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 Libav.
11 * Libav 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 * Libav 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 Libav; 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"
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 class = {
58 static void dnxhd_8bit_get_pixels_8x4_sym(int16_t *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 *restrict block,
83 const uint8_t *pixels,
90 for (i = 0; i < 4; i++) {
91 memcpy(block + i * 8, pixels + i * line_size, 8 * sizeof(*block));
92 memcpy(block - (i + 1) * 8, pixels + i * line_size, 8 * sizeof(*block));
96 static int dnxhd_10bit_dct_quantize(MpegEncContext *ctx, int16_t *block,
97 int n, int qscale, int *overflow)
99 const uint8_t *scantable= ctx->intra_scantable.scantable;
100 const int *qmat = ctx->q_intra_matrix[qscale];
101 int last_non_zero = 0;
104 ctx->fdsp.fdct(block);
106 // Divide by 4 with rounding, to compensate scaling of DCT coefficients
107 block[0] = (block[0] + 2) >> 2;
109 for (i = 1; i < 64; ++i) {
110 int j = scantable[i];
111 int sign = block[j] >> 31;
112 int level = (block[j] ^ sign) - sign;
113 level = level * qmat[j] >> DNX10BIT_QMAT_SHIFT;
114 block[j] = (level ^ sign) - sign;
119 return last_non_zero;
122 static av_cold int dnxhd_init_vlc(DNXHDEncContext *ctx)
124 int i, j, level, run;
125 int max_level = 1 << (ctx->cid_table->bit_depth + 2);
127 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->vlc_codes,
128 max_level * 4 * sizeof(*ctx->vlc_codes), fail);
129 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->vlc_bits,
130 max_level * 4 * sizeof(*ctx->vlc_bits), fail);
131 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_codes,
133 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_bits,
136 ctx->vlc_codes += max_level * 2;
137 ctx->vlc_bits += max_level * 2;
138 for (level = -max_level; level < max_level; level++) {
139 for (run = 0; run < 2; run++) {
140 int index = (level << 1) | run;
141 int sign, offset = 0, alevel = level;
143 MASK_ABS(sign, alevel);
145 offset = (alevel - 1) >> 6;
146 alevel -= offset << 6;
148 for (j = 0; j < 257; j++) {
149 if (ctx->cid_table->ac_level[j] == alevel &&
150 (!offset || (ctx->cid_table->ac_index_flag[j] && offset)) &&
151 (!run || (ctx->cid_table->ac_run_flag [j] && run))) {
152 assert(!ctx->vlc_codes[index]);
154 ctx->vlc_codes[index] =
155 (ctx->cid_table->ac_codes[j] << 1) | (sign & 1);
156 ctx->vlc_bits[index] = ctx->cid_table->ac_bits[j] + 1;
158 ctx->vlc_codes[index] = ctx->cid_table->ac_codes[j];
159 ctx->vlc_bits[index] = ctx->cid_table->ac_bits[j];
164 assert(!alevel || j < 257);
166 ctx->vlc_codes[index] =
167 (ctx->vlc_codes[index] << ctx->cid_table->index_bits) | offset;
168 ctx->vlc_bits[index] += ctx->cid_table->index_bits;
172 for (i = 0; i < 62; i++) {
173 int run = ctx->cid_table->run[i];
175 ctx->run_codes[run] = ctx->cid_table->run_codes[i];
176 ctx->run_bits[run] = ctx->cid_table->run_bits[i];
180 return AVERROR(ENOMEM);
183 static av_cold int dnxhd_init_qmat(DNXHDEncContext *ctx, int lbias, int cbias)
185 // init first elem to 1 to avoid div by 0 in convert_matrix
186 uint16_t weight_matrix[64] = { 1, }; // convert_matrix needs uint16_t*
188 const uint8_t *luma_weight_table = ctx->cid_table->luma_weight;
189 const uint8_t *chroma_weight_table = ctx->cid_table->chroma_weight;
191 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l,
192 (ctx->m.avctx->qmax + 1) * 64 * sizeof(int), fail);
193 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c,
194 (ctx->m.avctx->qmax + 1) * 64 * sizeof(int), fail);
195 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l16,
196 (ctx->m.avctx->qmax + 1) * 64 * 2 * sizeof(uint16_t),
198 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c16,
199 (ctx->m.avctx->qmax + 1) * 64 * 2 * sizeof(uint16_t),
202 if (ctx->cid_table->bit_depth == 8) {
203 for (i = 1; i < 64; i++) {
204 int j = ctx->m.idsp.idct_permutation[ff_zigzag_direct[i]];
205 weight_matrix[j] = ctx->cid_table->luma_weight[i];
207 ff_convert_matrix(&ctx->m, ctx->qmatrix_l, ctx->qmatrix_l16,
208 weight_matrix, ctx->m.intra_quant_bias, 1,
209 ctx->m.avctx->qmax, 1);
210 for (i = 1; i < 64; i++) {
211 int j = ctx->m.idsp.idct_permutation[ff_zigzag_direct[i]];
212 weight_matrix[j] = ctx->cid_table->chroma_weight[i];
214 ff_convert_matrix(&ctx->m, ctx->qmatrix_c, ctx->qmatrix_c16,
215 weight_matrix, ctx->m.intra_quant_bias, 1,
216 ctx->m.avctx->qmax, 1);
218 for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
219 for (i = 0; i < 64; i++) {
220 ctx->qmatrix_l[qscale][i] <<= 2;
221 ctx->qmatrix_c[qscale][i] <<= 2;
222 ctx->qmatrix_l16[qscale][0][i] <<= 2;
223 ctx->qmatrix_l16[qscale][1][i] <<= 2;
224 ctx->qmatrix_c16[qscale][0][i] <<= 2;
225 ctx->qmatrix_c16[qscale][1][i] <<= 2;
230 for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
231 for (i = 1; i < 64; i++) {
232 int j = ctx->m.idsp.idct_permutation[ff_zigzag_direct[i]];
234 /* The quantization formula from the VC-3 standard is:
235 * quantized = sign(block[i]) * floor(abs(block[i]/s) * p /
236 * (qscale * weight_table[i]))
237 * Where p is 32 for 8-bit samples and 8 for 10-bit ones.
238 * The s factor compensates scaling of DCT coefficients done by
239 * the DCT routines, and therefore is not present in standard.
240 * It's 8 for 8-bit samples and 4 for 10-bit ones.
241 * We want values of ctx->qtmatrix_l and ctx->qtmatrix_r to be:
242 * ((1 << DNX10BIT_QMAT_SHIFT) * (p / s)) /
243 * (qscale * weight_table[i])
244 * For 10-bit samples, p / s == 2 */
245 ctx->qmatrix_l[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) /
246 (qscale * luma_weight_table[i]);
247 ctx->qmatrix_c[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) /
248 (qscale * chroma_weight_table[i]);
255 return AVERROR(ENOMEM);
258 static av_cold int dnxhd_init_rc(DNXHDEncContext *ctx)
260 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_rc,
261 8160 * ctx->m.avctx->qmax * sizeof(RCEntry), fail);
262 if (ctx->m.avctx->mb_decision != FF_MB_DECISION_RD)
263 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_cmp,
264 ctx->m.mb_num * sizeof(RCCMPEntry), fail);
266 ctx->frame_bits = (ctx->cid_table->coding_unit_size -
267 640 - 4 - ctx->min_padding) * 8;
269 ctx->lambda = 2 << LAMBDA_FRAC_BITS; // qscale 2
272 return AVERROR(ENOMEM);
275 static av_cold int dnxhd_encode_init(AVCodecContext *avctx)
277 DNXHDEncContext *ctx = avctx->priv_data;
278 int i, index, bit_depth, ret;
280 switch (avctx->pix_fmt) {
281 case AV_PIX_FMT_YUV422P:
284 case AV_PIX_FMT_YUV422P10:
288 av_log(avctx, AV_LOG_ERROR,
289 "pixel format is incompatible with DNxHD\n");
290 return AVERROR(EINVAL);
293 ctx->cid = ff_dnxhd_find_cid(avctx, bit_depth);
295 av_log(avctx, AV_LOG_ERROR,
296 "video parameters incompatible with DNxHD\n");
297 return AVERROR(EINVAL);
299 av_log(avctx, AV_LOG_DEBUG, "cid %d\n", ctx->cid);
301 index = ff_dnxhd_get_cid_table(ctx->cid);
302 ctx->cid_table = &ff_dnxhd_cid_table[index];
304 ctx->m.avctx = avctx;
308 avctx->bits_per_raw_sample = ctx->cid_table->bit_depth;
310 ff_blockdsp_init(&ctx->bdsp, avctx);
311 ff_fdctdsp_init(&ctx->m.fdsp, avctx);
312 ff_mpv_idct_init(&ctx->m);
313 ff_mpegvideoencdsp_init(&ctx->m.mpvencdsp, avctx);
314 ff_pixblockdsp_init(&ctx->m.pdsp, avctx);
315 if (!ctx->m.dct_quantize)
316 ctx->m.dct_quantize = ff_dct_quantize_c;
318 if (ctx->cid_table->bit_depth == 10) {
319 ctx->m.dct_quantize = dnxhd_10bit_dct_quantize;
320 ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym;
321 ctx->block_width_l2 = 4;
323 ctx->get_pixels_8x4_sym = dnxhd_8bit_get_pixels_8x4_sym;
324 ctx->block_width_l2 = 3;
328 ff_dnxhdenc_init_x86(ctx);
330 ctx->m.mb_height = (avctx->height + 15) / 16;
331 ctx->m.mb_width = (avctx->width + 15) / 16;
333 if (avctx->flags & CODEC_FLAG_INTERLACED_DCT) {
335 ctx->m.mb_height /= 2;
338 ctx->m.mb_num = ctx->m.mb_height * ctx->m.mb_width;
340 if (avctx->intra_quant_bias != FF_DEFAULT_QUANT_BIAS)
341 ctx->m.intra_quant_bias = avctx->intra_quant_bias;
342 // XXX tune lbias/cbias
343 if ((ret = dnxhd_init_qmat(ctx, ctx->m.intra_quant_bias, 0)) < 0)
346 /* Avid Nitris hardware decoder requires a minimum amount of padding
347 * in the coding unit payload */
348 if (ctx->nitris_compat)
349 ctx->min_padding = 1600;
351 if ((ret = dnxhd_init_vlc(ctx)) < 0)
353 if ((ret = dnxhd_init_rc(ctx)) < 0)
356 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_size,
357 ctx->m.mb_height * sizeof(uint32_t), fail);
358 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_offs,
359 ctx->m.mb_height * sizeof(uint32_t), fail);
360 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_bits,
361 ctx->m.mb_num * sizeof(uint16_t), fail);
362 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_qscale,
363 ctx->m.mb_num * sizeof(uint8_t), fail);
365 avctx->coded_frame = av_frame_alloc();
366 if (!avctx->coded_frame)
367 return AVERROR(ENOMEM);
369 avctx->coded_frame->key_frame = 1;
370 avctx->coded_frame->pict_type = AV_PICTURE_TYPE_I;
372 if (avctx->thread_count > MAX_THREADS) {
373 av_log(avctx, AV_LOG_ERROR, "too many threads\n");
374 return AVERROR(EINVAL);
377 ctx->thread[0] = ctx;
378 for (i = 1; i < avctx->thread_count; i++) {
379 ctx->thread[i] = av_malloc(sizeof(DNXHDEncContext));
380 memcpy(ctx->thread[i], ctx, sizeof(DNXHDEncContext));
384 fail: // for FF_ALLOCZ_OR_GOTO
385 return AVERROR(ENOMEM);
388 static int dnxhd_write_header(AVCodecContext *avctx, uint8_t *buf)
390 DNXHDEncContext *ctx = avctx->priv_data;
391 const uint8_t header_prefix[5] = { 0x00, 0x00, 0x02, 0x80, 0x01 };
395 memcpy(buf, header_prefix, 5);
396 buf[5] = ctx->interlaced ? ctx->cur_field + 2 : 0x01;
397 buf[6] = 0x80; // crc flag off
398 buf[7] = 0xa0; // reserved
399 AV_WB16(buf + 0x18, avctx->height >> ctx->interlaced); // ALPF
400 AV_WB16(buf + 0x1a, avctx->width); // SPL
401 AV_WB16(buf + 0x1d, avctx->height >> ctx->interlaced); // NAL
403 buf[0x21] = ctx->cid_table->bit_depth == 10 ? 0x58 : 0x38;
404 buf[0x22] = 0x88 + (ctx->interlaced << 2);
405 AV_WB32(buf + 0x28, ctx->cid); // CID
406 buf[0x2c] = ctx->interlaced ? 0 : 0x80;
408 buf[0x5f] = 0x01; // UDL
410 buf[0x167] = 0x02; // reserved
411 AV_WB16(buf + 0x16a, ctx->m.mb_height * 4 + 4); // MSIPS
412 buf[0x16d] = ctx->m.mb_height; // Ns
413 buf[0x16f] = 0x10; // reserved
415 ctx->msip = buf + 0x170;
419 static av_always_inline void dnxhd_encode_dc(DNXHDEncContext *ctx, int diff)
423 nbits = av_log2_16bit(-2 * diff);
426 nbits = av_log2_16bit(2 * diff);
428 put_bits(&ctx->m.pb, ctx->cid_table->dc_bits[nbits] + nbits,
429 (ctx->cid_table->dc_codes[nbits] << nbits) +
430 (diff & ((1 << nbits) - 1)));
433 static av_always_inline
434 void dnxhd_encode_block(DNXHDEncContext *ctx, int16_t *block,
435 int last_index, int n)
437 int last_non_zero = 0;
440 dnxhd_encode_dc(ctx, block[0] - ctx->m.last_dc[n]);
441 ctx->m.last_dc[n] = block[0];
443 for (i = 1; i <= last_index; i++) {
444 j = ctx->m.intra_scantable.permutated[i];
447 int run_level = i - last_non_zero - 1;
448 int rlevel = (slevel << 1) | !!run_level;
449 put_bits(&ctx->m.pb, ctx->vlc_bits[rlevel], ctx->vlc_codes[rlevel]);
451 put_bits(&ctx->m.pb, ctx->run_bits[run_level],
452 ctx->run_codes[run_level]);
456 put_bits(&ctx->m.pb, ctx->vlc_bits[0], ctx->vlc_codes[0]); // EOB
459 static av_always_inline
460 void dnxhd_unquantize_c(DNXHDEncContext *ctx, int16_t *block, int n,
461 int qscale, int last_index)
463 const uint8_t *weight_matrix;
467 weight_matrix = (n & 2) ? ctx->cid_table->chroma_weight
468 : ctx->cid_table->luma_weight;
470 for (i = 1; i <= last_index; i++) {
471 int j = ctx->m.intra_scantable.permutated[i];
475 level = (1 - 2 * level) * qscale * weight_matrix[i];
476 if (ctx->cid_table->bit_depth == 10) {
477 if (weight_matrix[i] != 8)
481 if (weight_matrix[i] != 32)
487 level = (2 * level + 1) * qscale * weight_matrix[i];
488 if (ctx->cid_table->bit_depth == 10) {
489 if (weight_matrix[i] != 8)
493 if (weight_matrix[i] != 32)
503 static av_always_inline int dnxhd_ssd_block(int16_t *qblock, int16_t *block)
507 for (i = 0; i < 64; i++)
508 score += (block[i] - qblock[i]) * (block[i] - qblock[i]);
512 static av_always_inline
513 int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, int16_t *block, int last_index)
515 int last_non_zero = 0;
518 for (i = 1; i <= last_index; i++) {
519 j = ctx->m.intra_scantable.permutated[i];
522 int run_level = i - last_non_zero - 1;
523 bits += ctx->vlc_bits[(level << 1) |
524 !!run_level] + ctx->run_bits[run_level];
531 static av_always_inline
532 void dnxhd_get_blocks(DNXHDEncContext *ctx, int mb_x, int mb_y)
534 const int bs = ctx->block_width_l2;
535 const int bw = 1 << bs;
536 const uint8_t *ptr_y = ctx->thread[0]->src[0] +
537 ((mb_y << 4) * ctx->m.linesize) + (mb_x << bs + 1);
538 const uint8_t *ptr_u = ctx->thread[0]->src[1] +
539 ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs);
540 const uint8_t *ptr_v = ctx->thread[0]->src[2] +
541 ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs);
542 PixblockDSPContext *pdsp = &ctx->m.pdsp;
544 pdsp->get_pixels(ctx->blocks[0], ptr_y, ctx->m.linesize);
545 pdsp->get_pixels(ctx->blocks[1], ptr_y + bw, ctx->m.linesize);
546 pdsp->get_pixels(ctx->blocks[2], ptr_u, ctx->m.uvlinesize);
547 pdsp->get_pixels(ctx->blocks[3], ptr_v, ctx->m.uvlinesize);
549 if (mb_y + 1 == ctx->m.mb_height && ctx->m.avctx->height == 1080) {
550 if (ctx->interlaced) {
551 ctx->get_pixels_8x4_sym(ctx->blocks[4],
552 ptr_y + ctx->dct_y_offset,
554 ctx->get_pixels_8x4_sym(ctx->blocks[5],
555 ptr_y + ctx->dct_y_offset + bw,
557 ctx->get_pixels_8x4_sym(ctx->blocks[6],
558 ptr_u + ctx->dct_uv_offset,
560 ctx->get_pixels_8x4_sym(ctx->blocks[7],
561 ptr_v + ctx->dct_uv_offset,
564 ctx->bdsp.clear_block(ctx->blocks[4]);
565 ctx->bdsp.clear_block(ctx->blocks[5]);
566 ctx->bdsp.clear_block(ctx->blocks[6]);
567 ctx->bdsp.clear_block(ctx->blocks[7]);
570 pdsp->get_pixels(ctx->blocks[4],
571 ptr_y + ctx->dct_y_offset, ctx->m.linesize);
572 pdsp->get_pixels(ctx->blocks[5],
573 ptr_y + ctx->dct_y_offset + bw, ctx->m.linesize);
574 pdsp->get_pixels(ctx->blocks[6],
575 ptr_u + ctx->dct_uv_offset, ctx->m.uvlinesize);
576 pdsp->get_pixels(ctx->blocks[7],
577 ptr_v + ctx->dct_uv_offset, ctx->m.uvlinesize);
581 static av_always_inline
582 int dnxhd_switch_matrix(DNXHDEncContext *ctx, int i)
585 ctx->m.q_intra_matrix16 = ctx->qmatrix_c16;
586 ctx->m.q_intra_matrix = ctx->qmatrix_c;
589 ctx->m.q_intra_matrix16 = ctx->qmatrix_l16;
590 ctx->m.q_intra_matrix = ctx->qmatrix_l;
595 static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg,
596 int jobnr, int threadnr)
598 DNXHDEncContext *ctx = avctx->priv_data;
599 int mb_y = jobnr, mb_x;
600 int qscale = ctx->qscale;
601 LOCAL_ALIGNED_16(int16_t, block, [64]);
602 ctx = ctx->thread[threadnr];
606 ctx->m.last_dc[2] = 1 << (ctx->cid_table->bit_depth + 2);
608 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
609 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
615 dnxhd_get_blocks(ctx, mb_x, mb_y);
617 for (i = 0; i < 8; i++) {
618 int16_t *src_block = ctx->blocks[i];
619 int overflow, nbits, diff, last_index;
620 int n = dnxhd_switch_matrix(ctx, i);
622 memcpy(block, src_block, 64 * sizeof(*block));
623 last_index = ctx->m.dct_quantize(&ctx->m, block, i,
625 ac_bits += dnxhd_calc_ac_bits(ctx, block, last_index);
627 diff = block[0] - ctx->m.last_dc[n];
629 nbits = av_log2_16bit(-2 * diff);
631 nbits = av_log2_16bit(2 * diff);
633 assert(nbits < ctx->cid_table->bit_depth + 4);
634 dc_bits += ctx->cid_table->dc_bits[nbits] + nbits;
636 ctx->m.last_dc[n] = block[0];
638 if (avctx->mb_decision == FF_MB_DECISION_RD || !RC_VARIANCE) {
639 dnxhd_unquantize_c(ctx, block, i, qscale, last_index);
640 ctx->m.idsp.idct(block);
641 ssd += dnxhd_ssd_block(block, src_block);
644 ctx->mb_rc[qscale][mb].ssd = ssd;
645 ctx->mb_rc[qscale][mb].bits = ac_bits + dc_bits + 12 +
646 8 * ctx->vlc_bits[0];
651 static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg,
652 int jobnr, int threadnr)
654 DNXHDEncContext *ctx = avctx->priv_data;
655 int mb_y = jobnr, mb_x;
656 ctx = ctx->thread[threadnr];
657 init_put_bits(&ctx->m.pb, (uint8_t *)arg + 640 + ctx->slice_offs[jobnr],
658 ctx->slice_size[jobnr]);
662 ctx->m.last_dc[2] = 1 << (ctx->cid_table->bit_depth + 2);
663 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
664 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
665 int qscale = ctx->mb_qscale[mb];
668 put_bits(&ctx->m.pb, 12, qscale << 1);
670 dnxhd_get_blocks(ctx, mb_x, mb_y);
672 for (i = 0; i < 8; i++) {
673 int16_t *block = ctx->blocks[i];
674 int overflow, n = dnxhd_switch_matrix(ctx, i);
675 int last_index = ctx->m.dct_quantize(&ctx->m, block, i,
678 dnxhd_encode_block(ctx, block, last_index, n);
679 // STOP_TIMER("encode_block");
682 if (put_bits_count(&ctx->m.pb) & 31)
683 put_bits(&ctx->m.pb, 32 - (put_bits_count(&ctx->m.pb) & 31), 0);
684 flush_put_bits(&ctx->m.pb);
688 static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx)
692 for (mb_y = 0; mb_y < ctx->m.mb_height; mb_y++) {
694 ctx->slice_offs[mb_y] = offset;
695 ctx->slice_size[mb_y] = 0;
696 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
697 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
698 ctx->slice_size[mb_y] += ctx->mb_bits[mb];
700 ctx->slice_size[mb_y] = (ctx->slice_size[mb_y] + 31) & ~31;
701 ctx->slice_size[mb_y] >>= 3;
702 thread_size = ctx->slice_size[mb_y];
703 offset += thread_size;
707 static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg,
708 int jobnr, int threadnr)
710 DNXHDEncContext *ctx = avctx->priv_data;
711 int mb_y = jobnr, mb_x, x, y;
712 int partial_last_row = (mb_y == ctx->m.mb_height - 1) &&
713 ((avctx->height >> ctx->interlaced) & 0xF);
715 ctx = ctx->thread[threadnr];
716 if (ctx->cid_table->bit_depth == 8) {
717 uint8_t *pix = ctx->thread[0]->src[0] + ((mb_y << 4) * ctx->m.linesize);
718 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x, pix += 16) {
719 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
723 if (!partial_last_row && mb_x * 16 <= avctx->width - 16) {
724 sum = ctx->m.mpvencdsp.pix_sum(pix, ctx->m.linesize);
725 varc = ctx->m.mpvencdsp.pix_norm1(pix, ctx->m.linesize);
727 int bw = FFMIN(avctx->width - 16 * mb_x, 16);
728 int bh = FFMIN((avctx->height >> ctx->interlaced) - 16 * mb_y, 16);
730 for (y = 0; y < bh; y++) {
731 for (x = 0; x < bw; x++) {
732 uint8_t val = pix[x + y * ctx->m.linesize];
738 varc = (varc - (((unsigned) sum * sum) >> 8) + 128) >> 8;
740 ctx->mb_cmp[mb].value = varc;
741 ctx->mb_cmp[mb].mb = mb;
744 int const linesize = ctx->m.linesize >> 1;
745 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x) {
746 uint16_t *pix = (uint16_t *)ctx->thread[0]->src[0] +
747 ((mb_y << 4) * linesize) + (mb_x << 4);
748 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
753 // Macroblocks are 16x16 pixels, unlike DCT blocks which are 8x8.
754 for (i = 0; i < 16; ++i) {
755 for (j = 0; j < 16; ++j) {
756 // Turn 16-bit pixels into 10-bit ones.
757 int const sample = (unsigned) pix[j] >> 6;
759 sqsum += sample * sample;
760 // 2^10 * 2^10 * 16 * 16 = 2^28, which is less than INT_MAX
764 mean = sum >> 8; // 16*16 == 2^8
766 ctx->mb_cmp[mb].value = sqmean - mean * mean;
767 ctx->mb_cmp[mb].mb = mb;
773 static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx)
775 int lambda, up_step, down_step;
776 int last_lower = INT_MAX, last_higher = 0;
779 for (q = 1; q < avctx->qmax; q++) {
781 avctx->execute2(avctx, dnxhd_calc_bits_thread,
782 NULL, NULL, ctx->m.mb_height);
784 up_step = down_step = 2 << LAMBDA_FRAC_BITS;
785 lambda = ctx->lambda;
790 if (lambda == last_higher) {
792 end = 1; // need to set final qscales/bits
794 for (y = 0; y < ctx->m.mb_height; y++) {
795 for (x = 0; x < ctx->m.mb_width; x++) {
796 unsigned min = UINT_MAX;
798 int mb = y * ctx->m.mb_width + x;
799 for (q = 1; q < avctx->qmax; q++) {
800 unsigned score = ctx->mb_rc[q][mb].bits * lambda +
801 ((unsigned) ctx->mb_rc[q][mb].ssd << LAMBDA_FRAC_BITS);
807 bits += ctx->mb_rc[qscale][mb].bits;
808 ctx->mb_qscale[mb] = qscale;
809 ctx->mb_bits[mb] = ctx->mb_rc[qscale][mb].bits;
811 bits = (bits + 31) & ~31; // padding
812 if (bits > ctx->frame_bits)
815 // av_dlog(ctx->m.avctx,
816 // "lambda %d, up %u, down %u, bits %d, frame %d\n",
817 // lambda, last_higher, last_lower, bits, ctx->frame_bits);
819 if (bits > ctx->frame_bits)
820 return AVERROR(EINVAL);
823 if (bits < ctx->frame_bits) {
824 last_lower = FFMIN(lambda, last_lower);
825 if (last_higher != 0)
826 lambda = (lambda+last_higher)>>1;
829 down_step = FFMIN((int64_t)down_step*5, INT_MAX);
830 up_step = 1<<LAMBDA_FRAC_BITS;
831 lambda = FFMAX(1, lambda);
832 if (lambda == last_lower)
835 last_higher = FFMAX(lambda, last_higher);
836 if (last_lower != INT_MAX)
837 lambda = (lambda+last_lower)>>1;
838 else if ((int64_t)lambda + up_step > INT_MAX)
839 return AVERROR(EINVAL);
842 up_step = FFMIN((int64_t)up_step*5, INT_MAX);
843 down_step = 1<<LAMBDA_FRAC_BITS;
846 //av_dlog(ctx->m.avctx, "out lambda %d\n", lambda);
847 ctx->lambda = lambda;
851 static int dnxhd_find_qscale(DNXHDEncContext *ctx)
857 int last_lower = INT_MAX;
861 qscale = ctx->qscale;
864 ctx->qscale = qscale;
865 // XXX avoid recalculating bits
866 ctx->m.avctx->execute2(ctx->m.avctx, dnxhd_calc_bits_thread,
867 NULL, NULL, ctx->m.mb_height);
868 for (y = 0; y < ctx->m.mb_height; y++) {
869 for (x = 0; x < ctx->m.mb_width; x++)
870 bits += ctx->mb_rc[qscale][y*ctx->m.mb_width+x].bits;
871 bits = (bits+31)&~31; // padding
872 if (bits > ctx->frame_bits)
875 // av_dlog(ctx->m.avctx,
876 // "%d, qscale %d, bits %d, frame %d, higher %d, lower %d\n",
877 // ctx->m.avctx->frame_number, qscale, bits, ctx->frame_bits,
878 // last_higher, last_lower);
879 if (bits < ctx->frame_bits) {
882 if (last_higher == qscale - 1) {
883 qscale = last_higher;
886 last_lower = FFMIN(qscale, last_lower);
887 if (last_higher != 0)
888 qscale = (qscale + last_higher) >> 1;
890 qscale -= down_step++;
895 if (last_lower == qscale + 1)
897 last_higher = FFMAX(qscale, last_higher);
898 if (last_lower != INT_MAX)
899 qscale = (qscale + last_lower) >> 1;
903 if (qscale >= ctx->m.avctx->qmax)
904 return AVERROR(EINVAL);
907 //av_dlog(ctx->m.avctx, "out qscale %d\n", qscale);
908 ctx->qscale = qscale;
912 #define BUCKET_BITS 8
913 #define RADIX_PASSES 4
914 #define NBUCKETS (1 << BUCKET_BITS)
916 static inline int get_bucket(int value, int shift)
919 value &= NBUCKETS - 1;
920 return NBUCKETS - 1 - value;
923 static void radix_count(const RCCMPEntry *data, int size,
924 int buckets[RADIX_PASSES][NBUCKETS])
927 memset(buckets, 0, sizeof(buckets[0][0]) * RADIX_PASSES * NBUCKETS);
928 for (i = 0; i < size; i++) {
929 int v = data[i].value;
930 for (j = 0; j < RADIX_PASSES; j++) {
931 buckets[j][get_bucket(v, 0)]++;
936 for (j = 0; j < RADIX_PASSES; j++) {
938 for (i = NBUCKETS - 1; i >= 0; i--)
939 buckets[j][i] = offset -= buckets[j][i];
940 assert(!buckets[j][0]);
944 static void radix_sort_pass(RCCMPEntry *dst, const RCCMPEntry *data,
945 int size, int buckets[NBUCKETS], int pass)
947 int shift = pass * BUCKET_BITS;
949 for (i = 0; i < size; i++) {
950 int v = get_bucket(data[i].value, shift);
951 int pos = buckets[v]++;
956 static void radix_sort(RCCMPEntry *data, int size)
958 int buckets[RADIX_PASSES][NBUCKETS];
959 RCCMPEntry *tmp = av_malloc(sizeof(*tmp) * size);
960 radix_count(data, size, buckets);
961 radix_sort_pass(tmp, data, size, buckets[0], 0);
962 radix_sort_pass(data, tmp, size, buckets[1], 1);
963 if (buckets[2][NBUCKETS - 1] || buckets[3][NBUCKETS - 1]) {
964 radix_sort_pass(tmp, data, size, buckets[2], 2);
965 radix_sort_pass(data, tmp, size, buckets[3], 3);
970 static int dnxhd_encode_fast(AVCodecContext *avctx, DNXHDEncContext *ctx)
974 if ((ret = dnxhd_find_qscale(ctx)) < 0)
976 for (y = 0; y < ctx->m.mb_height; y++) {
977 for (x = 0; x < ctx->m.mb_width; x++) {
978 int mb = y * ctx->m.mb_width + x;
980 ctx->mb_qscale[mb] = ctx->qscale;
981 ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale][mb].bits;
982 max_bits += ctx->mb_rc[ctx->qscale][mb].bits;
984 delta_bits = ctx->mb_rc[ctx->qscale][mb].bits -
985 ctx->mb_rc[ctx->qscale + 1][mb].bits;
986 ctx->mb_cmp[mb].mb = mb;
987 ctx->mb_cmp[mb].value =
988 delta_bits ? ((ctx->mb_rc[ctx->qscale][mb].ssd -
989 ctx->mb_rc[ctx->qscale + 1][mb].ssd) * 100) /
991 : INT_MIN; // avoid increasing qscale
994 max_bits += 31; // worst padding
998 avctx->execute2(avctx, dnxhd_mb_var_thread,
999 NULL, NULL, ctx->m.mb_height);
1000 radix_sort(ctx->mb_cmp, ctx->m.mb_num);
1001 for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) {
1002 int mb = ctx->mb_cmp[x].mb;
1003 max_bits -= ctx->mb_rc[ctx->qscale][mb].bits -
1004 ctx->mb_rc[ctx->qscale + 1][mb].bits;
1005 ctx->mb_qscale[mb] = ctx->qscale + 1;
1006 ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale + 1][mb].bits;
1012 static void dnxhd_load_picture(DNXHDEncContext *ctx, const AVFrame *frame)
1016 for (i = 0; i < ctx->m.avctx->thread_count; i++) {
1017 ctx->thread[i]->m.linesize = frame->linesize[0] << ctx->interlaced;
1018 ctx->thread[i]->m.uvlinesize = frame->linesize[1] << ctx->interlaced;
1019 ctx->thread[i]->dct_y_offset = ctx->m.linesize *8;
1020 ctx->thread[i]->dct_uv_offset = ctx->m.uvlinesize*8;
1023 ctx->m.avctx->coded_frame->interlaced_frame = frame->interlaced_frame;
1024 ctx->cur_field = frame->interlaced_frame && !frame->top_field_first;
1027 static int dnxhd_encode_picture(AVCodecContext *avctx, AVPacket *pkt,
1028 const AVFrame *frame, int *got_packet)
1030 DNXHDEncContext *ctx = avctx->priv_data;
1031 int first_field = 1;
1035 if ((ret = ff_alloc_packet(pkt, ctx->cid_table->frame_size)) < 0) {
1036 av_log(avctx, AV_LOG_ERROR,
1037 "output buffer is too small to compress picture\n");
1042 dnxhd_load_picture(ctx, frame);
1045 for (i = 0; i < 3; i++) {
1046 ctx->src[i] = frame->data[i];
1047 if (ctx->interlaced && ctx->cur_field)
1048 ctx->src[i] += frame->linesize[i];
1051 dnxhd_write_header(avctx, buf);
1053 if (avctx->mb_decision == FF_MB_DECISION_RD)
1054 ret = dnxhd_encode_rdo(avctx, ctx);
1056 ret = dnxhd_encode_fast(avctx, ctx);
1058 av_log(avctx, AV_LOG_ERROR,
1059 "picture could not fit ratecontrol constraints, increase qmax\n");
1063 dnxhd_setup_threads_slices(ctx);
1066 for (i = 0; i < ctx->m.mb_height; i++) {
1067 AV_WB32(ctx->msip + i * 4, offset);
1068 offset += ctx->slice_size[i];
1069 assert(!(ctx->slice_size[i] & 3));
1072 avctx->execute2(avctx, dnxhd_encode_thread, buf, NULL, ctx->m.mb_height);
1074 assert(640 + offset + 4 <= ctx->cid_table->coding_unit_size);
1075 memset(buf + 640 + offset, 0,
1076 ctx->cid_table->coding_unit_size - 4 - offset - 640);
1078 AV_WB32(buf + ctx->cid_table->coding_unit_size - 4, 0x600DC0DE); // EOF
1080 if (ctx->interlaced && first_field) {
1082 ctx->cur_field ^= 1;
1083 buf += ctx->cid_table->coding_unit_size;
1084 goto encode_coding_unit;
1087 avctx->coded_frame->quality = ctx->qscale * FF_QP2LAMBDA;
1089 pkt->flags |= AV_PKT_FLAG_KEY;
1094 static av_cold int dnxhd_encode_end(AVCodecContext *avctx)
1096 DNXHDEncContext *ctx = avctx->priv_data;
1097 int max_level = 1 << (ctx->cid_table->bit_depth + 2);
1100 av_free(ctx->vlc_codes - max_level * 2);
1101 av_free(ctx->vlc_bits - max_level * 2);
1102 av_freep(&ctx->run_codes);
1103 av_freep(&ctx->run_bits);
1105 av_freep(&ctx->mb_bits);
1106 av_freep(&ctx->mb_qscale);
1107 av_freep(&ctx->mb_rc);
1108 av_freep(&ctx->mb_cmp);
1109 av_freep(&ctx->slice_size);
1110 av_freep(&ctx->slice_offs);
1112 av_freep(&ctx->qmatrix_c);
1113 av_freep(&ctx->qmatrix_l);
1114 av_freep(&ctx->qmatrix_c16);
1115 av_freep(&ctx->qmatrix_l16);
1117 for (i = 1; i < avctx->thread_count; i++)
1118 av_freep(&ctx->thread[i]);
1120 av_frame_free(&avctx->coded_frame);
1125 AVCodec ff_dnxhd_encoder = {
1127 .long_name = NULL_IF_CONFIG_SMALL("VC3/DNxHD"),
1128 .type = AVMEDIA_TYPE_VIDEO,
1129 .id = AV_CODEC_ID_DNXHD,
1130 .priv_data_size = sizeof(DNXHDEncContext),
1131 .init = dnxhd_encode_init,
1132 .encode2 = dnxhd_encode_picture,
1133 .close = dnxhd_encode_end,
1134 .capabilities = CODEC_CAP_SLICE_THREADS,
1135 .pix_fmts = (const enum AVPixelFormat[]) {
1137 AV_PIX_FMT_YUV422P10,
1140 .priv_class = &class,