4 * Copyright (c) 2012 Konstantin Shishkov
6 * This file is part of Libav.
8 * Libav is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
13 * Libav is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with Libav; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23 #include "libavutil/opt.h"
24 #include "libavutil/pixdesc.h"
28 #include "bytestream.h"
30 #include "proresdata.h"
32 #define CFACTOR_Y422 2
33 #define CFACTOR_Y444 3
35 #define MAX_MBS_PER_SLICE 8
40 PRORES_PROFILE_PROXY = 0,
42 PRORES_PROFILE_STANDARD,
55 static const uint8_t prores_quant_matrices[][64] = {
57 4, 7, 9, 11, 13, 14, 15, 63,
58 7, 7, 11, 12, 14, 15, 63, 63,
59 9, 11, 13, 14, 15, 63, 63, 63,
60 11, 11, 13, 14, 63, 63, 63, 63,
61 11, 13, 14, 63, 63, 63, 63, 63,
62 13, 14, 63, 63, 63, 63, 63, 63,
63 13, 63, 63, 63, 63, 63, 63, 63,
64 63, 63, 63, 63, 63, 63, 63, 63,
67 4, 5, 6, 7, 9, 11, 13, 15,
68 5, 5, 7, 8, 11, 13, 15, 17,
69 6, 7, 9, 11, 13, 15, 15, 17,
70 7, 7, 9, 11, 13, 15, 17, 19,
71 7, 9, 11, 13, 14, 16, 19, 23,
72 9, 11, 13, 14, 16, 19, 23, 29,
73 9, 11, 13, 15, 17, 21, 28, 35,
74 11, 13, 16, 17, 21, 28, 35, 41,
77 4, 4, 5, 5, 6, 7, 7, 9,
78 4, 4, 5, 6, 7, 7, 9, 9,
79 5, 5, 6, 7, 7, 9, 9, 10,
80 5, 5, 6, 7, 7, 9, 9, 10,
81 5, 6, 7, 7, 8, 9, 10, 12,
82 6, 7, 7, 8, 9, 10, 12, 15,
83 6, 7, 7, 9, 10, 11, 14, 17,
84 7, 7, 9, 10, 11, 14, 17, 21,
87 4, 4, 4, 4, 4, 4, 4, 4,
88 4, 4, 4, 4, 4, 4, 4, 4,
89 4, 4, 4, 4, 4, 4, 4, 4,
90 4, 4, 4, 4, 4, 4, 4, 5,
91 4, 4, 4, 4, 4, 4, 5, 5,
92 4, 4, 4, 4, 4, 5, 5, 6,
93 4, 4, 4, 4, 5, 5, 6, 7,
94 4, 4, 4, 4, 5, 6, 7, 7,
97 4, 4, 4, 4, 4, 4, 4, 4,
98 4, 4, 4, 4, 4, 4, 4, 4,
99 4, 4, 4, 4, 4, 4, 4, 4,
100 4, 4, 4, 4, 4, 4, 4, 4,
101 4, 4, 4, 4, 4, 4, 4, 4,
102 4, 4, 4, 4, 4, 4, 4, 4,
103 4, 4, 4, 4, 4, 4, 4, 4,
104 4, 4, 4, 4, 4, 4, 4, 4,
108 #define NUM_MB_LIMITS 4
109 static const int prores_mb_limits[NUM_MB_LIMITS] = {
110 1620, // up to 720x576
111 2700, // up to 960x720
112 6075, // up to 1440x1080
113 9216, // up to 2048x1152
116 static const struct prores_profile {
117 const char *full_name;
121 int br_tab[NUM_MB_LIMITS];
123 } prores_profile_info[5] = {
125 .full_name = "proxy",
126 .tag = MKTAG('a', 'p', 'c', 'o'),
129 .br_tab = { 300, 242, 220, 194 },
130 .quant = QUANT_MAT_PROXY,
134 .tag = MKTAG('a', 'p', 'c', 's'),
137 .br_tab = { 720, 560, 490, 440 },
138 .quant = QUANT_MAT_LT,
141 .full_name = "standard",
142 .tag = MKTAG('a', 'p', 'c', 'n'),
145 .br_tab = { 1050, 808, 710, 632 },
146 .quant = QUANT_MAT_STANDARD,
149 .full_name = "high quality",
150 .tag = MKTAG('a', 'p', 'c', 'h'),
153 .br_tab = { 1566, 1216, 1070, 950 },
154 .quant = QUANT_MAT_HQ,
158 .tag = MKTAG('a', 'p', '4', 'h'),
161 .br_tab = { 2350, 1828, 1600, 1425 },
162 .quant = QUANT_MAT_HQ,
166 #define TRELLIS_WIDTH 16
167 #define SCORE_LIMIT INT_MAX / 2
176 #define MAX_STORED_Q 16
178 typedef struct ProresThreadData {
179 DECLARE_ALIGNED(16, int16_t, blocks)[MAX_PLANES][64 * 4 * MAX_MBS_PER_SLICE];
180 DECLARE_ALIGNED(16, uint16_t, emu_buf)[16 * 16];
181 int16_t custom_q[64];
182 struct TrellisNode *nodes;
185 typedef struct ProresContext {
187 DECLARE_ALIGNED(16, int16_t, blocks)[MAX_PLANES][64 * 4 * MAX_MBS_PER_SLICE];
188 DECLARE_ALIGNED(16, uint16_t, emu_buf)[16*16];
189 int16_t quants[MAX_STORED_Q][64];
190 int16_t custom_q[64];
191 const uint8_t *quant_mat;
192 const uint8_t *scantable;
194 void (*fdct)(FDCTDSPContext *fdsp, const uint16_t *src,
195 int linesize, int16_t *block);
198 int mb_width, mb_height;
200 int num_chroma_blocks, chroma_factor;
202 int slices_per_picture;
203 int pictures_per_frame; // 1 for progressive, 2 for interlaced
214 int frame_size_upper_bound;
217 const struct prores_profile *profile_info;
221 ProresThreadData *tdata;
224 static void get_slice_data(ProresContext *ctx, const uint16_t *src,
225 int linesize, int x, int y, int w, int h,
226 int16_t *blocks, uint16_t *emu_buf,
227 int mbs_per_slice, int blocks_per_mb, int is_chroma)
229 const uint16_t *esrc;
230 const int mb_width = 4 * blocks_per_mb;
234 for (i = 0; i < mbs_per_slice; i++, src += mb_width) {
236 memset(blocks, 0, 64 * (mbs_per_slice - i) * blocks_per_mb
240 if (x + mb_width <= w && y + 16 <= h) {
242 elinesize = linesize;
247 elinesize = 16 * sizeof(*emu_buf);
249 bw = FFMIN(w - x, mb_width);
250 bh = FFMIN(h - y, 16);
252 for (j = 0; j < bh; j++) {
253 memcpy(emu_buf + j * 16,
254 (const uint8_t*)src + j * linesize,
256 pix = emu_buf[j * 16 + bw - 1];
257 for (k = bw; k < mb_width; k++)
258 emu_buf[j * 16 + k] = pix;
261 memcpy(emu_buf + j * 16,
262 emu_buf + (bh - 1) * 16,
263 mb_width * sizeof(*emu_buf));
266 ctx->fdct(&ctx->fdsp, esrc, elinesize, blocks);
268 if (blocks_per_mb > 2) {
269 ctx->fdct(&ctx->fdsp, esrc + 8, elinesize, blocks);
272 ctx->fdct(&ctx->fdsp, esrc + elinesize * 4, elinesize, blocks);
274 if (blocks_per_mb > 2) {
275 ctx->fdct(&ctx->fdsp, esrc + elinesize * 4 + 8, elinesize, blocks);
279 ctx->fdct(&ctx->fdsp, esrc, elinesize, blocks);
281 ctx->fdct(&ctx->fdsp, esrc + elinesize * 4, elinesize, blocks);
283 if (blocks_per_mb > 2) {
284 ctx->fdct(&ctx->fdsp, esrc + 8, elinesize, blocks);
286 ctx->fdct(&ctx->fdsp, esrc + elinesize * 4 + 8, elinesize, blocks);
295 static void get_alpha_data(ProresContext *ctx, const uint16_t *src,
296 int linesize, int x, int y, int w, int h,
297 int16_t *blocks, int mbs_per_slice, int abits)
299 const int slice_width = 16 * mbs_per_slice;
300 int i, j, copy_w, copy_h;
302 copy_w = FFMIN(w - x, slice_width);
303 copy_h = FFMIN(h - y, 16);
304 for (i = 0; i < copy_h; i++) {
305 memcpy(blocks, src, copy_w * sizeof(*src));
307 for (j = 0; j < copy_w; j++)
310 for (j = 0; j < copy_w; j++)
311 blocks[j] = (blocks[j] << 6) | (blocks[j] >> 4);
312 for (j = copy_w; j < slice_width; j++)
313 blocks[j] = blocks[copy_w - 1];
314 blocks += slice_width;
315 src += linesize >> 1;
317 for (; i < 16; i++) {
318 memcpy(blocks, blocks - slice_width, slice_width * sizeof(*blocks));
319 blocks += slice_width;
324 * Write an unsigned rice/exp golomb codeword.
326 static inline void encode_vlc_codeword(PutBitContext *pb, unsigned codebook, int val)
328 unsigned int rice_order, exp_order, switch_bits, switch_val;
331 /* number of prefix bits to switch between Rice and expGolomb */
332 switch_bits = (codebook & 3) + 1;
333 rice_order = codebook >> 5; /* rice code order */
334 exp_order = (codebook >> 2) & 7; /* exp golomb code order */
336 switch_val = switch_bits << rice_order;
338 if (val >= switch_val) {
339 val -= switch_val - (1 << exp_order);
340 exponent = av_log2(val);
342 put_bits(pb, exponent - exp_order + switch_bits, 0);
343 put_bits(pb, exponent + 1, val);
345 exponent = val >> rice_order;
348 put_bits(pb, exponent, 0);
351 put_sbits(pb, rice_order, val);
355 #define GET_SIGN(x) ((x) >> 31)
356 #define MAKE_CODE(x) (((x) << 1) ^ GET_SIGN(x))
358 static void encode_dcs(PutBitContext *pb, int16_t *blocks,
359 int blocks_per_slice, int scale)
362 int codebook = 3, code, dc, prev_dc, delta, sign, new_sign;
364 prev_dc = (blocks[0] - 0x4000) / scale;
365 encode_vlc_codeword(pb, FIRST_DC_CB, MAKE_CODE(prev_dc));
370 for (i = 1; i < blocks_per_slice; i++, blocks += 64) {
371 dc = (blocks[0] - 0x4000) / scale;
372 delta = dc - prev_dc;
373 new_sign = GET_SIGN(delta);
374 delta = (delta ^ sign) - sign;
375 code = MAKE_CODE(delta);
376 encode_vlc_codeword(pb, ff_prores_dc_codebook[codebook], code);
377 codebook = (code + (code & 1)) >> 1;
378 codebook = FFMIN(codebook, 3);
384 static void encode_acs(PutBitContext *pb, int16_t *blocks,
385 int blocks_per_slice,
386 int plane_size_factor,
387 const uint8_t *scan, const int16_t *qmat)
390 int run, level, run_cb, lev_cb;
391 int max_coeffs, abs_level;
393 max_coeffs = blocks_per_slice << 6;
394 run_cb = ff_prores_run_to_cb_index[4];
395 lev_cb = ff_prores_lev_to_cb_index[2];
398 for (i = 1; i < 64; i++) {
399 for (idx = scan[i]; idx < max_coeffs; idx += 64) {
400 level = blocks[idx] / qmat[scan[i]];
402 abs_level = FFABS(level);
403 encode_vlc_codeword(pb, ff_prores_ac_codebook[run_cb], run);
404 encode_vlc_codeword(pb, ff_prores_ac_codebook[lev_cb],
406 put_sbits(pb, 1, GET_SIGN(level));
408 run_cb = ff_prores_run_to_cb_index[FFMIN(run, 15)];
409 lev_cb = ff_prores_lev_to_cb_index[FFMIN(abs_level, 9)];
418 static int encode_slice_plane(ProresContext *ctx, PutBitContext *pb,
419 const uint16_t *src, int linesize,
420 int mbs_per_slice, int16_t *blocks,
421 int blocks_per_mb, int plane_size_factor,
424 int blocks_per_slice, saved_pos;
426 saved_pos = put_bits_count(pb);
427 blocks_per_slice = mbs_per_slice * blocks_per_mb;
429 encode_dcs(pb, blocks, blocks_per_slice, qmat[0]);
430 encode_acs(pb, blocks, blocks_per_slice, plane_size_factor,
431 ctx->scantable, qmat);
434 return (put_bits_count(pb) - saved_pos) >> 3;
437 static void put_alpha_diff(PutBitContext *pb, int cur, int prev, int abits)
439 const int mask = (1 << abits) - 1;
440 const int dbits = (abits == 8) ? 4 : 7;
441 const int dsize = 1 << dbits - 1;
442 int diff = cur - prev;
445 if (diff >= (1 << abits) - dsize)
447 if (diff < -dsize || diff > dsize || !diff) {
449 put_bits(pb, abits, diff);
452 put_bits(pb, dbits - 1, FFABS(diff) - 1);
453 put_bits(pb, 1, diff < 0);
457 static void put_alpha_run(PutBitContext *pb, int run)
462 put_bits(pb, 4, run);
464 put_bits(pb, 15, run);
470 // todo alpha quantisation for high quants
471 static int encode_alpha_plane(ProresContext *ctx, PutBitContext *pb,
472 int mbs_per_slice, uint16_t *blocks,
475 const int abits = ctx->alpha_bits;
476 const int mask = (1 << abits) - 1;
477 const int num_coeffs = mbs_per_slice * 256;
478 int saved_pos = put_bits_count(pb);
479 int prev = mask, cur;
484 put_alpha_diff(pb, cur, prev, abits);
489 put_alpha_run (pb, run);
490 put_alpha_diff(pb, cur, prev, abits);
496 } while (idx < num_coeffs);
498 put_alpha_run(pb, run);
500 return (put_bits_count(pb) - saved_pos) >> 3;
503 static int encode_slice(AVCodecContext *avctx, const AVFrame *pic,
505 int sizes[4], int x, int y, int quant,
508 ProresContext *ctx = avctx->priv_data;
512 int slice_width_factor = av_log2(mbs_per_slice);
513 int num_cblocks, pwidth, linesize, line_add;
514 int plane_factor, is_chroma;
517 if (ctx->pictures_per_frame == 1)
520 line_add = ctx->cur_picture_idx ^ !pic->top_field_first;
522 if (ctx->force_quant) {
523 qmat = ctx->quants[0];
524 } else if (quant < MAX_STORED_Q) {
525 qmat = ctx->quants[quant];
527 qmat = ctx->custom_q;
528 for (i = 0; i < 64; i++)
529 qmat[i] = ctx->quant_mat[i] * quant;
532 for (i = 0; i < ctx->num_planes; i++) {
533 is_chroma = (i == 1 || i == 2);
534 plane_factor = slice_width_factor + 2;
536 plane_factor += ctx->chroma_factor - 3;
537 if (!is_chroma || ctx->chroma_factor == CFACTOR_Y444) {
541 pwidth = avctx->width;
546 pwidth = avctx->width >> 1;
549 linesize = pic->linesize[i] * ctx->pictures_per_frame;
550 src = (const uint16_t*)(pic->data[i] + yp * linesize +
551 line_add * pic->linesize[i]) + xp;
554 get_slice_data(ctx, src, linesize, xp, yp,
555 pwidth, avctx->height / ctx->pictures_per_frame,
556 ctx->blocks[0], ctx->emu_buf,
557 mbs_per_slice, num_cblocks, is_chroma);
558 sizes[i] = encode_slice_plane(ctx, pb, src, linesize,
559 mbs_per_slice, ctx->blocks[0],
560 num_cblocks, plane_factor,
563 get_alpha_data(ctx, src, linesize, xp, yp,
564 pwidth, avctx->height / ctx->pictures_per_frame,
565 ctx->blocks[0], mbs_per_slice, ctx->alpha_bits);
566 sizes[i] = encode_alpha_plane(ctx, pb, mbs_per_slice,
567 ctx->blocks[0], quant);
569 total_size += sizes[i];
570 if (put_bits_left(pb) < 0) {
571 av_log(avctx, AV_LOG_ERROR,
572 "Underestimated required buffer size.\n");
579 static inline int estimate_vlc(unsigned codebook, int val)
581 unsigned int rice_order, exp_order, switch_bits, switch_val;
584 /* number of prefix bits to switch between Rice and expGolomb */
585 switch_bits = (codebook & 3) + 1;
586 rice_order = codebook >> 5; /* rice code order */
587 exp_order = (codebook >> 2) & 7; /* exp golomb code order */
589 switch_val = switch_bits << rice_order;
591 if (val >= switch_val) {
592 val -= switch_val - (1 << exp_order);
593 exponent = av_log2(val);
595 return exponent * 2 - exp_order + switch_bits + 1;
597 return (val >> rice_order) + rice_order + 1;
601 static int estimate_dcs(int *error, int16_t *blocks, int blocks_per_slice,
605 int codebook = 3, code, dc, prev_dc, delta, sign, new_sign;
608 prev_dc = (blocks[0] - 0x4000) / scale;
609 bits = estimate_vlc(FIRST_DC_CB, MAKE_CODE(prev_dc));
613 *error += FFABS(blocks[0] - 0x4000) % scale;
615 for (i = 1; i < blocks_per_slice; i++, blocks += 64) {
616 dc = (blocks[0] - 0x4000) / scale;
617 *error += FFABS(blocks[0] - 0x4000) % scale;
618 delta = dc - prev_dc;
619 new_sign = GET_SIGN(delta);
620 delta = (delta ^ sign) - sign;
621 code = MAKE_CODE(delta);
622 bits += estimate_vlc(ff_prores_dc_codebook[codebook], code);
623 codebook = (code + (code & 1)) >> 1;
624 codebook = FFMIN(codebook, 3);
632 static int estimate_acs(int *error, int16_t *blocks, int blocks_per_slice,
633 int plane_size_factor,
634 const uint8_t *scan, const int16_t *qmat)
637 int run, level, run_cb, lev_cb;
638 int max_coeffs, abs_level;
641 max_coeffs = blocks_per_slice << 6;
642 run_cb = ff_prores_run_to_cb_index[4];
643 lev_cb = ff_prores_lev_to_cb_index[2];
646 for (i = 1; i < 64; i++) {
647 for (idx = scan[i]; idx < max_coeffs; idx += 64) {
648 level = blocks[idx] / qmat[scan[i]];
649 *error += FFABS(blocks[idx]) % qmat[scan[i]];
651 abs_level = FFABS(level);
652 bits += estimate_vlc(ff_prores_ac_codebook[run_cb], run);
653 bits += estimate_vlc(ff_prores_ac_codebook[lev_cb],
656 run_cb = ff_prores_run_to_cb_index[FFMIN(run, 15)];
657 lev_cb = ff_prores_lev_to_cb_index[FFMIN(abs_level, 9)];
668 static int estimate_slice_plane(ProresContext *ctx, int *error, int plane,
669 const uint16_t *src, int linesize,
671 int blocks_per_mb, int plane_size_factor,
672 const int16_t *qmat, ProresThreadData *td)
674 int blocks_per_slice;
677 blocks_per_slice = mbs_per_slice * blocks_per_mb;
679 bits = estimate_dcs(error, td->blocks[plane], blocks_per_slice, qmat[0]);
680 bits += estimate_acs(error, td->blocks[plane], blocks_per_slice,
681 plane_size_factor, ctx->scantable, qmat);
683 return FFALIGN(bits, 8);
686 static int est_alpha_diff(int cur, int prev, int abits)
688 const int mask = (1 << abits) - 1;
689 const int dbits = (abits == 8) ? 4 : 7;
690 const int dsize = 1 << dbits - 1;
691 int diff = cur - prev;
694 if (diff >= (1 << abits) - dsize)
696 if (diff < -dsize || diff > dsize || !diff)
702 static int estimate_alpha_plane(ProresContext *ctx, int *error,
703 const uint16_t *src, int linesize,
704 int mbs_per_slice, int quant,
707 const int abits = ctx->alpha_bits;
708 const int mask = (1 << abits) - 1;
709 const int num_coeffs = mbs_per_slice * 256;
710 int prev = mask, cur;
717 bits = est_alpha_diff(cur, prev, abits);
728 bits += est_alpha_diff(cur, prev, abits);
734 } while (idx < num_coeffs);
746 static int find_slice_quant(AVCodecContext *avctx, const AVFrame *pic,
747 int trellis_node, int x, int y, int mbs_per_slice,
748 ProresThreadData *td)
750 ProresContext *ctx = avctx->priv_data;
751 int i, q, pq, xp, yp;
753 int slice_width_factor = av_log2(mbs_per_slice);
754 int num_cblocks[MAX_PLANES], pwidth;
755 int plane_factor[MAX_PLANES], is_chroma[MAX_PLANES];
756 const int min_quant = ctx->profile_info->min_quant;
757 const int max_quant = ctx->profile_info->max_quant;
758 int error, bits, bits_limit;
759 int mbs, prev, cur, new_score;
760 int slice_bits[TRELLIS_WIDTH], slice_score[TRELLIS_WIDTH];
763 int linesize[4], line_add;
765 if (ctx->pictures_per_frame == 1)
768 line_add = ctx->cur_picture_idx ^ !pic->top_field_first;
769 mbs = x + mbs_per_slice;
771 for (i = 0; i < ctx->num_planes; i++) {
772 is_chroma[i] = (i == 1 || i == 2);
773 plane_factor[i] = slice_width_factor + 2;
775 plane_factor[i] += ctx->chroma_factor - 3;
776 if (!is_chroma[i] || ctx->chroma_factor == CFACTOR_Y444) {
780 pwidth = avctx->width;
785 pwidth = avctx->width >> 1;
788 linesize[i] = pic->linesize[i] * ctx->pictures_per_frame;
789 src = (const uint16_t*)(pic->data[i] + yp * linesize[i] +
790 line_add * pic->linesize[i]) + xp;
793 get_slice_data(ctx, src, linesize[i], xp, yp,
794 pwidth, avctx->height / ctx->pictures_per_frame,
795 td->blocks[i], td->emu_buf,
796 mbs_per_slice, num_cblocks[i], is_chroma[i]);
798 get_alpha_data(ctx, src, linesize[i], xp, yp,
799 pwidth, avctx->height / ctx->pictures_per_frame,
800 td->blocks[i], mbs_per_slice, ctx->alpha_bits);
804 for (q = min_quant; q < max_quant + 2; q++) {
805 td->nodes[trellis_node + q].prev_node = -1;
806 td->nodes[trellis_node + q].quant = q;
809 // todo: maybe perform coarser quantising to fit into frame size when needed
810 for (q = min_quant; q <= max_quant; q++) {
813 for (i = 0; i < ctx->num_planes - !!ctx->alpha_bits; i++) {
814 bits += estimate_slice_plane(ctx, &error, i,
817 num_cblocks[i], plane_factor[i],
821 bits += estimate_alpha_plane(ctx, &error, src, linesize[3],
822 mbs_per_slice, q, td->blocks[3]);
823 if (bits > 65000 * 8)
826 slice_bits[q] = bits;
827 slice_score[q] = error;
829 if (slice_bits[max_quant] <= ctx->bits_per_mb * mbs_per_slice) {
830 slice_bits[max_quant + 1] = slice_bits[max_quant];
831 slice_score[max_quant + 1] = slice_score[max_quant] + 1;
832 overquant = max_quant;
834 for (q = max_quant + 1; q < 128; q++) {
837 if (q < MAX_STORED_Q) {
838 qmat = ctx->quants[q];
841 for (i = 0; i < 64; i++)
842 qmat[i] = ctx->quant_mat[i] * q;
844 for (i = 0; i < ctx->num_planes - !!ctx->alpha_bits; i++) {
845 bits += estimate_slice_plane(ctx, &error, i,
848 num_cblocks[i], plane_factor[i],
852 bits += estimate_alpha_plane(ctx, &error, src, linesize[3],
853 mbs_per_slice, q, td->blocks[3]);
854 if (bits <= ctx->bits_per_mb * mbs_per_slice)
858 slice_bits[max_quant + 1] = bits;
859 slice_score[max_quant + 1] = error;
862 td->nodes[trellis_node + max_quant + 1].quant = overquant;
864 bits_limit = mbs * ctx->bits_per_mb;
865 for (pq = min_quant; pq < max_quant + 2; pq++) {
866 prev = trellis_node - TRELLIS_WIDTH + pq;
868 for (q = min_quant; q < max_quant + 2; q++) {
869 cur = trellis_node + q;
871 bits = td->nodes[prev].bits + slice_bits[q];
872 error = slice_score[q];
873 if (bits > bits_limit)
876 if (td->nodes[prev].score < SCORE_LIMIT && error < SCORE_LIMIT)
877 new_score = td->nodes[prev].score + error;
879 new_score = SCORE_LIMIT;
880 if (td->nodes[cur].prev_node == -1 ||
881 td->nodes[cur].score >= new_score) {
883 td->nodes[cur].bits = bits;
884 td->nodes[cur].score = new_score;
885 td->nodes[cur].prev_node = prev;
890 error = td->nodes[trellis_node + min_quant].score;
891 pq = trellis_node + min_quant;
892 for (q = min_quant + 1; q < max_quant + 2; q++) {
893 if (td->nodes[trellis_node + q].score <= error) {
894 error = td->nodes[trellis_node + q].score;
895 pq = trellis_node + q;
902 static int find_quant_thread(AVCodecContext *avctx, void *arg,
903 int jobnr, int threadnr)
905 ProresContext *ctx = avctx->priv_data;
906 ProresThreadData *td = ctx->tdata + threadnr;
907 int mbs_per_slice = ctx->mbs_per_slice;
908 int x, y = jobnr, mb, q = 0;
910 for (x = mb = 0; x < ctx->mb_width; x += mbs_per_slice, mb++) {
911 while (ctx->mb_width - x < mbs_per_slice)
913 q = find_slice_quant(avctx, avctx->coded_frame,
914 (mb + 1) * TRELLIS_WIDTH, x, y,
918 for (x = ctx->slices_width - 1; x >= 0; x--) {
919 ctx->slice_q[x + y * ctx->slices_width] = td->nodes[q].quant;
920 q = td->nodes[q].prev_node;
926 static int encode_frame(AVCodecContext *avctx, AVPacket *pkt,
927 const AVFrame *pic, int *got_packet)
929 ProresContext *ctx = avctx->priv_data;
930 uint8_t *orig_buf, *buf, *slice_hdr, *slice_sizes, *tmp;
931 uint8_t *picture_size_pos;
933 int x, y, i, mb, q = 0;
934 int sizes[4] = { 0 };
935 int slice_hdr_size = 2 + 2 * (ctx->num_planes - 1);
936 int frame_size, picture_size, slice_size;
937 int pkt_size, ret, max_slice_size = 0;
940 *avctx->coded_frame = *pic;
941 avctx->coded_frame->pict_type = AV_PICTURE_TYPE_I;
942 avctx->coded_frame->key_frame = 1;
944 pkt_size = ctx->frame_size_upper_bound;
946 if ((ret = ff_alloc_packet(pkt, pkt_size + FF_MIN_BUFFER_SIZE)) < 0) {
947 av_log(avctx, AV_LOG_ERROR, "Error getting output packet.\n");
951 orig_buf = pkt->data;
954 orig_buf += 4; // frame size
955 bytestream_put_be32 (&orig_buf, FRAME_ID); // frame container ID
960 buf += 2; // frame header size will be stored here
961 bytestream_put_be16 (&buf, 0); // version 1
962 bytestream_put_buffer(&buf, ctx->vendor, 4);
963 bytestream_put_be16 (&buf, avctx->width);
964 bytestream_put_be16 (&buf, avctx->height);
966 frame_flags = ctx->chroma_factor << 6;
967 if (avctx->flags & CODEC_FLAG_INTERLACED_DCT)
968 frame_flags |= pic->top_field_first ? 0x04 : 0x08;
969 bytestream_put_byte (&buf, frame_flags);
971 bytestream_put_byte (&buf, 0); // reserved
972 bytestream_put_byte (&buf, avctx->color_primaries);
973 bytestream_put_byte (&buf, avctx->color_trc);
974 bytestream_put_byte (&buf, avctx->colorspace);
975 bytestream_put_byte (&buf, 0x40 | (ctx->alpha_bits >> 3));
976 bytestream_put_byte (&buf, 0); // reserved
977 if (ctx->quant_sel != QUANT_MAT_DEFAULT) {
978 bytestream_put_byte (&buf, 0x03); // matrix flags - both matrices are present
979 // luma quantisation matrix
980 for (i = 0; i < 64; i++)
981 bytestream_put_byte(&buf, ctx->quant_mat[i]);
982 // chroma quantisation matrix
983 for (i = 0; i < 64; i++)
984 bytestream_put_byte(&buf, ctx->quant_mat[i]);
986 bytestream_put_byte (&buf, 0x00); // matrix flags - default matrices are used
988 bytestream_put_be16 (&tmp, buf - orig_buf); // write back frame header size
990 for (ctx->cur_picture_idx = 0;
991 ctx->cur_picture_idx < ctx->pictures_per_frame;
992 ctx->cur_picture_idx++) {
994 picture_size_pos = buf + 1;
995 bytestream_put_byte (&buf, 0x40); // picture header size (in bits)
996 buf += 4; // picture data size will be stored here
997 bytestream_put_be16 (&buf, ctx->slices_per_picture);
998 bytestream_put_byte (&buf, av_log2(ctx->mbs_per_slice) << 4); // slice width and height in MBs
1000 // seek table - will be filled during slice encoding
1002 buf += ctx->slices_per_picture * 2;
1005 if (!ctx->force_quant) {
1006 ret = avctx->execute2(avctx, find_quant_thread, NULL, NULL,
1012 for (y = 0; y < ctx->mb_height; y++) {
1013 int mbs_per_slice = ctx->mbs_per_slice;
1014 for (x = mb = 0; x < ctx->mb_width; x += mbs_per_slice, mb++) {
1015 q = ctx->force_quant ? ctx->force_quant
1016 : ctx->slice_q[mb + y * ctx->slices_width];
1018 while (ctx->mb_width - x < mbs_per_slice)
1019 mbs_per_slice >>= 1;
1021 bytestream_put_byte(&buf, slice_hdr_size << 3);
1023 buf += slice_hdr_size - 1;
1024 if (pkt_size <= buf - orig_buf + 2 * max_slice_size) {
1025 uint8_t *start = pkt->data;
1026 // Recompute new size according to max_slice_size
1028 int delta = 200 + ctx->pictures_per_frame *
1029 ctx->slices_per_picture * max_slice_size -
1032 delta = FFMAX(delta, 2 * max_slice_size);
1033 ctx->frame_size_upper_bound += delta;
1036 avpriv_request_sample(avctx,
1037 "Packet too small: is %i,"
1038 " needs %i (slice: %i). "
1039 "Correct allocation",
1040 pkt_size, delta, max_slice_size);
1044 ret = av_grow_packet(pkt, delta);
1050 orig_buf = pkt->data + (orig_buf - start);
1051 buf = pkt->data + (buf - start);
1052 picture_size_pos = pkt->data + (picture_size_pos - start);
1053 slice_sizes = pkt->data + (slice_sizes - start);
1054 slice_hdr = pkt->data + (slice_hdr - start);
1055 tmp = pkt->data + (tmp - start);
1057 init_put_bits(&pb, buf, (pkt_size - (buf - orig_buf)) * 8);
1058 ret = encode_slice(avctx, pic, &pb, sizes, x, y, q,
1063 bytestream_put_byte(&slice_hdr, q);
1064 slice_size = slice_hdr_size + sizes[ctx->num_planes - 1];
1065 for (i = 0; i < ctx->num_planes - 1; i++) {
1066 bytestream_put_be16(&slice_hdr, sizes[i]);
1067 slice_size += sizes[i];
1069 bytestream_put_be16(&slice_sizes, slice_size);
1070 buf += slice_size - slice_hdr_size;
1071 if (max_slice_size < slice_size)
1072 max_slice_size = slice_size;
1076 if (ctx->pictures_per_frame == 1)
1077 picture_size = buf - picture_size_pos - 6;
1079 picture_size = buf - picture_size_pos + 1;
1080 bytestream_put_be32(&picture_size_pos, picture_size);
1084 frame_size = buf - orig_buf;
1085 bytestream_put_be32(&orig_buf, frame_size);
1087 pkt->size = frame_size;
1088 pkt->flags |= AV_PKT_FLAG_KEY;
1094 static av_cold int encode_close(AVCodecContext *avctx)
1096 ProresContext *ctx = avctx->priv_data;
1099 av_freep(&avctx->coded_frame);
1102 for (i = 0; i < avctx->thread_count; i++)
1103 av_free(ctx->tdata[i].nodes);
1105 av_freep(&ctx->tdata);
1106 av_freep(&ctx->slice_q);
1111 static void prores_fdct(FDCTDSPContext *fdsp, const uint16_t *src,
1112 int linesize, int16_t *block)
1115 const uint16_t *tsrc = src;
1117 for (y = 0; y < 8; y++) {
1118 for (x = 0; x < 8; x++)
1119 block[y * 8 + x] = tsrc[x];
1120 tsrc += linesize >> 1;
1125 static av_cold int encode_init(AVCodecContext *avctx)
1127 ProresContext *ctx = avctx->priv_data;
1130 int min_quant, max_quant;
1131 int interlaced = !!(avctx->flags & CODEC_FLAG_INTERLACED_DCT);
1133 avctx->bits_per_raw_sample = 10;
1134 avctx->coded_frame = av_frame_alloc();
1135 if (!avctx->coded_frame)
1136 return AVERROR(ENOMEM);
1138 ctx->fdct = prores_fdct;
1139 ctx->scantable = interlaced ? ff_prores_interlaced_scan
1140 : ff_prores_progressive_scan;
1141 ff_fdctdsp_init(&ctx->fdsp, avctx);
1143 mps = ctx->mbs_per_slice;
1144 if (mps & (mps - 1)) {
1145 av_log(avctx, AV_LOG_ERROR,
1146 "there should be an integer power of two MBs per slice\n");
1147 return AVERROR(EINVAL);
1149 if (av_pix_fmt_desc_get(avctx->pix_fmt)->flags & AV_PIX_FMT_FLAG_ALPHA) {
1150 if (ctx->alpha_bits & 7) {
1151 av_log(avctx, AV_LOG_ERROR, "alpha bits should be 0, 8 or 16\n");
1152 return AVERROR(EINVAL);
1155 ctx->alpha_bits = 0;
1158 ctx->chroma_factor = avctx->pix_fmt == AV_PIX_FMT_YUV422P10
1161 ctx->profile_info = prores_profile_info + ctx->profile;
1162 ctx->num_planes = 3 + !!ctx->alpha_bits;
1164 ctx->mb_width = FFALIGN(avctx->width, 16) >> 4;
1167 ctx->mb_height = FFALIGN(avctx->height, 32) >> 5;
1169 ctx->mb_height = FFALIGN(avctx->height, 16) >> 4;
1171 ctx->slices_width = ctx->mb_width / mps;
1172 ctx->slices_width += av_popcount(ctx->mb_width - ctx->slices_width * mps);
1173 ctx->slices_per_picture = ctx->mb_height * ctx->slices_width;
1174 ctx->pictures_per_frame = 1 + interlaced;
1176 if (ctx->quant_sel == -1)
1177 ctx->quant_mat = prores_quant_matrices[ctx->profile_info->quant];
1179 ctx->quant_mat = prores_quant_matrices[ctx->quant_sel];
1181 if (strlen(ctx->vendor) != 4) {
1182 av_log(avctx, AV_LOG_ERROR, "vendor ID should be 4 bytes\n");
1183 return AVERROR_INVALIDDATA;
1186 ctx->force_quant = avctx->global_quality / FF_QP2LAMBDA;
1187 if (!ctx->force_quant) {
1188 if (!ctx->bits_per_mb) {
1189 for (i = 0; i < NUM_MB_LIMITS - 1; i++)
1190 if (prores_mb_limits[i] >= ctx->mb_width * ctx->mb_height *
1191 ctx->pictures_per_frame)
1193 ctx->bits_per_mb = ctx->profile_info->br_tab[i];
1194 } else if (ctx->bits_per_mb < 128) {
1195 av_log(avctx, AV_LOG_ERROR, "too few bits per MB, please set at least 128\n");
1196 return AVERROR_INVALIDDATA;
1199 min_quant = ctx->profile_info->min_quant;
1200 max_quant = ctx->profile_info->max_quant;
1201 for (i = min_quant; i < MAX_STORED_Q; i++) {
1202 for (j = 0; j < 64; j++)
1203 ctx->quants[i][j] = ctx->quant_mat[j] * i;
1206 ctx->slice_q = av_malloc(ctx->slices_per_picture * sizeof(*ctx->slice_q));
1207 if (!ctx->slice_q) {
1208 encode_close(avctx);
1209 return AVERROR(ENOMEM);
1212 ctx->tdata = av_mallocz(avctx->thread_count * sizeof(*ctx->tdata));
1214 encode_close(avctx);
1215 return AVERROR(ENOMEM);
1218 for (j = 0; j < avctx->thread_count; j++) {
1219 ctx->tdata[j].nodes = av_malloc((ctx->slices_width + 1)
1221 * sizeof(*ctx->tdata->nodes));
1222 if (!ctx->tdata[j].nodes) {
1223 encode_close(avctx);
1224 return AVERROR(ENOMEM);
1226 for (i = min_quant; i < max_quant + 2; i++) {
1227 ctx->tdata[j].nodes[i].prev_node = -1;
1228 ctx->tdata[j].nodes[i].bits = 0;
1229 ctx->tdata[j].nodes[i].score = 0;
1235 if (ctx->force_quant > 64) {
1236 av_log(avctx, AV_LOG_ERROR, "too large quantiser, maximum is 64\n");
1237 return AVERROR_INVALIDDATA;
1240 for (j = 0; j < 64; j++) {
1241 ctx->quants[0][j] = ctx->quant_mat[j] * ctx->force_quant;
1242 ls += av_log2((1 << 11) / ctx->quants[0][j]) * 2 + 1;
1245 ctx->bits_per_mb = ls * 8;
1246 if (ctx->chroma_factor == CFACTOR_Y444)
1247 ctx->bits_per_mb += ls * 4;
1250 ctx->frame_size_upper_bound = ctx->pictures_per_frame *
1251 ctx->slices_per_picture *
1252 (2 + 2 * ctx->num_planes +
1253 (mps * ctx->bits_per_mb) / 8)
1256 if (ctx->alpha_bits) {
1257 // The alpha plane is run-coded and might exceed the bit budget.
1258 ctx->frame_size_upper_bound += ctx->pictures_per_frame *
1259 ctx->slices_per_picture *
1260 /* num pixels per slice */ (ctx->mbs_per_slice * 256 *
1261 /* bits per pixel */ (1 + ctx->alpha_bits + 1) + 7 >> 3);
1264 avctx->codec_tag = ctx->profile_info->tag;
1266 av_log(avctx, AV_LOG_DEBUG,
1267 "profile %d, %d slices, interlacing: %s, %d bits per MB\n",
1268 ctx->profile, ctx->slices_per_picture * ctx->pictures_per_frame,
1269 interlaced ? "yes" : "no", ctx->bits_per_mb);
1270 av_log(avctx, AV_LOG_DEBUG, "frame size upper bound: %d\n",
1271 ctx->frame_size_upper_bound);
1276 #define OFFSET(x) offsetof(ProresContext, x)
1277 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
1279 static const AVOption options[] = {
1280 { "mbs_per_slice", "macroblocks per slice", OFFSET(mbs_per_slice),
1281 AV_OPT_TYPE_INT, { .i64 = 8 }, 1, MAX_MBS_PER_SLICE, VE },
1282 { "profile", NULL, OFFSET(profile), AV_OPT_TYPE_INT,
1283 { .i64 = PRORES_PROFILE_STANDARD },
1284 PRORES_PROFILE_PROXY, PRORES_PROFILE_4444, VE, "profile" },
1285 { "proxy", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = PRORES_PROFILE_PROXY },
1286 0, 0, VE, "profile" },
1287 { "lt", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = PRORES_PROFILE_LT },
1288 0, 0, VE, "profile" },
1289 { "standard", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = PRORES_PROFILE_STANDARD },
1290 0, 0, VE, "profile" },
1291 { "hq", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = PRORES_PROFILE_HQ },
1292 0, 0, VE, "profile" },
1293 { "4444", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = PRORES_PROFILE_4444 },
1294 0, 0, VE, "profile" },
1295 { "vendor", "vendor ID", OFFSET(vendor),
1296 AV_OPT_TYPE_STRING, { .str = "Lavc" }, CHAR_MIN, CHAR_MAX, VE },
1297 { "bits_per_mb", "desired bits per macroblock", OFFSET(bits_per_mb),
1298 AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 8192, VE },
1299 { "quant_mat", "quantiser matrix", OFFSET(quant_sel), AV_OPT_TYPE_INT,
1300 { .i64 = -1 }, -1, QUANT_MAT_DEFAULT, VE, "quant_mat" },
1301 { "auto", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = -1 },
1302 0, 0, VE, "quant_mat" },
1303 { "proxy", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = QUANT_MAT_PROXY },
1304 0, 0, VE, "quant_mat" },
1305 { "lt", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = QUANT_MAT_LT },
1306 0, 0, VE, "quant_mat" },
1307 { "standard", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = QUANT_MAT_STANDARD },
1308 0, 0, VE, "quant_mat" },
1309 { "hq", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = QUANT_MAT_HQ },
1310 0, 0, VE, "quant_mat" },
1311 { "default", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = QUANT_MAT_DEFAULT },
1312 0, 0, VE, "quant_mat" },
1313 { "alpha_bits", "bits for alpha plane", OFFSET(alpha_bits), AV_OPT_TYPE_INT,
1314 { .i64 = 16 }, 0, 16, VE },
1318 static const AVClass proresenc_class = {
1319 .class_name = "ProRes encoder",
1320 .item_name = av_default_item_name,
1322 .version = LIBAVUTIL_VERSION_INT,
1325 AVCodec ff_prores_encoder = {
1327 .long_name = NULL_IF_CONFIG_SMALL("Apple ProRes (iCodec Pro)"),
1328 .type = AVMEDIA_TYPE_VIDEO,
1329 .id = AV_CODEC_ID_PRORES,
1330 .priv_data_size = sizeof(ProresContext),
1331 .init = encode_init,
1332 .close = encode_close,
1333 .encode2 = encode_frame,
1334 .capabilities = CODEC_CAP_SLICE_THREADS,
1335 .pix_fmts = (const enum AVPixelFormat[]) {
1336 AV_PIX_FMT_YUV422P10, AV_PIX_FMT_YUV444P10,
1337 AV_PIX_FMT_YUVA444P10, AV_PIX_FMT_NONE
1339 .priv_class = &proresenc_class,