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 10-bit 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, use yuv422p or yuv422p10.\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, available CIDs:\n");
300 ff_dnxhd_list_cid(avctx);
301 return AVERROR(EINVAL);
303 av_log(avctx, AV_LOG_DEBUG, "cid %d\n", ctx->cid);
305 index = ff_dnxhd_get_cid_table(ctx->cid);
308 ctx->cid_table = &ff_dnxhd_cid_table[index];
310 ctx->m.avctx = avctx;
314 avctx->bits_per_raw_sample = ctx->cid_table->bit_depth;
316 ff_blockdsp_init(&ctx->bdsp, avctx);
317 ff_fdctdsp_init(&ctx->m.fdsp, avctx);
318 ff_mpv_idct_init(&ctx->m);
319 ff_mpegvideoencdsp_init(&ctx->m.mpvencdsp, avctx);
320 ff_pixblockdsp_init(&ctx->m.pdsp, avctx);
321 if (!ctx->m.dct_quantize)
322 ctx->m.dct_quantize = ff_dct_quantize_c;
324 if (ctx->cid_table->bit_depth == 10) {
325 ctx->m.dct_quantize = dnxhd_10bit_dct_quantize;
326 ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym;
327 ctx->block_width_l2 = 4;
329 ctx->get_pixels_8x4_sym = dnxhd_8bit_get_pixels_8x4_sym;
330 ctx->block_width_l2 = 3;
334 ff_dnxhdenc_init_x86(ctx);
336 ctx->m.mb_height = (avctx->height + 15) / 16;
337 ctx->m.mb_width = (avctx->width + 15) / 16;
339 if (avctx->flags & AV_CODEC_FLAG_INTERLACED_DCT) {
341 ctx->m.mb_height /= 2;
344 ctx->m.mb_num = ctx->m.mb_height * ctx->m.mb_width;
346 #if FF_API_QUANT_BIAS
347 FF_DISABLE_DEPRECATION_WARNINGS
348 if (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 #if FF_API_CODED_FRAME
376 FF_DISABLE_DEPRECATION_WARNINGS
377 avctx->coded_frame->key_frame = 1;
378 avctx->coded_frame->pict_type = AV_PICTURE_TYPE_I;
379 FF_ENABLE_DEPRECATION_WARNINGS
382 if (avctx->thread_count > MAX_THREADS) {
383 av_log(avctx, AV_LOG_ERROR, "too many threads\n");
384 return AVERROR(EINVAL);
387 ctx->thread[0] = ctx;
388 for (i = 1; i < avctx->thread_count; i++) {
389 ctx->thread[i] = av_malloc(sizeof(DNXHDEncContext));
390 memcpy(ctx->thread[i], ctx, sizeof(DNXHDEncContext));
394 fail: // for FF_ALLOCZ_OR_GOTO
395 return AVERROR(ENOMEM);
398 static int dnxhd_write_header(AVCodecContext *avctx, uint8_t *buf)
400 DNXHDEncContext *ctx = avctx->priv_data;
401 static const uint8_t header_prefix[5] = { 0x00, 0x00, 0x02, 0x80, 0x01 };
405 memcpy(buf, header_prefix, 5);
406 buf[5] = ctx->interlaced ? ctx->cur_field + 2 : 0x01;
407 buf[6] = 0x80; // crc flag off
408 buf[7] = 0xa0; // reserved
409 AV_WB16(buf + 0x18, avctx->height >> ctx->interlaced); // ALPF
410 AV_WB16(buf + 0x1a, avctx->width); // SPL
411 AV_WB16(buf + 0x1d, avctx->height >> ctx->interlaced); // NAL
413 buf[0x21] = ctx->cid_table->bit_depth == 10 ? 0x58 : 0x38;
414 buf[0x22] = 0x88 + (ctx->interlaced << 2);
415 AV_WB32(buf + 0x28, ctx->cid); // CID
416 buf[0x2c] = ctx->interlaced ? 0 : 0x80;
418 buf[0x5f] = 0x01; // UDL
420 buf[0x167] = 0x02; // reserved
421 AV_WB16(buf + 0x16a, ctx->m.mb_height * 4 + 4); // MSIPS
422 buf[0x16d] = ctx->m.mb_height; // Ns
423 buf[0x16f] = 0x10; // reserved
425 ctx->msip = buf + 0x170;
429 static av_always_inline void dnxhd_encode_dc(DNXHDEncContext *ctx, int diff)
433 nbits = av_log2_16bit(-2 * diff);
436 nbits = av_log2_16bit(2 * diff);
438 put_bits(&ctx->m.pb, ctx->cid_table->dc_bits[nbits] + nbits,
439 (ctx->cid_table->dc_codes[nbits] << nbits) +
440 (diff & ((1 << nbits) - 1)));
443 static av_always_inline
444 void dnxhd_encode_block(DNXHDEncContext *ctx, int16_t *block,
445 int last_index, int n)
447 int last_non_zero = 0;
450 dnxhd_encode_dc(ctx, block[0] - ctx->m.last_dc[n]);
451 ctx->m.last_dc[n] = block[0];
453 for (i = 1; i <= last_index; i++) {
454 j = ctx->m.intra_scantable.permutated[i];
457 int run_level = i - last_non_zero - 1;
458 int rlevel = (slevel << 1) | !!run_level;
459 put_bits(&ctx->m.pb, ctx->vlc_bits[rlevel], ctx->vlc_codes[rlevel]);
461 put_bits(&ctx->m.pb, ctx->run_bits[run_level],
462 ctx->run_codes[run_level]);
466 put_bits(&ctx->m.pb, ctx->vlc_bits[0], ctx->vlc_codes[0]); // EOB
469 static av_always_inline
470 void dnxhd_unquantize_c(DNXHDEncContext *ctx, int16_t *block, int n,
471 int qscale, int last_index)
473 const uint8_t *weight_matrix;
477 weight_matrix = (n & 2) ? ctx->cid_table->chroma_weight
478 : ctx->cid_table->luma_weight;
480 for (i = 1; i <= last_index; i++) {
481 int j = ctx->m.intra_scantable.permutated[i];
485 level = (1 - 2 * level) * qscale * weight_matrix[i];
486 if (ctx->cid_table->bit_depth == 10) {
487 if (weight_matrix[i] != 8)
491 if (weight_matrix[i] != 32)
497 level = (2 * level + 1) * qscale * weight_matrix[i];
498 if (ctx->cid_table->bit_depth == 10) {
499 if (weight_matrix[i] != 8)
503 if (weight_matrix[i] != 32)
513 static av_always_inline int dnxhd_ssd_block(int16_t *qblock, int16_t *block)
517 for (i = 0; i < 64; i++)
518 score += (block[i] - qblock[i]) * (block[i] - qblock[i]);
522 static av_always_inline
523 int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, int16_t *block, int last_index)
525 int last_non_zero = 0;
528 for (i = 1; i <= last_index; i++) {
529 j = ctx->m.intra_scantable.permutated[i];
532 int run_level = i - last_non_zero - 1;
533 bits += ctx->vlc_bits[(level << 1) |
534 !!run_level] + ctx->run_bits[run_level];
541 static av_always_inline
542 void dnxhd_get_blocks(DNXHDEncContext *ctx, int mb_x, int mb_y)
544 const int bs = ctx->block_width_l2;
545 const int bw = 1 << bs;
546 const uint8_t *ptr_y = ctx->thread[0]->src[0] +
547 ((mb_y << 4) * ctx->m.linesize) + (mb_x << bs + 1);
548 const uint8_t *ptr_u = ctx->thread[0]->src[1] +
549 ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs);
550 const uint8_t *ptr_v = ctx->thread[0]->src[2] +
551 ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs);
552 PixblockDSPContext *pdsp = &ctx->m.pdsp;
554 pdsp->get_pixels(ctx->blocks[0], ptr_y, ctx->m.linesize);
555 pdsp->get_pixels(ctx->blocks[1], ptr_y + bw, ctx->m.linesize);
556 pdsp->get_pixels(ctx->blocks[2], ptr_u, ctx->m.uvlinesize);
557 pdsp->get_pixels(ctx->blocks[3], ptr_v, ctx->m.uvlinesize);
559 if (mb_y + 1 == ctx->m.mb_height && ctx->m.avctx->height == 1080) {
560 if (ctx->interlaced) {
561 ctx->get_pixels_8x4_sym(ctx->blocks[4],
562 ptr_y + ctx->dct_y_offset,
564 ctx->get_pixels_8x4_sym(ctx->blocks[5],
565 ptr_y + ctx->dct_y_offset + bw,
567 ctx->get_pixels_8x4_sym(ctx->blocks[6],
568 ptr_u + ctx->dct_uv_offset,
570 ctx->get_pixels_8x4_sym(ctx->blocks[7],
571 ptr_v + ctx->dct_uv_offset,
574 ctx->bdsp.clear_block(ctx->blocks[4]);
575 ctx->bdsp.clear_block(ctx->blocks[5]);
576 ctx->bdsp.clear_block(ctx->blocks[6]);
577 ctx->bdsp.clear_block(ctx->blocks[7]);
580 pdsp->get_pixels(ctx->blocks[4],
581 ptr_y + ctx->dct_y_offset, ctx->m.linesize);
582 pdsp->get_pixels(ctx->blocks[5],
583 ptr_y + ctx->dct_y_offset + bw, ctx->m.linesize);
584 pdsp->get_pixels(ctx->blocks[6],
585 ptr_u + ctx->dct_uv_offset, ctx->m.uvlinesize);
586 pdsp->get_pixels(ctx->blocks[7],
587 ptr_v + ctx->dct_uv_offset, ctx->m.uvlinesize);
591 static av_always_inline
592 int dnxhd_switch_matrix(DNXHDEncContext *ctx, int i)
595 ctx->m.q_intra_matrix16 = ctx->qmatrix_c16;
596 ctx->m.q_intra_matrix = ctx->qmatrix_c;
599 ctx->m.q_intra_matrix16 = ctx->qmatrix_l16;
600 ctx->m.q_intra_matrix = ctx->qmatrix_l;
605 static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg,
606 int jobnr, int threadnr)
608 DNXHDEncContext *ctx = avctx->priv_data;
609 int mb_y = jobnr, mb_x;
610 int qscale = ctx->qscale;
611 LOCAL_ALIGNED_16(int16_t, block, [64]);
612 ctx = ctx->thread[threadnr];
616 ctx->m.last_dc[2] = 1 << (ctx->cid_table->bit_depth + 2);
618 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
619 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
625 dnxhd_get_blocks(ctx, mb_x, mb_y);
627 for (i = 0; i < 8; i++) {
628 int16_t *src_block = ctx->blocks[i];
629 int overflow, nbits, diff, last_index;
630 int n = dnxhd_switch_matrix(ctx, i);
632 memcpy(block, src_block, 64 * sizeof(*block));
633 last_index = ctx->m.dct_quantize(&ctx->m, block, i,
635 ac_bits += dnxhd_calc_ac_bits(ctx, block, last_index);
637 diff = block[0] - ctx->m.last_dc[n];
639 nbits = av_log2_16bit(-2 * diff);
641 nbits = av_log2_16bit(2 * diff);
643 assert(nbits < ctx->cid_table->bit_depth + 4);
644 dc_bits += ctx->cid_table->dc_bits[nbits] + nbits;
646 ctx->m.last_dc[n] = block[0];
648 if (avctx->mb_decision == FF_MB_DECISION_RD || !RC_VARIANCE) {
649 dnxhd_unquantize_c(ctx, block, i, qscale, last_index);
650 ctx->m.idsp.idct(block);
651 ssd += dnxhd_ssd_block(block, src_block);
654 ctx->mb_rc[qscale][mb].ssd = ssd;
655 ctx->mb_rc[qscale][mb].bits = ac_bits + dc_bits + 12 +
656 8 * ctx->vlc_bits[0];
661 static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg,
662 int jobnr, int threadnr)
664 DNXHDEncContext *ctx = avctx->priv_data;
665 int mb_y = jobnr, mb_x;
666 ctx = ctx->thread[threadnr];
667 init_put_bits(&ctx->m.pb, (uint8_t *)arg + 640 + ctx->slice_offs[jobnr],
668 ctx->slice_size[jobnr]);
672 ctx->m.last_dc[2] = 1 << (ctx->cid_table->bit_depth + 2);
673 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
674 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
675 int qscale = ctx->mb_qscale[mb];
678 put_bits(&ctx->m.pb, 12, qscale << 1);
680 dnxhd_get_blocks(ctx, mb_x, mb_y);
682 for (i = 0; i < 8; i++) {
683 int16_t *block = ctx->blocks[i];
684 int overflow, n = dnxhd_switch_matrix(ctx, i);
685 int last_index = ctx->m.dct_quantize(&ctx->m, block, i,
688 dnxhd_encode_block(ctx, block, last_index, n);
689 // STOP_TIMER("encode_block");
692 if (put_bits_count(&ctx->m.pb) & 31)
693 put_bits(&ctx->m.pb, 32 - (put_bits_count(&ctx->m.pb) & 31), 0);
694 flush_put_bits(&ctx->m.pb);
698 static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx)
702 for (mb_y = 0; mb_y < ctx->m.mb_height; mb_y++) {
704 ctx->slice_offs[mb_y] = offset;
705 ctx->slice_size[mb_y] = 0;
706 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
707 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
708 ctx->slice_size[mb_y] += ctx->mb_bits[mb];
710 ctx->slice_size[mb_y] = (ctx->slice_size[mb_y] + 31) & ~31;
711 ctx->slice_size[mb_y] >>= 3;
712 thread_size = ctx->slice_size[mb_y];
713 offset += thread_size;
717 static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg,
718 int jobnr, int threadnr)
720 DNXHDEncContext *ctx = avctx->priv_data;
721 int mb_y = jobnr, mb_x, x, y;
722 int partial_last_row = (mb_y == ctx->m.mb_height - 1) &&
723 ((avctx->height >> ctx->interlaced) & 0xF);
725 ctx = ctx->thread[threadnr];
726 if (ctx->cid_table->bit_depth == 8) {
727 uint8_t *pix = ctx->thread[0]->src[0] + ((mb_y << 4) * ctx->m.linesize);
728 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x, pix += 16) {
729 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
733 if (!partial_last_row && mb_x * 16 <= avctx->width - 16) {
734 sum = ctx->m.mpvencdsp.pix_sum(pix, ctx->m.linesize);
735 varc = ctx->m.mpvencdsp.pix_norm1(pix, ctx->m.linesize);
737 int bw = FFMIN(avctx->width - 16 * mb_x, 16);
738 int bh = FFMIN((avctx->height >> ctx->interlaced) - 16 * mb_y, 16);
740 for (y = 0; y < bh; y++) {
741 for (x = 0; x < bw; x++) {
742 uint8_t val = pix[x + y * ctx->m.linesize];
748 varc = (varc - (((unsigned) sum * sum) >> 8) + 128) >> 8;
750 ctx->mb_cmp[mb].value = varc;
751 ctx->mb_cmp[mb].mb = mb;
754 const int linesize = ctx->m.linesize >> 1;
755 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x) {
756 uint16_t *pix = (uint16_t *)ctx->thread[0]->src[0] +
757 ((mb_y << 4) * linesize) + (mb_x << 4);
758 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
763 // Macroblocks are 16x16 pixels, unlike DCT blocks which are 8x8.
764 for (i = 0; i < 16; ++i) {
765 for (j = 0; j < 16; ++j) {
766 // Turn 16-bit pixels into 10-bit ones.
767 const int sample = (unsigned) pix[j] >> 6;
769 sqsum += sample * sample;
770 // 2^10 * 2^10 * 16 * 16 = 2^28, which is less than INT_MAX
774 mean = sum >> 8; // 16*16 == 2^8
776 ctx->mb_cmp[mb].value = sqmean - mean * mean;
777 ctx->mb_cmp[mb].mb = mb;
783 static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx)
785 int lambda, up_step, down_step;
786 int last_lower = INT_MAX, last_higher = 0;
789 for (q = 1; q < avctx->qmax; q++) {
791 avctx->execute2(avctx, dnxhd_calc_bits_thread,
792 NULL, NULL, ctx->m.mb_height);
794 up_step = down_step = 2 << LAMBDA_FRAC_BITS;
795 lambda = ctx->lambda;
800 if (lambda == last_higher) {
802 end = 1; // need to set final qscales/bits
804 for (y = 0; y < ctx->m.mb_height; y++) {
805 for (x = 0; x < ctx->m.mb_width; x++) {
806 unsigned min = UINT_MAX;
808 int mb = y * ctx->m.mb_width + x;
809 for (q = 1; q < avctx->qmax; q++) {
810 unsigned score = ctx->mb_rc[q][mb].bits * lambda +
811 ((unsigned) ctx->mb_rc[q][mb].ssd << LAMBDA_FRAC_BITS);
817 bits += ctx->mb_rc[qscale][mb].bits;
818 ctx->mb_qscale[mb] = qscale;
819 ctx->mb_bits[mb] = ctx->mb_rc[qscale][mb].bits;
821 bits = (bits + 31) & ~31; // padding
822 if (bits > ctx->frame_bits)
826 if (bits > ctx->frame_bits)
827 return AVERROR(EINVAL);
830 if (bits < ctx->frame_bits) {
831 last_lower = FFMIN(lambda, last_lower);
832 if (last_higher != 0)
833 lambda = (lambda+last_higher)>>1;
836 down_step = FFMIN((int64_t)down_step*5, INT_MAX);
837 up_step = 1<<LAMBDA_FRAC_BITS;
838 lambda = FFMAX(1, lambda);
839 if (lambda == last_lower)
842 last_higher = FFMAX(lambda, last_higher);
843 if (last_lower != INT_MAX)
844 lambda = (lambda+last_lower)>>1;
845 else if ((int64_t)lambda + up_step > INT_MAX)
846 return AVERROR(EINVAL);
849 up_step = FFMIN((int64_t)up_step*5, INT_MAX);
850 down_step = 1<<LAMBDA_FRAC_BITS;
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 if (bits < ctx->frame_bits) {
884 if (last_higher == qscale - 1) {
885 qscale = last_higher;
888 last_lower = FFMIN(qscale, last_lower);
889 if (last_higher != 0)
890 qscale = (qscale + last_higher) >> 1;
892 qscale -= down_step++;
897 if (last_lower == qscale + 1)
899 last_higher = FFMAX(qscale, last_higher);
900 if (last_lower != INT_MAX)
901 qscale = (qscale + last_lower) >> 1;
905 if (qscale >= ctx->m.avctx->qmax)
906 return AVERROR(EINVAL);
909 ctx->qscale = qscale;
913 #define BUCKET_BITS 8
914 #define RADIX_PASSES 4
915 #define NBUCKETS (1 << BUCKET_BITS)
917 static inline int get_bucket(int value, int shift)
920 value &= NBUCKETS - 1;
921 return NBUCKETS - 1 - value;
924 static void radix_count(const RCCMPEntry *data, int size,
925 int buckets[RADIX_PASSES][NBUCKETS])
928 memset(buckets, 0, sizeof(buckets[0][0]) * RADIX_PASSES * NBUCKETS);
929 for (i = 0; i < size; i++) {
930 int v = data[i].value;
931 for (j = 0; j < RADIX_PASSES; j++) {
932 buckets[j][get_bucket(v, 0)]++;
937 for (j = 0; j < RADIX_PASSES; j++) {
939 for (i = NBUCKETS - 1; i >= 0; i--)
940 buckets[j][i] = offset -= buckets[j][i];
941 assert(!buckets[j][0]);
945 static void radix_sort_pass(RCCMPEntry *dst, const RCCMPEntry *data,
946 int size, int buckets[NBUCKETS], int pass)
948 int shift = pass * BUCKET_BITS;
950 for (i = 0; i < size; i++) {
951 int v = get_bucket(data[i].value, shift);
952 int pos = buckets[v]++;
957 static void radix_sort(RCCMPEntry *data, int size)
959 int buckets[RADIX_PASSES][NBUCKETS];
960 RCCMPEntry *tmp = av_malloc(sizeof(*tmp) * size);
961 radix_count(data, size, buckets);
962 radix_sort_pass(tmp, data, size, buckets[0], 0);
963 radix_sort_pass(data, tmp, size, buckets[1], 1);
964 if (buckets[2][NBUCKETS - 1] || buckets[3][NBUCKETS - 1]) {
965 radix_sort_pass(tmp, data, size, buckets[2], 2);
966 radix_sort_pass(data, tmp, size, buckets[3], 3);
971 static int dnxhd_encode_fast(AVCodecContext *avctx, DNXHDEncContext *ctx)
975 if ((ret = dnxhd_find_qscale(ctx)) < 0)
977 for (y = 0; y < ctx->m.mb_height; y++) {
978 for (x = 0; x < ctx->m.mb_width; x++) {
979 int mb = y * ctx->m.mb_width + x;
981 ctx->mb_qscale[mb] = ctx->qscale;
982 ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale][mb].bits;
983 max_bits += ctx->mb_rc[ctx->qscale][mb].bits;
985 delta_bits = ctx->mb_rc[ctx->qscale][mb].bits -
986 ctx->mb_rc[ctx->qscale + 1][mb].bits;
987 ctx->mb_cmp[mb].mb = mb;
988 ctx->mb_cmp[mb].value =
989 delta_bits ? ((ctx->mb_rc[ctx->qscale][mb].ssd -
990 ctx->mb_rc[ctx->qscale + 1][mb].ssd) * 100) /
992 : INT_MIN; // avoid increasing qscale
995 max_bits += 31; // worst padding
999 avctx->execute2(avctx, dnxhd_mb_var_thread,
1000 NULL, NULL, ctx->m.mb_height);
1001 radix_sort(ctx->mb_cmp, ctx->m.mb_num);
1002 for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) {
1003 int mb = ctx->mb_cmp[x].mb;
1004 max_bits -= ctx->mb_rc[ctx->qscale][mb].bits -
1005 ctx->mb_rc[ctx->qscale + 1][mb].bits;
1006 ctx->mb_qscale[mb] = ctx->qscale + 1;
1007 ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale + 1][mb].bits;
1013 static void dnxhd_load_picture(DNXHDEncContext *ctx, const AVFrame *frame)
1017 for (i = 0; i < ctx->m.avctx->thread_count; i++) {
1018 ctx->thread[i]->m.linesize = frame->linesize[0] << ctx->interlaced;
1019 ctx->thread[i]->m.uvlinesize = frame->linesize[1] << ctx->interlaced;
1020 ctx->thread[i]->dct_y_offset = ctx->m.linesize *8;
1021 ctx->thread[i]->dct_uv_offset = ctx->m.uvlinesize*8;
1024 #if FF_API_CODED_FRAME
1025 FF_DISABLE_DEPRECATION_WARNINGS
1026 ctx->m.avctx->coded_frame->interlaced_frame = frame->interlaced_frame;
1027 FF_ENABLE_DEPRECATION_WARNINGS
1029 ctx->cur_field = frame->interlaced_frame && !frame->top_field_first;
1032 static int dnxhd_encode_picture(AVCodecContext *avctx, AVPacket *pkt,
1033 const AVFrame *frame, int *got_packet)
1035 DNXHDEncContext *ctx = avctx->priv_data;
1036 int first_field = 1;
1040 if ((ret = ff_alloc_packet(pkt, ctx->cid_table->frame_size)) < 0) {
1041 av_log(avctx, AV_LOG_ERROR,
1042 "output buffer is too small to compress picture\n");
1047 dnxhd_load_picture(ctx, frame);
1050 for (i = 0; i < 3; i++) {
1051 ctx->src[i] = frame->data[i];
1052 if (ctx->interlaced && ctx->cur_field)
1053 ctx->src[i] += frame->linesize[i];
1056 dnxhd_write_header(avctx, buf);
1058 if (avctx->mb_decision == FF_MB_DECISION_RD)
1059 ret = dnxhd_encode_rdo(avctx, ctx);
1061 ret = dnxhd_encode_fast(avctx, ctx);
1063 av_log(avctx, AV_LOG_ERROR,
1064 "picture could not fit ratecontrol constraints, increase qmax\n");
1068 dnxhd_setup_threads_slices(ctx);
1071 for (i = 0; i < ctx->m.mb_height; i++) {
1072 AV_WB32(ctx->msip + i * 4, offset);
1073 offset += ctx->slice_size[i];
1074 assert(!(ctx->slice_size[i] & 3));
1077 avctx->execute2(avctx, dnxhd_encode_thread, buf, NULL, ctx->m.mb_height);
1079 assert(640 + offset + 4 <= ctx->cid_table->coding_unit_size);
1080 memset(buf + 640 + offset, 0,
1081 ctx->cid_table->coding_unit_size - 4 - offset - 640);
1083 AV_WB32(buf + ctx->cid_table->coding_unit_size - 4, 0x600DC0DE); // EOF
1085 if (ctx->interlaced && first_field) {
1087 ctx->cur_field ^= 1;
1088 buf += ctx->cid_table->coding_unit_size;
1089 goto encode_coding_unit;
1092 #if FF_API_CODED_FRAME
1093 FF_DISABLE_DEPRECATION_WARNINGS
1094 avctx->coded_frame->quality = ctx->qscale * FF_QP2LAMBDA;
1095 FF_ENABLE_DEPRECATION_WARNINGS
1098 sd = av_packet_new_side_data(pkt, AV_PKT_DATA_QUALITY_FACTOR, sizeof(int));
1100 return AVERROR(ENOMEM);
1101 *(int *)sd = ctx->qscale * FF_QP2LAMBDA;
1103 pkt->flags |= AV_PKT_FLAG_KEY;
1108 static av_cold int dnxhd_encode_end(AVCodecContext *avctx)
1110 DNXHDEncContext *ctx = avctx->priv_data;
1111 int max_level = 1 << (ctx->cid_table->bit_depth + 2);
1114 av_free(ctx->vlc_codes - max_level * 2);
1115 av_free(ctx->vlc_bits - max_level * 2);
1116 av_freep(&ctx->run_codes);
1117 av_freep(&ctx->run_bits);
1119 av_freep(&ctx->mb_bits);
1120 av_freep(&ctx->mb_qscale);
1121 av_freep(&ctx->mb_rc);
1122 av_freep(&ctx->mb_cmp);
1123 av_freep(&ctx->slice_size);
1124 av_freep(&ctx->slice_offs);
1126 av_freep(&ctx->qmatrix_c);
1127 av_freep(&ctx->qmatrix_l);
1128 av_freep(&ctx->qmatrix_c16);
1129 av_freep(&ctx->qmatrix_l16);
1131 for (i = 1; i < avctx->thread_count; i++)
1132 av_freep(&ctx->thread[i]);
1137 AVCodec ff_dnxhd_encoder = {
1139 .long_name = NULL_IF_CONFIG_SMALL("VC3/DNxHD"),
1140 .type = AVMEDIA_TYPE_VIDEO,
1141 .id = AV_CODEC_ID_DNXHD,
1142 .priv_data_size = sizeof(DNXHDEncContext),
1143 .init = dnxhd_encode_init,
1144 .encode2 = dnxhd_encode_picture,
1145 .close = dnxhd_encode_end,
1146 .capabilities = AV_CODEC_CAP_SLICE_THREADS,
1147 .pix_fmts = (const enum AVPixelFormat[]) {
1149 AV_PIX_FMT_YUV422P10,
1152 .priv_class = &class,