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 },
48 { "ibias", "intra quant bias",
49 offsetof(DNXHDEncContext, intra_quant_bias), AV_OPT_TYPE_INT,
50 { .i64 = FF_DEFAULT_QUANT_BIAS }, INT_MIN, INT_MAX, VE },
54 static const AVClass class = {
61 static void dnxhd_8bit_get_pixels_8x4_sym(int16_t *restrict block,
62 const uint8_t *pixels,
66 for (i = 0; i < 4; i++) {
78 memcpy(block, block - 8, sizeof(*block) * 8);
79 memcpy(block + 8, block - 16, sizeof(*block) * 8);
80 memcpy(block + 16, block - 24, sizeof(*block) * 8);
81 memcpy(block + 24, block - 32, sizeof(*block) * 8);
84 static av_always_inline
85 void dnxhd_10bit_get_pixels_8x4_sym(int16_t *restrict block,
86 const uint8_t *pixels,
93 for (i = 0; i < 4; i++) {
94 memcpy(block + i * 8, pixels + i * line_size, 8 * sizeof(*block));
95 memcpy(block - (i + 1) * 8, pixels + i * line_size, 8 * sizeof(*block));
99 static int dnxhd_10bit_dct_quantize(MpegEncContext *ctx, int16_t *block,
100 int n, int qscale, int *overflow)
102 const uint8_t *scantable= ctx->intra_scantable.scantable;
103 const int *qmat = ctx->q_intra_matrix[qscale];
104 int last_non_zero = 0;
107 ctx->fdsp.fdct(block);
109 // Divide by 4 with rounding, to compensate scaling of DCT coefficients
110 block[0] = (block[0] + 2) >> 2;
112 for (i = 1; i < 64; ++i) {
113 int j = scantable[i];
114 int sign = FF_SIGNBIT(block[j]);
115 int level = (block[j] ^ sign) - sign;
116 level = level * qmat[j] >> DNX10BIT_QMAT_SHIFT;
117 block[j] = (level ^ sign) - sign;
122 return last_non_zero;
125 static av_cold int dnxhd_init_vlc(DNXHDEncContext *ctx)
127 int i, j, level, run;
128 int max_level = 1 << (ctx->cid_table->bit_depth + 2);
130 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->vlc_codes,
131 max_level * 4 * sizeof(*ctx->vlc_codes), fail);
132 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->vlc_bits,
133 max_level * 4 * sizeof(*ctx->vlc_bits), fail);
134 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_codes,
136 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_bits,
139 ctx->vlc_codes += max_level * 2;
140 ctx->vlc_bits += max_level * 2;
141 for (level = -max_level; level < max_level; level++) {
142 for (run = 0; run < 2; run++) {
143 int index = (level << 1) | run;
144 int sign, offset = 0, alevel = level;
146 MASK_ABS(sign, alevel);
148 offset = (alevel - 1) >> 6;
149 alevel -= offset << 6;
151 for (j = 0; j < 257; j++) {
152 if (ctx->cid_table->ac_level[j] == alevel &&
153 (!offset || (ctx->cid_table->ac_index_flag[j] && offset)) &&
154 (!run || (ctx->cid_table->ac_run_flag [j] && run))) {
155 assert(!ctx->vlc_codes[index]);
157 ctx->vlc_codes[index] =
158 (ctx->cid_table->ac_codes[j] << 1) | (sign & 1);
159 ctx->vlc_bits[index] = ctx->cid_table->ac_bits[j] + 1;
161 ctx->vlc_codes[index] = ctx->cid_table->ac_codes[j];
162 ctx->vlc_bits[index] = ctx->cid_table->ac_bits[j];
167 assert(!alevel || j < 257);
169 ctx->vlc_codes[index] =
170 (ctx->vlc_codes[index] << ctx->cid_table->index_bits) | offset;
171 ctx->vlc_bits[index] += ctx->cid_table->index_bits;
175 for (i = 0; i < 62; i++) {
176 int run = ctx->cid_table->run[i];
178 ctx->run_codes[run] = ctx->cid_table->run_codes[i];
179 ctx->run_bits[run] = ctx->cid_table->run_bits[i];
183 return AVERROR(ENOMEM);
186 static av_cold int dnxhd_init_qmat(DNXHDEncContext *ctx, int lbias, int cbias)
188 // init first elem to 1 to avoid div by 0 in convert_matrix
189 uint16_t weight_matrix[64] = { 1, }; // convert_matrix needs uint16_t*
191 const uint8_t *luma_weight_table = ctx->cid_table->luma_weight;
192 const uint8_t *chroma_weight_table = ctx->cid_table->chroma_weight;
194 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l,
195 (ctx->m.avctx->qmax + 1) * 64 * sizeof(int), fail);
196 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c,
197 (ctx->m.avctx->qmax + 1) * 64 * sizeof(int), fail);
198 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l16,
199 (ctx->m.avctx->qmax + 1) * 64 * 2 * sizeof(uint16_t),
201 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c16,
202 (ctx->m.avctx->qmax + 1) * 64 * 2 * sizeof(uint16_t),
205 if (ctx->cid_table->bit_depth == 8) {
206 for (i = 1; i < 64; i++) {
207 int j = ctx->m.idsp.idct_permutation[ff_zigzag_direct[i]];
208 weight_matrix[j] = ctx->cid_table->luma_weight[i];
210 ff_convert_matrix(&ctx->m, ctx->qmatrix_l, ctx->qmatrix_l16,
211 weight_matrix, ctx->intra_quant_bias, 1,
212 ctx->m.avctx->qmax, 1);
213 for (i = 1; i < 64; i++) {
214 int j = ctx->m.idsp.idct_permutation[ff_zigzag_direct[i]];
215 weight_matrix[j] = ctx->cid_table->chroma_weight[i];
217 ff_convert_matrix(&ctx->m, ctx->qmatrix_c, ctx->qmatrix_c16,
218 weight_matrix, ctx->intra_quant_bias, 1,
219 ctx->m.avctx->qmax, 1);
221 for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
222 for (i = 0; i < 64; i++) {
223 ctx->qmatrix_l[qscale][i] <<= 2;
224 ctx->qmatrix_c[qscale][i] <<= 2;
225 ctx->qmatrix_l16[qscale][0][i] <<= 2;
226 ctx->qmatrix_l16[qscale][1][i] <<= 2;
227 ctx->qmatrix_c16[qscale][0][i] <<= 2;
228 ctx->qmatrix_c16[qscale][1][i] <<= 2;
233 for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
234 for (i = 1; i < 64; i++) {
235 int j = ctx->m.idsp.idct_permutation[ff_zigzag_direct[i]];
237 /* The quantization formula from the VC-3 standard is:
238 * quantized = sign(block[i]) * floor(abs(block[i]/s) * p /
239 * (qscale * weight_table[i]))
240 * Where p is 32 for 8-bit samples and 8 for 10-bit ones.
241 * The s factor compensates scaling of DCT coefficients done by
242 * the DCT routines, and therefore is not present in standard.
243 * It's 8 for 8-bit samples and 4 for 10-bit ones.
244 * We want values of ctx->qtmatrix_l and ctx->qtmatrix_r to be:
245 * ((1 << DNX10BIT_QMAT_SHIFT) * (p / s)) /
246 * (qscale * weight_table[i])
247 * For 10-bit samples, p / s == 2 */
248 ctx->qmatrix_l[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) /
249 (qscale * luma_weight_table[i]);
250 ctx->qmatrix_c[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) /
251 (qscale * chroma_weight_table[i]);
258 return AVERROR(ENOMEM);
261 static av_cold int dnxhd_init_rc(DNXHDEncContext *ctx)
263 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_rc,
264 8160 * ctx->m.avctx->qmax * sizeof(RCEntry), fail);
265 if (ctx->m.avctx->mb_decision != FF_MB_DECISION_RD)
266 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_cmp,
267 ctx->m.mb_num * sizeof(RCCMPEntry), fail);
269 ctx->frame_bits = (ctx->cid_table->coding_unit_size -
270 640 - 4 - ctx->min_padding) * 8;
272 ctx->lambda = 2 << LAMBDA_FRAC_BITS; // qscale 2
275 return AVERROR(ENOMEM);
278 static av_cold int dnxhd_encode_init(AVCodecContext *avctx)
280 DNXHDEncContext *ctx = avctx->priv_data;
281 int i, index, bit_depth, ret;
283 switch (avctx->pix_fmt) {
284 case AV_PIX_FMT_YUV422P:
287 case AV_PIX_FMT_YUV422P10:
291 av_log(avctx, AV_LOG_ERROR,
292 "pixel format is incompatible with DNxHD\n");
293 return AVERROR(EINVAL);
296 ctx->cid = ff_dnxhd_find_cid(avctx, bit_depth);
298 av_log(avctx, AV_LOG_ERROR,
299 "video parameters incompatible with DNxHD\n");
300 return AVERROR(EINVAL);
302 av_log(avctx, AV_LOG_DEBUG, "cid %d\n", ctx->cid);
304 index = ff_dnxhd_get_cid_table(ctx->cid);
307 ctx->cid_table = &ff_dnxhd_cid_table[index];
309 ctx->m.avctx = avctx;
313 avctx->bits_per_raw_sample = ctx->cid_table->bit_depth;
315 ff_blockdsp_init(&ctx->bdsp, avctx);
316 ff_fdctdsp_init(&ctx->m.fdsp, avctx);
317 ff_mpv_idct_init(&ctx->m);
318 ff_mpegvideoencdsp_init(&ctx->m.mpvencdsp, avctx);
319 ff_pixblockdsp_init(&ctx->m.pdsp, avctx);
320 if (!ctx->m.dct_quantize)
321 ctx->m.dct_quantize = ff_dct_quantize_c;
323 if (ctx->cid_table->bit_depth == 10) {
324 ctx->m.dct_quantize = dnxhd_10bit_dct_quantize;
325 ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym;
326 ctx->block_width_l2 = 4;
328 ctx->get_pixels_8x4_sym = dnxhd_8bit_get_pixels_8x4_sym;
329 ctx->block_width_l2 = 3;
333 ff_dnxhdenc_init_x86(ctx);
335 ctx->m.mb_height = (avctx->height + 15) / 16;
336 ctx->m.mb_width = (avctx->width + 15) / 16;
338 if (avctx->flags & CODEC_FLAG_INTERLACED_DCT) {
340 ctx->m.mb_height /= 2;
343 ctx->m.mb_num = ctx->m.mb_height * ctx->m.mb_width;
345 #if FF_API_QUANT_BIAS
346 FF_DISABLE_DEPRECATION_WARNINGS
347 if (ctx->intra_quant_bias == FF_DEFAULT_QUANT_BIAS &&
348 avctx->intra_quant_bias != FF_DEFAULT_QUANT_BIAS)
349 ctx->intra_quant_bias = avctx->intra_quant_bias;
350 FF_ENABLE_DEPRECATION_WARNINGS
352 // XXX tune lbias/cbias
353 if ((ret = dnxhd_init_qmat(ctx, ctx->intra_quant_bias, 0)) < 0)
356 /* Avid Nitris hardware decoder requires a minimum amount of padding
357 * in the coding unit payload */
358 if (ctx->nitris_compat)
359 ctx->min_padding = 1600;
361 if ((ret = dnxhd_init_vlc(ctx)) < 0)
363 if ((ret = dnxhd_init_rc(ctx)) < 0)
366 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_size,
367 ctx->m.mb_height * sizeof(uint32_t), fail);
368 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_offs,
369 ctx->m.mb_height * sizeof(uint32_t), fail);
370 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_bits,
371 ctx->m.mb_num * sizeof(uint16_t), fail);
372 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_qscale,
373 ctx->m.mb_num * sizeof(uint8_t), fail);
375 avctx->coded_frame->key_frame = 1;
376 avctx->coded_frame->pict_type = AV_PICTURE_TYPE_I;
378 if (avctx->thread_count > MAX_THREADS) {
379 av_log(avctx, AV_LOG_ERROR, "too many threads\n");
380 return AVERROR(EINVAL);
383 ctx->thread[0] = ctx;
384 for (i = 1; i < avctx->thread_count; i++) {
385 ctx->thread[i] = av_malloc(sizeof(DNXHDEncContext));
386 memcpy(ctx->thread[i], ctx, sizeof(DNXHDEncContext));
390 fail: // for FF_ALLOCZ_OR_GOTO
391 return AVERROR(ENOMEM);
394 static int dnxhd_write_header(AVCodecContext *avctx, uint8_t *buf)
396 DNXHDEncContext *ctx = avctx->priv_data;
397 const uint8_t header_prefix[5] = { 0x00, 0x00, 0x02, 0x80, 0x01 };
401 memcpy(buf, header_prefix, 5);
402 buf[5] = ctx->interlaced ? ctx->cur_field + 2 : 0x01;
403 buf[6] = 0x80; // crc flag off
404 buf[7] = 0xa0; // reserved
405 AV_WB16(buf + 0x18, avctx->height >> ctx->interlaced); // ALPF
406 AV_WB16(buf + 0x1a, avctx->width); // SPL
407 AV_WB16(buf + 0x1d, avctx->height >> ctx->interlaced); // NAL
409 buf[0x21] = ctx->cid_table->bit_depth == 10 ? 0x58 : 0x38;
410 buf[0x22] = 0x88 + (ctx->interlaced << 2);
411 AV_WB32(buf + 0x28, ctx->cid); // CID
412 buf[0x2c] = ctx->interlaced ? 0 : 0x80;
414 buf[0x5f] = 0x01; // UDL
416 buf[0x167] = 0x02; // reserved
417 AV_WB16(buf + 0x16a, ctx->m.mb_height * 4 + 4); // MSIPS
418 buf[0x16d] = ctx->m.mb_height; // Ns
419 buf[0x16f] = 0x10; // reserved
421 ctx->msip = buf + 0x170;
425 static av_always_inline void dnxhd_encode_dc(DNXHDEncContext *ctx, int diff)
429 nbits = av_log2_16bit(-2 * diff);
432 nbits = av_log2_16bit(2 * diff);
434 put_bits(&ctx->m.pb, ctx->cid_table->dc_bits[nbits] + nbits,
435 (ctx->cid_table->dc_codes[nbits] << nbits) +
436 (diff & ((1 << nbits) - 1)));
439 static av_always_inline
440 void dnxhd_encode_block(DNXHDEncContext *ctx, int16_t *block,
441 int last_index, int n)
443 int last_non_zero = 0;
446 dnxhd_encode_dc(ctx, block[0] - ctx->m.last_dc[n]);
447 ctx->m.last_dc[n] = block[0];
449 for (i = 1; i <= last_index; i++) {
450 j = ctx->m.intra_scantable.permutated[i];
453 int run_level = i - last_non_zero - 1;
454 int rlevel = (slevel << 1) | !!run_level;
455 put_bits(&ctx->m.pb, ctx->vlc_bits[rlevel], ctx->vlc_codes[rlevel]);
457 put_bits(&ctx->m.pb, ctx->run_bits[run_level],
458 ctx->run_codes[run_level]);
462 put_bits(&ctx->m.pb, ctx->vlc_bits[0], ctx->vlc_codes[0]); // EOB
465 static av_always_inline
466 void dnxhd_unquantize_c(DNXHDEncContext *ctx, int16_t *block, int n,
467 int qscale, int last_index)
469 const uint8_t *weight_matrix;
473 weight_matrix = (n & 2) ? ctx->cid_table->chroma_weight
474 : ctx->cid_table->luma_weight;
476 for (i = 1; i <= last_index; i++) {
477 int j = ctx->m.intra_scantable.permutated[i];
481 level = (1 - 2 * level) * qscale * weight_matrix[i];
482 if (ctx->cid_table->bit_depth == 10) {
483 if (weight_matrix[i] != 8)
487 if (weight_matrix[i] != 32)
493 level = (2 * level + 1) * qscale * weight_matrix[i];
494 if (ctx->cid_table->bit_depth == 10) {
495 if (weight_matrix[i] != 8)
499 if (weight_matrix[i] != 32)
509 static av_always_inline int dnxhd_ssd_block(int16_t *qblock, int16_t *block)
513 for (i = 0; i < 64; i++)
514 score += (block[i] - qblock[i]) * (block[i] - qblock[i]);
518 static av_always_inline
519 int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, int16_t *block, int last_index)
521 int last_non_zero = 0;
524 for (i = 1; i <= last_index; i++) {
525 j = ctx->m.intra_scantable.permutated[i];
528 int run_level = i - last_non_zero - 1;
529 bits += ctx->vlc_bits[(level << 1) |
530 !!run_level] + ctx->run_bits[run_level];
537 static av_always_inline
538 void dnxhd_get_blocks(DNXHDEncContext *ctx, int mb_x, int mb_y)
540 const int bs = ctx->block_width_l2;
541 const int bw = 1 << bs;
542 const uint8_t *ptr_y = ctx->thread[0]->src[0] +
543 ((mb_y << 4) * ctx->m.linesize) + (mb_x << bs + 1);
544 const uint8_t *ptr_u = ctx->thread[0]->src[1] +
545 ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs);
546 const uint8_t *ptr_v = ctx->thread[0]->src[2] +
547 ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs);
548 PixblockDSPContext *pdsp = &ctx->m.pdsp;
550 pdsp->get_pixels(ctx->blocks[0], ptr_y, ctx->m.linesize);
551 pdsp->get_pixels(ctx->blocks[1], ptr_y + bw, ctx->m.linesize);
552 pdsp->get_pixels(ctx->blocks[2], ptr_u, ctx->m.uvlinesize);
553 pdsp->get_pixels(ctx->blocks[3], ptr_v, ctx->m.uvlinesize);
555 if (mb_y + 1 == ctx->m.mb_height && ctx->m.avctx->height == 1080) {
556 if (ctx->interlaced) {
557 ctx->get_pixels_8x4_sym(ctx->blocks[4],
558 ptr_y + ctx->dct_y_offset,
560 ctx->get_pixels_8x4_sym(ctx->blocks[5],
561 ptr_y + ctx->dct_y_offset + bw,
563 ctx->get_pixels_8x4_sym(ctx->blocks[6],
564 ptr_u + ctx->dct_uv_offset,
566 ctx->get_pixels_8x4_sym(ctx->blocks[7],
567 ptr_v + ctx->dct_uv_offset,
570 ctx->bdsp.clear_block(ctx->blocks[4]);
571 ctx->bdsp.clear_block(ctx->blocks[5]);
572 ctx->bdsp.clear_block(ctx->blocks[6]);
573 ctx->bdsp.clear_block(ctx->blocks[7]);
576 pdsp->get_pixels(ctx->blocks[4],
577 ptr_y + ctx->dct_y_offset, ctx->m.linesize);
578 pdsp->get_pixels(ctx->blocks[5],
579 ptr_y + ctx->dct_y_offset + bw, ctx->m.linesize);
580 pdsp->get_pixels(ctx->blocks[6],
581 ptr_u + ctx->dct_uv_offset, ctx->m.uvlinesize);
582 pdsp->get_pixels(ctx->blocks[7],
583 ptr_v + ctx->dct_uv_offset, ctx->m.uvlinesize);
587 static av_always_inline
588 int dnxhd_switch_matrix(DNXHDEncContext *ctx, int i)
591 ctx->m.q_intra_matrix16 = ctx->qmatrix_c16;
592 ctx->m.q_intra_matrix = ctx->qmatrix_c;
595 ctx->m.q_intra_matrix16 = ctx->qmatrix_l16;
596 ctx->m.q_intra_matrix = ctx->qmatrix_l;
601 static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg,
602 int jobnr, int threadnr)
604 DNXHDEncContext *ctx = avctx->priv_data;
605 int mb_y = jobnr, mb_x;
606 int qscale = ctx->qscale;
607 LOCAL_ALIGNED_16(int16_t, block, [64]);
608 ctx = ctx->thread[threadnr];
612 ctx->m.last_dc[2] = 1 << (ctx->cid_table->bit_depth + 2);
614 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
615 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
621 dnxhd_get_blocks(ctx, mb_x, mb_y);
623 for (i = 0; i < 8; i++) {
624 int16_t *src_block = ctx->blocks[i];
625 int overflow, nbits, diff, last_index;
626 int n = dnxhd_switch_matrix(ctx, i);
628 memcpy(block, src_block, 64 * sizeof(*block));
629 last_index = ctx->m.dct_quantize(&ctx->m, block, i,
631 ac_bits += dnxhd_calc_ac_bits(ctx, block, last_index);
633 diff = block[0] - ctx->m.last_dc[n];
635 nbits = av_log2_16bit(-2 * diff);
637 nbits = av_log2_16bit(2 * diff);
639 assert(nbits < ctx->cid_table->bit_depth + 4);
640 dc_bits += ctx->cid_table->dc_bits[nbits] + nbits;
642 ctx->m.last_dc[n] = block[0];
644 if (avctx->mb_decision == FF_MB_DECISION_RD || !RC_VARIANCE) {
645 dnxhd_unquantize_c(ctx, block, i, qscale, last_index);
646 ctx->m.idsp.idct(block);
647 ssd += dnxhd_ssd_block(block, src_block);
650 ctx->mb_rc[qscale][mb].ssd = ssd;
651 ctx->mb_rc[qscale][mb].bits = ac_bits + dc_bits + 12 +
652 8 * ctx->vlc_bits[0];
657 static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg,
658 int jobnr, int threadnr)
660 DNXHDEncContext *ctx = avctx->priv_data;
661 int mb_y = jobnr, mb_x;
662 ctx = ctx->thread[threadnr];
663 init_put_bits(&ctx->m.pb, (uint8_t *)arg + 640 + ctx->slice_offs[jobnr],
664 ctx->slice_size[jobnr]);
668 ctx->m.last_dc[2] = 1 << (ctx->cid_table->bit_depth + 2);
669 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
670 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
671 int qscale = ctx->mb_qscale[mb];
674 put_bits(&ctx->m.pb, 12, qscale << 1);
676 dnxhd_get_blocks(ctx, mb_x, mb_y);
678 for (i = 0; i < 8; i++) {
679 int16_t *block = ctx->blocks[i];
680 int overflow, n = dnxhd_switch_matrix(ctx, i);
681 int last_index = ctx->m.dct_quantize(&ctx->m, block, i,
684 dnxhd_encode_block(ctx, block, last_index, n);
685 // STOP_TIMER("encode_block");
688 if (put_bits_count(&ctx->m.pb) & 31)
689 put_bits(&ctx->m.pb, 32 - (put_bits_count(&ctx->m.pb) & 31), 0);
690 flush_put_bits(&ctx->m.pb);
694 static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx)
698 for (mb_y = 0; mb_y < ctx->m.mb_height; mb_y++) {
700 ctx->slice_offs[mb_y] = offset;
701 ctx->slice_size[mb_y] = 0;
702 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
703 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
704 ctx->slice_size[mb_y] += ctx->mb_bits[mb];
706 ctx->slice_size[mb_y] = (ctx->slice_size[mb_y] + 31) & ~31;
707 ctx->slice_size[mb_y] >>= 3;
708 thread_size = ctx->slice_size[mb_y];
709 offset += thread_size;
713 static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg,
714 int jobnr, int threadnr)
716 DNXHDEncContext *ctx = avctx->priv_data;
717 int mb_y = jobnr, mb_x, x, y;
718 int partial_last_row = (mb_y == ctx->m.mb_height - 1) &&
719 ((avctx->height >> ctx->interlaced) & 0xF);
721 ctx = ctx->thread[threadnr];
722 if (ctx->cid_table->bit_depth == 8) {
723 uint8_t *pix = ctx->thread[0]->src[0] + ((mb_y << 4) * ctx->m.linesize);
724 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x, pix += 16) {
725 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
729 if (!partial_last_row && mb_x * 16 <= avctx->width - 16) {
730 sum = ctx->m.mpvencdsp.pix_sum(pix, ctx->m.linesize);
731 varc = ctx->m.mpvencdsp.pix_norm1(pix, ctx->m.linesize);
733 int bw = FFMIN(avctx->width - 16 * mb_x, 16);
734 int bh = FFMIN((avctx->height >> ctx->interlaced) - 16 * mb_y, 16);
736 for (y = 0; y < bh; y++) {
737 for (x = 0; x < bw; x++) {
738 uint8_t val = pix[x + y * ctx->m.linesize];
744 varc = (varc - (((unsigned) sum * sum) >> 8) + 128) >> 8;
746 ctx->mb_cmp[mb].value = varc;
747 ctx->mb_cmp[mb].mb = mb;
750 int const linesize = ctx->m.linesize >> 1;
751 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x) {
752 uint16_t *pix = (uint16_t *)ctx->thread[0]->src[0] +
753 ((mb_y << 4) * linesize) + (mb_x << 4);
754 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
759 // Macroblocks are 16x16 pixels, unlike DCT blocks which are 8x8.
760 for (i = 0; i < 16; ++i) {
761 for (j = 0; j < 16; ++j) {
762 // Turn 16-bit pixels into 10-bit ones.
763 int const sample = (unsigned) pix[j] >> 6;
765 sqsum += sample * sample;
766 // 2^10 * 2^10 * 16 * 16 = 2^28, which is less than INT_MAX
770 mean = sum >> 8; // 16*16 == 2^8
772 ctx->mb_cmp[mb].value = sqmean - mean * mean;
773 ctx->mb_cmp[mb].mb = mb;
779 static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx)
781 int lambda, up_step, down_step;
782 int last_lower = INT_MAX, last_higher = 0;
785 for (q = 1; q < avctx->qmax; q++) {
787 avctx->execute2(avctx, dnxhd_calc_bits_thread,
788 NULL, NULL, ctx->m.mb_height);
790 up_step = down_step = 2 << LAMBDA_FRAC_BITS;
791 lambda = ctx->lambda;
796 if (lambda == last_higher) {
798 end = 1; // need to set final qscales/bits
800 for (y = 0; y < ctx->m.mb_height; y++) {
801 for (x = 0; x < ctx->m.mb_width; x++) {
802 unsigned min = UINT_MAX;
804 int mb = y * ctx->m.mb_width + x;
805 for (q = 1; q < avctx->qmax; q++) {
806 unsigned score = ctx->mb_rc[q][mb].bits * lambda +
807 ((unsigned) ctx->mb_rc[q][mb].ssd << LAMBDA_FRAC_BITS);
813 bits += ctx->mb_rc[qscale][mb].bits;
814 ctx->mb_qscale[mb] = qscale;
815 ctx->mb_bits[mb] = ctx->mb_rc[qscale][mb].bits;
817 bits = (bits + 31) & ~31; // padding
818 if (bits > ctx->frame_bits)
821 // ff_dlog(ctx->m.avctx,
822 // "lambda %d, up %u, down %u, bits %d, frame %d\n",
823 // lambda, last_higher, last_lower, bits, ctx->frame_bits);
825 if (bits > ctx->frame_bits)
826 return AVERROR(EINVAL);
829 if (bits < ctx->frame_bits) {
830 last_lower = FFMIN(lambda, last_lower);
831 if (last_higher != 0)
832 lambda = (lambda+last_higher)>>1;
835 down_step = FFMIN((int64_t)down_step*5, INT_MAX);
836 up_step = 1<<LAMBDA_FRAC_BITS;
837 lambda = FFMAX(1, lambda);
838 if (lambda == last_lower)
841 last_higher = FFMAX(lambda, last_higher);
842 if (last_lower != INT_MAX)
843 lambda = (lambda+last_lower)>>1;
844 else if ((int64_t)lambda + up_step > INT_MAX)
845 return AVERROR(EINVAL);
848 up_step = FFMIN((int64_t)up_step*5, INT_MAX);
849 down_step = 1<<LAMBDA_FRAC_BITS;
852 //ff_dlog(ctx->m.avctx, "out lambda %d\n", lambda);
853 ctx->lambda = lambda;
857 static int dnxhd_find_qscale(DNXHDEncContext *ctx)
863 int last_lower = INT_MAX;
867 qscale = ctx->qscale;
870 ctx->qscale = qscale;
871 // XXX avoid recalculating bits
872 ctx->m.avctx->execute2(ctx->m.avctx, dnxhd_calc_bits_thread,
873 NULL, NULL, ctx->m.mb_height);
874 for (y = 0; y < ctx->m.mb_height; y++) {
875 for (x = 0; x < ctx->m.mb_width; x++)
876 bits += ctx->mb_rc[qscale][y*ctx->m.mb_width+x].bits;
877 bits = (bits+31)&~31; // padding
878 if (bits > ctx->frame_bits)
881 // ff_dlog(ctx->m.avctx,
882 // "%d, qscale %d, bits %d, frame %d, higher %d, lower %d\n",
883 // ctx->m.avctx->frame_number, qscale, bits, ctx->frame_bits,
884 // last_higher, last_lower);
885 if (bits < ctx->frame_bits) {
888 if (last_higher == qscale - 1) {
889 qscale = last_higher;
892 last_lower = FFMIN(qscale, last_lower);
893 if (last_higher != 0)
894 qscale = (qscale + last_higher) >> 1;
896 qscale -= down_step++;
901 if (last_lower == qscale + 1)
903 last_higher = FFMAX(qscale, last_higher);
904 if (last_lower != INT_MAX)
905 qscale = (qscale + last_lower) >> 1;
909 if (qscale >= ctx->m.avctx->qmax)
910 return AVERROR(EINVAL);
913 //ff_dlog(ctx->m.avctx, "out qscale %d\n", qscale);
914 ctx->qscale = qscale;
918 #define BUCKET_BITS 8
919 #define RADIX_PASSES 4
920 #define NBUCKETS (1 << BUCKET_BITS)
922 static inline int get_bucket(int value, int shift)
925 value &= NBUCKETS - 1;
926 return NBUCKETS - 1 - value;
929 static void radix_count(const RCCMPEntry *data, int size,
930 int buckets[RADIX_PASSES][NBUCKETS])
933 memset(buckets, 0, sizeof(buckets[0][0]) * RADIX_PASSES * NBUCKETS);
934 for (i = 0; i < size; i++) {
935 int v = data[i].value;
936 for (j = 0; j < RADIX_PASSES; j++) {
937 buckets[j][get_bucket(v, 0)]++;
942 for (j = 0; j < RADIX_PASSES; j++) {
944 for (i = NBUCKETS - 1; i >= 0; i--)
945 buckets[j][i] = offset -= buckets[j][i];
946 assert(!buckets[j][0]);
950 static void radix_sort_pass(RCCMPEntry *dst, const RCCMPEntry *data,
951 int size, int buckets[NBUCKETS], int pass)
953 int shift = pass * BUCKET_BITS;
955 for (i = 0; i < size; i++) {
956 int v = get_bucket(data[i].value, shift);
957 int pos = buckets[v]++;
962 static void radix_sort(RCCMPEntry *data, int size)
964 int buckets[RADIX_PASSES][NBUCKETS];
965 RCCMPEntry *tmp = av_malloc(sizeof(*tmp) * size);
966 radix_count(data, size, buckets);
967 radix_sort_pass(tmp, data, size, buckets[0], 0);
968 radix_sort_pass(data, tmp, size, buckets[1], 1);
969 if (buckets[2][NBUCKETS - 1] || buckets[3][NBUCKETS - 1]) {
970 radix_sort_pass(tmp, data, size, buckets[2], 2);
971 radix_sort_pass(data, tmp, size, buckets[3], 3);
976 static int dnxhd_encode_fast(AVCodecContext *avctx, DNXHDEncContext *ctx)
980 if ((ret = dnxhd_find_qscale(ctx)) < 0)
982 for (y = 0; y < ctx->m.mb_height; y++) {
983 for (x = 0; x < ctx->m.mb_width; x++) {
984 int mb = y * ctx->m.mb_width + x;
986 ctx->mb_qscale[mb] = ctx->qscale;
987 ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale][mb].bits;
988 max_bits += ctx->mb_rc[ctx->qscale][mb].bits;
990 delta_bits = ctx->mb_rc[ctx->qscale][mb].bits -
991 ctx->mb_rc[ctx->qscale + 1][mb].bits;
992 ctx->mb_cmp[mb].mb = mb;
993 ctx->mb_cmp[mb].value =
994 delta_bits ? ((ctx->mb_rc[ctx->qscale][mb].ssd -
995 ctx->mb_rc[ctx->qscale + 1][mb].ssd) * 100) /
997 : INT_MIN; // avoid increasing qscale
1000 max_bits += 31; // worst padding
1004 avctx->execute2(avctx, dnxhd_mb_var_thread,
1005 NULL, NULL, ctx->m.mb_height);
1006 radix_sort(ctx->mb_cmp, ctx->m.mb_num);
1007 for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) {
1008 int mb = ctx->mb_cmp[x].mb;
1009 max_bits -= ctx->mb_rc[ctx->qscale][mb].bits -
1010 ctx->mb_rc[ctx->qscale + 1][mb].bits;
1011 ctx->mb_qscale[mb] = ctx->qscale + 1;
1012 ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale + 1][mb].bits;
1018 static void dnxhd_load_picture(DNXHDEncContext *ctx, const AVFrame *frame)
1022 for (i = 0; i < ctx->m.avctx->thread_count; i++) {
1023 ctx->thread[i]->m.linesize = frame->linesize[0] << ctx->interlaced;
1024 ctx->thread[i]->m.uvlinesize = frame->linesize[1] << ctx->interlaced;
1025 ctx->thread[i]->dct_y_offset = ctx->m.linesize *8;
1026 ctx->thread[i]->dct_uv_offset = ctx->m.uvlinesize*8;
1029 ctx->m.avctx->coded_frame->interlaced_frame = frame->interlaced_frame;
1030 ctx->cur_field = frame->interlaced_frame && !frame->top_field_first;
1033 static int dnxhd_encode_picture(AVCodecContext *avctx, AVPacket *pkt,
1034 const AVFrame *frame, int *got_packet)
1036 DNXHDEncContext *ctx = avctx->priv_data;
1037 int first_field = 1;
1041 if ((ret = ff_alloc_packet(pkt, ctx->cid_table->frame_size)) < 0) {
1042 av_log(avctx, AV_LOG_ERROR,
1043 "output buffer is too small to compress picture\n");
1048 dnxhd_load_picture(ctx, frame);
1051 for (i = 0; i < 3; i++) {
1052 ctx->src[i] = frame->data[i];
1053 if (ctx->interlaced && ctx->cur_field)
1054 ctx->src[i] += frame->linesize[i];
1057 dnxhd_write_header(avctx, buf);
1059 if (avctx->mb_decision == FF_MB_DECISION_RD)
1060 ret = dnxhd_encode_rdo(avctx, ctx);
1062 ret = dnxhd_encode_fast(avctx, ctx);
1064 av_log(avctx, AV_LOG_ERROR,
1065 "picture could not fit ratecontrol constraints, increase qmax\n");
1069 dnxhd_setup_threads_slices(ctx);
1072 for (i = 0; i < ctx->m.mb_height; i++) {
1073 AV_WB32(ctx->msip + i * 4, offset);
1074 offset += ctx->slice_size[i];
1075 assert(!(ctx->slice_size[i] & 3));
1078 avctx->execute2(avctx, dnxhd_encode_thread, buf, NULL, ctx->m.mb_height);
1080 assert(640 + offset + 4 <= ctx->cid_table->coding_unit_size);
1081 memset(buf + 640 + offset, 0,
1082 ctx->cid_table->coding_unit_size - 4 - offset - 640);
1084 AV_WB32(buf + ctx->cid_table->coding_unit_size - 4, 0x600DC0DE); // EOF
1086 if (ctx->interlaced && first_field) {
1088 ctx->cur_field ^= 1;
1089 buf += ctx->cid_table->coding_unit_size;
1090 goto encode_coding_unit;
1093 avctx->coded_frame->quality = ctx->qscale * FF_QP2LAMBDA;
1095 sd = av_packet_new_side_data(pkt, AV_PKT_DATA_QUALITY_FACTOR, sizeof(int));
1097 return AVERROR(ENOMEM);
1098 *(int *)sd = ctx->qscale * FF_QP2LAMBDA;
1100 pkt->flags |= AV_PKT_FLAG_KEY;
1105 static av_cold int dnxhd_encode_end(AVCodecContext *avctx)
1107 DNXHDEncContext *ctx = avctx->priv_data;
1108 int max_level = 1 << (ctx->cid_table->bit_depth + 2);
1111 av_free(ctx->vlc_codes - max_level * 2);
1112 av_free(ctx->vlc_bits - max_level * 2);
1113 av_freep(&ctx->run_codes);
1114 av_freep(&ctx->run_bits);
1116 av_freep(&ctx->mb_bits);
1117 av_freep(&ctx->mb_qscale);
1118 av_freep(&ctx->mb_rc);
1119 av_freep(&ctx->mb_cmp);
1120 av_freep(&ctx->slice_size);
1121 av_freep(&ctx->slice_offs);
1123 av_freep(&ctx->qmatrix_c);
1124 av_freep(&ctx->qmatrix_l);
1125 av_freep(&ctx->qmatrix_c16);
1126 av_freep(&ctx->qmatrix_l16);
1128 for (i = 1; i < avctx->thread_count; i++)
1129 av_freep(&ctx->thread[i]);
1134 AVCodec ff_dnxhd_encoder = {
1136 .long_name = NULL_IF_CONFIG_SMALL("VC3/DNxHD"),
1137 .type = AVMEDIA_TYPE_VIDEO,
1138 .id = AV_CODEC_ID_DNXHD,
1139 .priv_data_size = sizeof(DNXHDEncContext),
1140 .init = dnxhd_encode_init,
1141 .encode2 = dnxhd_encode_picture,
1142 .close = dnxhd_encode_end,
1143 .capabilities = CODEC_CAP_SLICE_THREADS,
1144 .pix_fmts = (const enum AVPixelFormat[]) {
1146 AV_PIX_FMT_YUV422P10,
1149 .priv_class = &class,