx->m.mb_height);
879 for (y = 0; y < ctx->m.mb_height; y++) {
880 for (x = 0; x < ctx->m.mb_width; x++)
881 bits += ctx->mb_rc[qscale][y*ctx->m.mb_width+x].bits;
882 bits = (bits+31)&~31; // padding
883 if (bits > ctx->frame_bits)
886 // av_dlog(ctx->m.avctx,
887 // "%d, qscale %d, bits %d, frame %d, higher %d, lower %d\n",
888 // ctx->m.avctx->frame_number, qscale, bits, ctx->frame_bits,
889 // last_higher, last_lower);
890 if (bits < ctx->frame_bits) {
893 if (last_higher == qscale - 1) {
894 qscale = last_higher;
897 last_lower = FFMIN(qscale, last_lower);
898 if (last_higher != 0)
899 qscale = (qscale + last_higher) >> 1;
901 qscale -= down_step++;
906 if (last_lower == qscale + 1)
908 last_higher = FFMAX(qscale, last_higher);
909 if (last_lower != INT_MAX)
910 qscale = (qscale + last_lower) >> 1;
914 if (qscale >= ctx->m.avctx->qmax)
915 return AVERROR(EINVAL);
918 //av_dlog(ctx->m.avctx, "out qscale %d\n", qscale);
919 ctx->qscale = qscale;
923 #define BUCKET_BITS 8
924 #define RADIX_PASSES 4
925 #define NBUCKETS (1 << BUCKET_BITS)
927 static inline int get_bucket(int value, int shift)
930 value &= NBUCKETS - 1;
931 return NBUCKETS - 1 - value;
934 static void radix_count(const RCCMPEntry *data, int size,
935 int buckets[RADIX_PASSES][NBUCKETS])
938 memset(buckets, 0, sizeof(buckets[0][0]) * RADIX_PASSES * NBUCKETS);
939 for (i = 0; i < size; i++) {
940 int v = data[i].value;
941 for (j = 0; j < RADIX_PASSES; j++) {
942 buckets[j][get_bucket(v, 0)]++;
947 for (j = 0; j < RADIX_PASSES; j++) {
949 for (i = NBUCKETS - 1; i >= 0; i--)
950 buckets[j][i] = offset -= buckets[j][i];
951 av_assert1(!buckets[j][0]);
955 static void radix_sort_pass(RCCMPEntry *dst, const RCCMPEntry *data,
956 int size, int buckets[NBUCKETS], int pass)
958 int shift = pass * BUCKET_BITS;
960 for (i = 0; i < size; i++) {
961 int v = get_bucket(data[i].value, shift);
962 int pos = buckets[v]++;
967 static void radix_sort(RCCMPEntry *data, int size)
969 int buckets[RADIX_PASSES][NBUCKETS];
970 RCCMPEntry *tmp = av_malloc_array(size, sizeof(*tmp));
971 radix_count(data, size, buckets);
972 radix_sort_pass(tmp, data, size, buckets[0], 0);
973 radix_sort_pass(data, tmp, size, buckets[1], 1);
974 if (buckets[2][NBUCKETS - 1] || buckets[3][NBUCKETS - 1]) {
975 radix_sort_pass(tmp, data, size, buckets[2], 2);
976 radix_sort_pass(data, tmp, size, buckets[3], 3);
981 static int dnxhd_encode_fast(AVCodecContext *avctx, DNXHDEncContext *ctx)
985 if ((ret = dnxhd_find_qscale(ctx)) < 0)
987 for (y = 0; y < ctx->m.mb_height; y++) {
988 for (x = 0; x < ctx->m.mb_width; x++) {
989 int mb = y * ctx->m.mb_width + x;
991 ctx->mb_qscale[mb] = ctx->qscale;
992 ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale][mb].bits;
993 max_bits += ctx->mb_rc[ctx->qscale][mb].bits;
995 delta_bits = ctx->mb_rc[ctx->qscale][mb].bits -
996 ctx->mb_rc[ctx->qscale + 1][mb].bits;
997 ctx->mb_cmp[mb].mb = mb;
998 ctx->mb_cmp[mb].value =
999 delta_bits ? ((ctx->mb_rc[ctx->qscale][mb].ssd -
1000 ctx->mb_rc[ctx->qscale + 1][mb].ssd) * 100) /
1002 : INT_MIN; // avoid increasing qscale
1005 max_bits += 31; // worst padding
1009 avctx->execute2(avctx, dnxhd_mb_var_thread,
1010 NULL, NULL, ctx->m.mb_height);
1011 radix_sort(ctx->mb_cmp, ctx->m.mb_num);
1012 for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) {
1013 int mb = ctx->mb_cmp[x].mb;
1014 max_bits -= ctx->mb_rc[ctx->qscale][mb].bits -
1015 ctx->mb_rc[ctx->qscale + 1][mb].bits;
1016 ctx->mb_qscale[mb] = ctx->qscale + 1;
1017 ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale + 1][mb].bits;
1023 static void dnxhd_load_picture(DNXHDEncContext *ctx, const AVFrame *frame)
1027 for (i = 0; i < ctx->m.avctx->thread_count; i++) {
1028 ctx->thread[i]->m.linesize = frame->linesize[0] << ctx->interlaced;
1029 ctx->thread[i]->m.uvlinesize = frame->linesize[1] << ctx->interlaced;
1030 ctx->thread[i]->dct_y_offset = ctx->m.linesize *8;
1031 ctx->thread[i]->dct_uv_offset = ctx->m.uvlinesize*8;
1034 ctx->m.avctx->coded_frame->interlaced_frame = frame->interlaced_frame;
1035 ctx->cur_field = frame->interlaced_frame && !frame->top_field_first;
1038 static int dnxhd_encode_picture(AVCodecContext *avctx, AVPacket *pkt,
1039 const AVFrame *frame, int *got_packet)
1041 DNXHDEncContext *ctx = avctx->priv_data;
1042 int first_field = 1;
1046 if ((ret = ff_alloc_packet2(avctx, pkt, ctx->cid_table->frame_size)) < 0)
1050 dnxhd_load_picture(ctx, frame);
1053 for (i = 0; i < 3; i++) {
1054 ctx->src[i] = frame->data[i];
1055 if (ctx->interlaced && ctx->cur_field)
1056 ctx->src[i] += frame->linesize[i];
1059 dnxhd_write_header(avctx, buf);
1061 if (avctx->mb_decision == FF_MB_DECISION_RD)
1062 ret = dnxhd_encode_rdo(avctx, ctx);
1064 ret = dnxhd_encode_fast(avctx, ctx);
1066 av_log(avctx, AV_LOG_ERROR,
1067 "picture could not fit ratecontrol constraints, increase qmax\n");
1071 dnxhd_setup_threads_slices(ctx);
1074 for (i = 0; i < ctx->m.mb_height; i++) {
1075 AV_WB32(ctx->msip + i * 4, offset);
1076 offset += ctx->slice_size[i];
1077 av_assert1(!(ctx->slice_size[i] & 3));
1080 avctx->execute2(avctx, dnxhd_encode_thread, buf, NULL, ctx->m.mb_height);
1082 av_assert1(640 + offset + 4 <= ctx->cid_table->coding_unit_size);
1083 memset(buf + 640 + offset, 0,
1084 ctx->cid_table->coding_unit_size - 4 - offset - 640);
1086 AV_WB32(buf + ctx->cid_table->coding_unit_size - 4, 0x600DC0DE); // EOF
1088 if (ctx->interlaced && first_field) {
1090 ctx->cur_field ^= 1;
1091 buf += ctx->cid_table->coding_unit_size;
1092 goto encode_coding_unit;
1095 avctx->coded_frame->quality = ctx->qscale * FF_QP2LAMBDA;
1097 pkt->flags |= AV_PKT_FLAG_KEY;
1102 static av_cold int dnxhd_encode_end(AVCodecContext *avctx)
1104 DNXHDEncContext *ctx = avctx->priv_data;
1105 int max_level = 1 << (ctx->cid_table->bit_depth + 2);
1108 av_free(ctx->vlc_codes - max_level * 2);
1109 av_free(ctx->vlc_bits - max_level * 2);
1110 av_freep(&ctx->run_codes);
1111 av_freep(&ctx->run_bits);
1113 av_freep(&ctx->mb_bits);
1114 av_freep(&ctx->mb_qscale);
1115 av_freep(&ctx->mb_rc);
1116 av_freep(&ctx->mb_cmp);
1117 av_freep(&ctx->slice_size);
1118 av_freep(&ctx->slice_offs);
1120 av_freep(&ctx->qmatrix_c);
1121 av_freep(&ctx->qmatrix_l);
1122 av_freep(&ctx->qmatrix_c16);
1123 av_freep(&ctx->qmatrix_l16);
1125 for (i = 1; i < avctx->thread_count; i++)
1126 av_freep(&ctx->thread[i]);
1128 av_frame_free(&avctx->coded_frame);
1133 static const AVCodecDefault dnxhd_defaults[] = {
1134 { "qmax", "1024" }, /* Maximum quantization scale factor allowed for VC-3 */
1138 AVCodec ff_dnxhd_encoder = {
1140 .long_name = NULL_IF_CONFIG_SMALL("VC3/DNxHD"),
1141 .type = AVMEDIA_TYPE_VIDEO,
1142 .id = AV_CODEC_ID_DNXHD,
1143 .priv_data_size = sizeof(DNXHDEncContext),
1144 .init = dnxhd_encode_init,
1145 .encode2 = dnxhd_encode_picture,
1146 .close = dnxhd_encode_end,
1147 .capabilities = CODEC_CAP_SLICE_THREADS,
1148 .pix_fmts = (const enum AVPixelFormat[]) {
1150 AV_PIX_FMT_YUV422P10,
1153 .priv_class = &dnxhd_class,
1154 .defaults = dnxhd_defaults,