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 #define RC_VARIANCE 1 // use variance or ssd for fast rc
28 #include "libavutil/attributes.h"
29 #include "libavutil/internal.h"
30 #include "libavutil/opt.h"
31 #include "libavutil/timer.h"
35 #include "mpegvideo.h"
38 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
39 #define DNX10BIT_QMAT_SHIFT 18 // The largest value that will not lead to overflow for 10bit samples.
41 static const AVOption options[]={
42 {"nitris_compat", "encode with Avid Nitris compatibility", offsetof(DNXHDEncContext, nitris_compat), AV_OPT_TYPE_INT, {.i64 = 0}, 0, 1, VE},
45 static const AVClass class = { "dnxhd", av_default_item_name, options, LIBAVUTIL_VERSION_INT };
47 #define LAMBDA_FRAC_BITS 10
49 static void dnxhd_8bit_get_pixels_8x4_sym(int16_t *restrict block, const uint8_t *pixels, int line_size)
52 for (i = 0; i < 4; i++) {
53 block[0] = pixels[0]; block[1] = pixels[1];
54 block[2] = pixels[2]; block[3] = pixels[3];
55 block[4] = pixels[4]; block[5] = pixels[5];
56 block[6] = pixels[6]; block[7] = pixels[7];
60 memcpy(block, block - 8, sizeof(*block) * 8);
61 memcpy(block + 8, block - 16, sizeof(*block) * 8);
62 memcpy(block + 16, block - 24, sizeof(*block) * 8);
63 memcpy(block + 24, block - 32, sizeof(*block) * 8);
66 static av_always_inline void dnxhd_10bit_get_pixels_8x4_sym(int16_t *restrict block, const uint8_t *pixels, int line_size)
72 for (i = 0; i < 4; i++) {
73 memcpy(block + i * 8, pixels + i * line_size, 8 * sizeof(*block));
74 memcpy(block - (i+1) * 8, pixels + i * line_size, 8 * sizeof(*block));
78 static int dnxhd_10bit_dct_quantize(MpegEncContext *ctx, int16_t *block,
79 int n, int qscale, int *overflow)
81 const uint8_t *scantable= ctx->intra_scantable.scantable;
82 const int *qmat = ctx->q_intra_matrix[qscale];
83 int last_non_zero = 0;
88 // Divide by 4 with rounding, to compensate scaling of DCT coefficients
89 block[0] = (block[0] + 2) >> 2;
91 for (i = 1; i < 64; ++i) {
93 int sign = block[j] >> 31;
94 int level = (block[j] ^ sign) - sign;
95 level = level * qmat[j] >> DNX10BIT_QMAT_SHIFT;
96 block[j] = (level ^ sign) - sign;
101 return last_non_zero;
104 static av_cold int dnxhd_init_vlc(DNXHDEncContext *ctx)
106 int i, j, level, run;
107 int max_level = 1<<(ctx->cid_table->bit_depth+2);
109 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->vlc_codes, max_level*4*sizeof(*ctx->vlc_codes), fail);
110 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->vlc_bits, max_level*4*sizeof(*ctx->vlc_bits) , fail);
111 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_codes, 63*2, fail);
112 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_bits, 63, fail);
114 ctx->vlc_codes += max_level*2;
115 ctx->vlc_bits += max_level*2;
116 for (level = -max_level; level < max_level; level++) {
117 for (run = 0; run < 2; run++) {
118 int index = (level<<1)|run;
119 int sign, offset = 0, alevel = level;
121 MASK_ABS(sign, alevel);
123 offset = (alevel-1)>>6;
126 for (j = 0; j < 257; j++) {
127 if (ctx->cid_table->ac_level[j] == alevel &&
128 (!offset || (ctx->cid_table->ac_index_flag[j] && offset)) &&
129 (!run || (ctx->cid_table->ac_run_flag [j] && run))) {
130 assert(!ctx->vlc_codes[index]);
132 ctx->vlc_codes[index] = (ctx->cid_table->ac_codes[j]<<1)|(sign&1);
133 ctx->vlc_bits [index] = ctx->cid_table->ac_bits[j]+1;
135 ctx->vlc_codes[index] = ctx->cid_table->ac_codes[j];
136 ctx->vlc_bits [index] = ctx->cid_table->ac_bits [j];
141 assert(!alevel || j < 257);
143 ctx->vlc_codes[index] = (ctx->vlc_codes[index]<<ctx->cid_table->index_bits)|offset;
144 ctx->vlc_bits [index]+= ctx->cid_table->index_bits;
148 for (i = 0; i < 62; i++) {
149 int run = ctx->cid_table->run[i];
151 ctx->run_codes[run] = ctx->cid_table->run_codes[i];
152 ctx->run_bits [run] = ctx->cid_table->run_bits[i];
159 static av_cold int dnxhd_init_qmat(DNXHDEncContext *ctx, int lbias, int cbias)
161 // init first elem to 1 to avoid div by 0 in convert_matrix
162 uint16_t weight_matrix[64] = {1,}; // convert_matrix needs uint16_t*
164 const uint8_t *luma_weight_table = ctx->cid_table->luma_weight;
165 const uint8_t *chroma_weight_table = ctx->cid_table->chroma_weight;
167 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l, (ctx->m.avctx->qmax+1) * 64 * sizeof(int), fail);
168 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c, (ctx->m.avctx->qmax+1) * 64 * sizeof(int), fail);
169 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l16, (ctx->m.avctx->qmax+1) * 64 * 2 * sizeof(uint16_t), fail);
170 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c16, (ctx->m.avctx->qmax+1) * 64 * 2 * sizeof(uint16_t), fail);
172 if (ctx->cid_table->bit_depth == 8) {
173 for (i = 1; i < 64; i++) {
174 int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]];
175 weight_matrix[j] = ctx->cid_table->luma_weight[i];
177 ff_convert_matrix(&ctx->m.dsp, ctx->qmatrix_l, ctx->qmatrix_l16, weight_matrix,
178 ctx->m.intra_quant_bias, 1, ctx->m.avctx->qmax, 1);
179 for (i = 1; i < 64; i++) {
180 int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]];
181 weight_matrix[j] = ctx->cid_table->chroma_weight[i];
183 ff_convert_matrix(&ctx->m.dsp, ctx->qmatrix_c, ctx->qmatrix_c16, weight_matrix,
184 ctx->m.intra_quant_bias, 1, ctx->m.avctx->qmax, 1);
186 for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
187 for (i = 0; i < 64; i++) {
188 ctx->qmatrix_l [qscale] [i] <<= 2; ctx->qmatrix_c [qscale] [i] <<= 2;
189 ctx->qmatrix_l16[qscale][0][i] <<= 2; ctx->qmatrix_l16[qscale][1][i] <<= 2;
190 ctx->qmatrix_c16[qscale][0][i] <<= 2; ctx->qmatrix_c16[qscale][1][i] <<= 2;
195 for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
196 for (i = 1; i < 64; i++) {
197 int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]];
199 // The quantization formula from the VC-3 standard is:
200 // quantized = sign(block[i]) * floor(abs(block[i]/s) * p / (qscale * weight_table[i]))
201 // Where p is 32 for 8-bit samples and 8 for 10-bit ones.
202 // The s factor compensates scaling of DCT coefficients done by the DCT routines,
203 // and therefore is not present in standard. It's 8 for 8-bit samples and 4 for 10-bit ones.
204 // We want values of ctx->qtmatrix_l and ctx->qtmatrix_r to be:
205 // ((1 << DNX10BIT_QMAT_SHIFT) * (p / s)) / (qscale * weight_table[i])
206 // For 10-bit samples, p / s == 2
207 ctx->qmatrix_l[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) / (qscale * luma_weight_table[i]);
208 ctx->qmatrix_c[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) / (qscale * chroma_weight_table[i]);
218 static av_cold int dnxhd_init_rc(DNXHDEncContext *ctx)
220 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_rc, 8160*ctx->m.avctx->qmax*sizeof(RCEntry), fail);
221 if (ctx->m.avctx->mb_decision != FF_MB_DECISION_RD)
222 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_cmp, ctx->m.mb_num*sizeof(RCCMPEntry), fail);
224 ctx->frame_bits = (ctx->cid_table->coding_unit_size - 640 - 4 - ctx->min_padding) * 8;
226 ctx->lambda = 2<<LAMBDA_FRAC_BITS; // qscale 2
232 static av_cold int dnxhd_encode_init(AVCodecContext *avctx)
234 DNXHDEncContext *ctx = avctx->priv_data;
235 int i, index, bit_depth;
237 switch (avctx->pix_fmt) {
238 case AV_PIX_FMT_YUV422P:
241 case AV_PIX_FMT_YUV422P10:
245 av_log(avctx, AV_LOG_ERROR, "pixel format is incompatible with DNxHD\n");
249 ctx->cid = ff_dnxhd_find_cid(avctx, bit_depth);
251 av_log(avctx, AV_LOG_ERROR, "video parameters incompatible with DNxHD\n");
254 av_log(avctx, AV_LOG_DEBUG, "cid %d\n", ctx->cid);
256 index = ff_dnxhd_get_cid_table(ctx->cid);
257 ctx->cid_table = &ff_dnxhd_cid_table[index];
259 ctx->m.avctx = avctx;
263 avctx->bits_per_raw_sample = ctx->cid_table->bit_depth;
265 ff_dsputil_init(&ctx->m.dsp, avctx);
266 ff_dct_common_init(&ctx->m);
267 if (!ctx->m.dct_quantize)
268 ctx->m.dct_quantize = ff_dct_quantize_c;
270 if (ctx->cid_table->bit_depth == 10) {
271 ctx->m.dct_quantize = dnxhd_10bit_dct_quantize;
272 ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym;
273 ctx->block_width_l2 = 4;
275 ctx->get_pixels_8x4_sym = dnxhd_8bit_get_pixels_8x4_sym;
276 ctx->block_width_l2 = 3;
280 ff_dnxhdenc_init_x86(ctx);
282 ctx->m.mb_height = (avctx->height + 15) / 16;
283 ctx->m.mb_width = (avctx->width + 15) / 16;
285 if (avctx->flags & CODEC_FLAG_INTERLACED_DCT) {
287 ctx->m.mb_height /= 2;
290 ctx->m.mb_num = ctx->m.mb_height * ctx->m.mb_width;
292 if (avctx->intra_quant_bias != FF_DEFAULT_QUANT_BIAS)
293 ctx->m.intra_quant_bias = avctx->intra_quant_bias;
294 if (dnxhd_init_qmat(ctx, ctx->m.intra_quant_bias, 0) < 0) // XXX tune lbias/cbias
297 // Avid Nitris hardware decoder requires a minimum amount of padding in the coding unit payload
298 if (ctx->nitris_compat)
299 ctx->min_padding = 1600;
301 if (dnxhd_init_vlc(ctx) < 0)
303 if (dnxhd_init_rc(ctx) < 0)
306 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_size, ctx->m.mb_height*sizeof(uint32_t), fail);
307 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_offs, ctx->m.mb_height*sizeof(uint32_t), fail);
308 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_bits, ctx->m.mb_num *sizeof(uint16_t), fail);
309 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_qscale, ctx->m.mb_num *sizeof(uint8_t), fail);
311 avctx->coded_frame = av_frame_alloc();
312 if (!avctx->coded_frame)
313 return AVERROR(ENOMEM);
315 avctx->coded_frame->key_frame = 1;
316 avctx->coded_frame->pict_type = AV_PICTURE_TYPE_I;
318 if (avctx->thread_count > MAX_THREADS) {
319 av_log(avctx, AV_LOG_ERROR, "too many threads\n");
323 ctx->thread[0] = ctx;
324 for (i = 1; i < avctx->thread_count; i++) {
325 ctx->thread[i] = av_malloc(sizeof(DNXHDEncContext));
326 memcpy(ctx->thread[i], ctx, sizeof(DNXHDEncContext));
330 fail: //for FF_ALLOCZ_OR_GOTO
334 static int dnxhd_write_header(AVCodecContext *avctx, uint8_t *buf)
336 DNXHDEncContext *ctx = avctx->priv_data;
337 const uint8_t header_prefix[5] = { 0x00,0x00,0x02,0x80,0x01 };
341 memcpy(buf, header_prefix, 5);
342 buf[5] = ctx->interlaced ? ctx->cur_field+2 : 0x01;
343 buf[6] = 0x80; // crc flag off
344 buf[7] = 0xa0; // reserved
345 AV_WB16(buf + 0x18, avctx->height>>ctx->interlaced); // ALPF
346 AV_WB16(buf + 0x1a, avctx->width); // SPL
347 AV_WB16(buf + 0x1d, avctx->height>>ctx->interlaced); // NAL
349 buf[0x21] = ctx->cid_table->bit_depth == 10 ? 0x58 : 0x38;
350 buf[0x22] = 0x88 + (ctx->interlaced<<2);
351 AV_WB32(buf + 0x28, ctx->cid); // CID
352 buf[0x2c] = ctx->interlaced ? 0 : 0x80;
354 buf[0x5f] = 0x01; // UDL
356 buf[0x167] = 0x02; // reserved
357 AV_WB16(buf + 0x16a, ctx->m.mb_height * 4 + 4); // MSIPS
358 buf[0x16d] = ctx->m.mb_height; // Ns
359 buf[0x16f] = 0x10; // reserved
361 ctx->msip = buf + 0x170;
365 static av_always_inline void dnxhd_encode_dc(DNXHDEncContext *ctx, int diff)
369 nbits = av_log2_16bit(-2*diff);
372 nbits = av_log2_16bit(2*diff);
374 put_bits(&ctx->m.pb, ctx->cid_table->dc_bits[nbits] + nbits,
375 (ctx->cid_table->dc_codes[nbits]<<nbits) + (diff & ((1 << nbits) - 1)));
378 static av_always_inline void dnxhd_encode_block(DNXHDEncContext *ctx, int16_t *block, int last_index, int n)
380 int last_non_zero = 0;
383 dnxhd_encode_dc(ctx, block[0] - ctx->m.last_dc[n]);
384 ctx->m.last_dc[n] = block[0];
386 for (i = 1; i <= last_index; i++) {
387 j = ctx->m.intra_scantable.permutated[i];
390 int run_level = i - last_non_zero - 1;
391 int rlevel = (slevel<<1)|!!run_level;
392 put_bits(&ctx->m.pb, ctx->vlc_bits[rlevel], ctx->vlc_codes[rlevel]);
394 put_bits(&ctx->m.pb, ctx->run_bits[run_level], ctx->run_codes[run_level]);
398 put_bits(&ctx->m.pb, ctx->vlc_bits[0], ctx->vlc_codes[0]); // EOB
401 static av_always_inline void dnxhd_unquantize_c(DNXHDEncContext *ctx, int16_t *block, int n, int qscale, int last_index)
403 const uint8_t *weight_matrix;
407 weight_matrix = (n&2) ? ctx->cid_table->chroma_weight : ctx->cid_table->luma_weight;
409 for (i = 1; i <= last_index; i++) {
410 int j = ctx->m.intra_scantable.permutated[i];
414 level = (1-2*level) * qscale * weight_matrix[i];
415 if (ctx->cid_table->bit_depth == 10) {
416 if (weight_matrix[i] != 8)
420 if (weight_matrix[i] != 32)
426 level = (2*level+1) * qscale * weight_matrix[i];
427 if (ctx->cid_table->bit_depth == 10) {
428 if (weight_matrix[i] != 8)
432 if (weight_matrix[i] != 32)
442 static av_always_inline int dnxhd_ssd_block(int16_t *qblock, int16_t *block)
446 for (i = 0; i < 64; i++)
447 score += (block[i] - qblock[i]) * (block[i] - qblock[i]);
451 static av_always_inline int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, int16_t *block, int last_index)
453 int last_non_zero = 0;
456 for (i = 1; i <= last_index; i++) {
457 j = ctx->m.intra_scantable.permutated[i];
460 int run_level = i - last_non_zero - 1;
461 bits += ctx->vlc_bits[(level<<1)|!!run_level]+ctx->run_bits[run_level];
468 static av_always_inline void dnxhd_get_blocks(DNXHDEncContext *ctx, int mb_x, int mb_y)
470 const int bs = ctx->block_width_l2;
471 const int bw = 1 << bs;
472 const uint8_t *ptr_y = ctx->thread[0]->src[0] + ((mb_y << 4) * ctx->m.linesize) + (mb_x << bs+1);
473 const uint8_t *ptr_u = ctx->thread[0]->src[1] + ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs);
474 const uint8_t *ptr_v = ctx->thread[0]->src[2] + ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs);
475 DSPContext *dsp = &ctx->m.dsp;
477 dsp->get_pixels(ctx->blocks[0], ptr_y, ctx->m.linesize);
478 dsp->get_pixels(ctx->blocks[1], ptr_y + bw, ctx->m.linesize);
479 dsp->get_pixels(ctx->blocks[2], ptr_u, ctx->m.uvlinesize);
480 dsp->get_pixels(ctx->blocks[3], ptr_v, ctx->m.uvlinesize);
482 if (mb_y+1 == ctx->m.mb_height && ctx->m.avctx->height == 1080) {
483 if (ctx->interlaced) {
484 ctx->get_pixels_8x4_sym(ctx->blocks[4], ptr_y + ctx->dct_y_offset, ctx->m.linesize);
485 ctx->get_pixels_8x4_sym(ctx->blocks[5], ptr_y + ctx->dct_y_offset + bw, ctx->m.linesize);
486 ctx->get_pixels_8x4_sym(ctx->blocks[6], ptr_u + ctx->dct_uv_offset, ctx->m.uvlinesize);
487 ctx->get_pixels_8x4_sym(ctx->blocks[7], ptr_v + ctx->dct_uv_offset, ctx->m.uvlinesize);
489 dsp->clear_block(ctx->blocks[4]);
490 dsp->clear_block(ctx->blocks[5]);
491 dsp->clear_block(ctx->blocks[6]);
492 dsp->clear_block(ctx->blocks[7]);
495 dsp->get_pixels(ctx->blocks[4], ptr_y + ctx->dct_y_offset, ctx->m.linesize);
496 dsp->get_pixels(ctx->blocks[5], ptr_y + ctx->dct_y_offset + bw, ctx->m.linesize);
497 dsp->get_pixels(ctx->blocks[6], ptr_u + ctx->dct_uv_offset, ctx->m.uvlinesize);
498 dsp->get_pixels(ctx->blocks[7], ptr_v + ctx->dct_uv_offset, ctx->m.uvlinesize);
502 static av_always_inline int dnxhd_switch_matrix(DNXHDEncContext *ctx, int i)
505 ctx->m.q_intra_matrix16 = ctx->qmatrix_c16;
506 ctx->m.q_intra_matrix = ctx->qmatrix_c;
509 ctx->m.q_intra_matrix16 = ctx->qmatrix_l16;
510 ctx->m.q_intra_matrix = ctx->qmatrix_l;
515 static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg, int jobnr, int threadnr)
517 DNXHDEncContext *ctx = avctx->priv_data;
518 int mb_y = jobnr, mb_x;
519 int qscale = ctx->qscale;
520 LOCAL_ALIGNED_16(int16_t, block, [64]);
521 ctx = ctx->thread[threadnr];
525 ctx->m.last_dc[2] = 1 << (ctx->cid_table->bit_depth + 2);
527 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
528 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
534 dnxhd_get_blocks(ctx, mb_x, mb_y);
536 for (i = 0; i < 8; i++) {
537 int16_t *src_block = ctx->blocks[i];
538 int overflow, nbits, diff, last_index;
539 int n = dnxhd_switch_matrix(ctx, i);
541 memcpy(block, src_block, 64*sizeof(*block));
542 last_index = ctx->m.dct_quantize(&ctx->m, block, i, qscale, &overflow);
543 ac_bits += dnxhd_calc_ac_bits(ctx, block, last_index);
545 diff = block[0] - ctx->m.last_dc[n];
546 if (diff < 0) nbits = av_log2_16bit(-2*diff);
547 else nbits = av_log2_16bit( 2*diff);
549 assert(nbits < ctx->cid_table->bit_depth + 4);
550 dc_bits += ctx->cid_table->dc_bits[nbits] + nbits;
552 ctx->m.last_dc[n] = block[0];
554 if (avctx->mb_decision == FF_MB_DECISION_RD || !RC_VARIANCE) {
555 dnxhd_unquantize_c(ctx, block, i, qscale, last_index);
556 ctx->m.dsp.idct(block);
557 ssd += dnxhd_ssd_block(block, src_block);
560 ctx->mb_rc[qscale][mb].ssd = ssd;
561 ctx->mb_rc[qscale][mb].bits = ac_bits+dc_bits+12+8*ctx->vlc_bits[0];
566 static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg, int jobnr, int threadnr)
568 DNXHDEncContext *ctx = avctx->priv_data;
569 int mb_y = jobnr, mb_x;
570 ctx = ctx->thread[threadnr];
571 init_put_bits(&ctx->m.pb, (uint8_t *)arg + 640 + ctx->slice_offs[jobnr], ctx->slice_size[jobnr]);
575 ctx->m.last_dc[2] = 1 << (ctx->cid_table->bit_depth + 2);
576 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
577 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
578 int qscale = ctx->mb_qscale[mb];
581 put_bits(&ctx->m.pb, 12, qscale<<1);
583 dnxhd_get_blocks(ctx, mb_x, mb_y);
585 for (i = 0; i < 8; i++) {
586 int16_t *block = ctx->blocks[i];
587 int overflow, n = dnxhd_switch_matrix(ctx, i);
588 int last_index = ctx->m.dct_quantize(&ctx->m, block, i,
591 dnxhd_encode_block(ctx, block, last_index, n);
592 //STOP_TIMER("encode_block");
595 if (put_bits_count(&ctx->m.pb)&31)
596 put_bits(&ctx->m.pb, 32-(put_bits_count(&ctx->m.pb)&31), 0);
597 flush_put_bits(&ctx->m.pb);
601 static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx)
605 for (mb_y = 0; mb_y < ctx->m.mb_height; mb_y++) {
607 ctx->slice_offs[mb_y] = offset;
608 ctx->slice_size[mb_y] = 0;
609 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
610 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
611 ctx->slice_size[mb_y] += ctx->mb_bits[mb];
613 ctx->slice_size[mb_y] = (ctx->slice_size[mb_y]+31)&~31;
614 ctx->slice_size[mb_y] >>= 3;
615 thread_size = ctx->slice_size[mb_y];
616 offset += thread_size;
620 static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg, int jobnr, int threadnr)
622 DNXHDEncContext *ctx = avctx->priv_data;
623 int mb_y = jobnr, mb_x, x, y;
624 int partial_last_row = (mb_y == ctx->m.mb_height - 1) &&
625 ((avctx->height >> ctx->interlaced) & 0xF);
627 ctx = ctx->thread[threadnr];
628 if (ctx->cid_table->bit_depth == 8) {
629 uint8_t *pix = ctx->thread[0]->src[0] + ((mb_y<<4) * ctx->m.linesize);
630 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x, pix += 16) {
631 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
635 if (!partial_last_row && mb_x * 16 <= avctx->width - 16) {
636 sum = ctx->m.dsp.pix_sum(pix, ctx->m.linesize);
637 varc = ctx->m.dsp.pix_norm1(pix, ctx->m.linesize);
639 int bw = FFMIN(avctx->width - 16 * mb_x, 16);
640 int bh = FFMIN((avctx->height >> ctx->interlaced) - 16 * mb_y, 16);
642 for (y = 0; y < bh; y++) {
643 for (x = 0; x < bw; x++) {
644 uint8_t val = pix[x + y * ctx->m.linesize];
650 varc = (varc - (((unsigned)sum * sum) >> 8) + 128) >> 8;
652 ctx->mb_cmp[mb].value = varc;
653 ctx->mb_cmp[mb].mb = mb;
656 int const linesize = ctx->m.linesize >> 1;
657 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x) {
658 uint16_t *pix = (uint16_t*)ctx->thread[0]->src[0] + ((mb_y << 4) * linesize) + (mb_x << 4);
659 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
664 // Macroblocks are 16x16 pixels, unlike DCT blocks which are 8x8.
665 for (i = 0; i < 16; ++i) {
666 for (j = 0; j < 16; ++j) {
667 // Turn 16-bit pixels into 10-bit ones.
668 int const sample = (unsigned)pix[j] >> 6;
670 sqsum += sample * sample;
671 // 2^10 * 2^10 * 16 * 16 = 2^28, which is less than INT_MAX
675 mean = sum >> 8; // 16*16 == 2^8
677 ctx->mb_cmp[mb].value = sqmean - mean * mean;
678 ctx->mb_cmp[mb].mb = mb;
684 static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx)
686 int lambda, up_step, down_step;
687 int last_lower = INT_MAX, last_higher = 0;
690 for (q = 1; q < avctx->qmax; q++) {
692 avctx->execute2(avctx, dnxhd_calc_bits_thread, NULL, NULL, ctx->m.mb_height);
694 up_step = down_step = 2<<LAMBDA_FRAC_BITS;
695 lambda = ctx->lambda;
700 if (lambda == last_higher) {
702 end = 1; // need to set final qscales/bits
704 for (y = 0; y < ctx->m.mb_height; y++) {
705 for (x = 0; x < ctx->m.mb_width; x++) {
706 unsigned min = UINT_MAX;
708 int mb = y*ctx->m.mb_width+x;
709 for (q = 1; q < avctx->qmax; q++) {
710 unsigned score = ctx->mb_rc[q][mb].bits*lambda+
711 ((unsigned)ctx->mb_rc[q][mb].ssd<<LAMBDA_FRAC_BITS);
717 bits += ctx->mb_rc[qscale][mb].bits;
718 ctx->mb_qscale[mb] = qscale;
719 ctx->mb_bits[mb] = ctx->mb_rc[qscale][mb].bits;
721 bits = (bits+31)&~31; // padding
722 if (bits > ctx->frame_bits)
725 //av_dlog(ctx->m.avctx, "lambda %d, up %u, down %u, bits %d, frame %d\n",
726 // lambda, last_higher, last_lower, bits, ctx->frame_bits);
728 if (bits > ctx->frame_bits)
732 if (bits < ctx->frame_bits) {
733 last_lower = FFMIN(lambda, last_lower);
734 if (last_higher != 0)
735 lambda = (lambda+last_higher)>>1;
738 down_step = FFMIN((int64_t)down_step*5, INT_MAX);
739 up_step = 1<<LAMBDA_FRAC_BITS;
740 lambda = FFMAX(1, lambda);
741 if (lambda == last_lower)
744 last_higher = FFMAX(lambda, last_higher);
745 if (last_lower != INT_MAX)
746 lambda = (lambda+last_lower)>>1;
747 else if ((int64_t)lambda + up_step > INT_MAX)
751 up_step = FFMIN((int64_t)up_step*5, INT_MAX);
752 down_step = 1<<LAMBDA_FRAC_BITS;
755 //av_dlog(ctx->m.avctx, "out lambda %d\n", lambda);
756 ctx->lambda = lambda;
760 static int dnxhd_find_qscale(DNXHDEncContext *ctx)
766 int last_lower = INT_MAX;
770 qscale = ctx->qscale;
773 ctx->qscale = qscale;
774 // XXX avoid recalculating bits
775 ctx->m.avctx->execute2(ctx->m.avctx, dnxhd_calc_bits_thread, NULL, NULL, ctx->m.mb_height);
776 for (y = 0; y < ctx->m.mb_height; y++) {
777 for (x = 0; x < ctx->m.mb_width; x++)
778 bits += ctx->mb_rc[qscale][y*ctx->m.mb_width+x].bits;
779 bits = (bits+31)&~31; // padding
780 if (bits > ctx->frame_bits)
783 //av_dlog(ctx->m.avctx, "%d, qscale %d, bits %d, frame %d, higher %d, lower %d\n",
784 // ctx->m.avctx->frame_number, qscale, bits, ctx->frame_bits, last_higher, last_lower);
785 if (bits < ctx->frame_bits) {
788 if (last_higher == qscale - 1) {
789 qscale = last_higher;
792 last_lower = FFMIN(qscale, last_lower);
793 if (last_higher != 0)
794 qscale = (qscale+last_higher)>>1;
796 qscale -= down_step++;
801 if (last_lower == qscale + 1)
803 last_higher = FFMAX(qscale, last_higher);
804 if (last_lower != INT_MAX)
805 qscale = (qscale+last_lower)>>1;
809 if (qscale >= ctx->m.avctx->qmax)
813 //av_dlog(ctx->m.avctx, "out qscale %d\n", qscale);
814 ctx->qscale = qscale;
818 #define BUCKET_BITS 8
819 #define RADIX_PASSES 4
820 #define NBUCKETS (1 << BUCKET_BITS)
822 static inline int get_bucket(int value, int shift)
825 value &= NBUCKETS - 1;
826 return NBUCKETS - 1 - value;
829 static void radix_count(const RCCMPEntry *data, int size, int buckets[RADIX_PASSES][NBUCKETS])
832 memset(buckets, 0, sizeof(buckets[0][0]) * RADIX_PASSES * NBUCKETS);
833 for (i = 0; i < size; i++) {
834 int v = data[i].value;
835 for (j = 0; j < RADIX_PASSES; j++) {
836 buckets[j][get_bucket(v, 0)]++;
841 for (j = 0; j < RADIX_PASSES; j++) {
843 for (i = NBUCKETS - 1; i >= 0; i--)
844 buckets[j][i] = offset -= buckets[j][i];
845 assert(!buckets[j][0]);
849 static void radix_sort_pass(RCCMPEntry *dst, const RCCMPEntry *data, int size, int buckets[NBUCKETS], int pass)
851 int shift = pass * BUCKET_BITS;
853 for (i = 0; i < size; i++) {
854 int v = get_bucket(data[i].value, shift);
855 int pos = buckets[v]++;
860 static void radix_sort(RCCMPEntry *data, int size)
862 int buckets[RADIX_PASSES][NBUCKETS];
863 RCCMPEntry *tmp = av_malloc(sizeof(*tmp) * size);
864 radix_count(data, size, buckets);
865 radix_sort_pass(tmp, data, size, buckets[0], 0);
866 radix_sort_pass(data, tmp, size, buckets[1], 1);
867 if (buckets[2][NBUCKETS - 1] || buckets[3][NBUCKETS - 1]) {
868 radix_sort_pass(tmp, data, size, buckets[2], 2);
869 radix_sort_pass(data, tmp, size, buckets[3], 3);
874 static int dnxhd_encode_fast(AVCodecContext *avctx, DNXHDEncContext *ctx)
878 if ((ret = dnxhd_find_qscale(ctx)) < 0)
880 for (y = 0; y < ctx->m.mb_height; y++) {
881 for (x = 0; x < ctx->m.mb_width; x++) {
882 int mb = y*ctx->m.mb_width+x;
884 ctx->mb_qscale[mb] = ctx->qscale;
885 ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale][mb].bits;
886 max_bits += ctx->mb_rc[ctx->qscale][mb].bits;
888 delta_bits = ctx->mb_rc[ctx->qscale][mb].bits-ctx->mb_rc[ctx->qscale+1][mb].bits;
889 ctx->mb_cmp[mb].mb = mb;
890 ctx->mb_cmp[mb].value = delta_bits ?
891 ((ctx->mb_rc[ctx->qscale][mb].ssd-ctx->mb_rc[ctx->qscale+1][mb].ssd)*100)/delta_bits
892 : INT_MIN; //avoid increasing qscale
895 max_bits += 31; //worst padding
899 avctx->execute2(avctx, dnxhd_mb_var_thread, NULL, NULL, ctx->m.mb_height);
900 radix_sort(ctx->mb_cmp, ctx->m.mb_num);
901 for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) {
902 int mb = ctx->mb_cmp[x].mb;
903 max_bits -= ctx->mb_rc[ctx->qscale][mb].bits - ctx->mb_rc[ctx->qscale+1][mb].bits;
904 ctx->mb_qscale[mb] = ctx->qscale+1;
905 ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale+1][mb].bits;
911 static void dnxhd_load_picture(DNXHDEncContext *ctx, const AVFrame *frame)
915 for (i = 0; i < ctx->m.avctx->thread_count; i++) {
916 ctx->thread[i]->m.linesize = frame->linesize[0] << ctx->interlaced;
917 ctx->thread[i]->m.uvlinesize = frame->linesize[1] << ctx->interlaced;
918 ctx->thread[i]->dct_y_offset = ctx->m.linesize *8;
919 ctx->thread[i]->dct_uv_offset = ctx->m.uvlinesize*8;
922 ctx->m.avctx->coded_frame->interlaced_frame = frame->interlaced_frame;
923 ctx->cur_field = frame->interlaced_frame && !frame->top_field_first;
926 static int dnxhd_encode_picture(AVCodecContext *avctx, AVPacket *pkt,
927 const AVFrame *frame, int *got_packet)
929 DNXHDEncContext *ctx = avctx->priv_data;
934 if ((ret = ff_alloc_packet(pkt, ctx->cid_table->frame_size)) < 0) {
935 av_log(avctx, AV_LOG_ERROR, "output buffer is too small to compress picture\n");
940 dnxhd_load_picture(ctx, frame);
943 for (i = 0; i < 3; i++) {
944 ctx->src[i] = frame->data[i];
945 if (ctx->interlaced && ctx->cur_field)
946 ctx->src[i] += frame->linesize[i];
949 dnxhd_write_header(avctx, buf);
951 if (avctx->mb_decision == FF_MB_DECISION_RD)
952 ret = dnxhd_encode_rdo(avctx, ctx);
954 ret = dnxhd_encode_fast(avctx, ctx);
956 av_log(avctx, AV_LOG_ERROR,
957 "picture could not fit ratecontrol constraints, increase qmax\n");
961 dnxhd_setup_threads_slices(ctx);
964 for (i = 0; i < ctx->m.mb_height; i++) {
965 AV_WB32(ctx->msip + i * 4, offset);
966 offset += ctx->slice_size[i];
967 assert(!(ctx->slice_size[i] & 3));
970 avctx->execute2(avctx, dnxhd_encode_thread, buf, NULL, ctx->m.mb_height);
972 assert(640 + offset + 4 <= ctx->cid_table->coding_unit_size);
973 memset(buf + 640 + offset, 0, ctx->cid_table->coding_unit_size - 4 - offset - 640);
975 AV_WB32(buf + ctx->cid_table->coding_unit_size - 4, 0x600DC0DE); // EOF
977 if (ctx->interlaced && first_field) {
980 buf += ctx->cid_table->coding_unit_size;
981 goto encode_coding_unit;
984 avctx->coded_frame->quality = ctx->qscale * FF_QP2LAMBDA;
986 pkt->flags |= AV_PKT_FLAG_KEY;
991 static av_cold int dnxhd_encode_end(AVCodecContext *avctx)
993 DNXHDEncContext *ctx = avctx->priv_data;
994 int max_level = 1<<(ctx->cid_table->bit_depth+2);
997 av_free(ctx->vlc_codes-max_level*2);
998 av_free(ctx->vlc_bits -max_level*2);
999 av_freep(&ctx->run_codes);
1000 av_freep(&ctx->run_bits);
1002 av_freep(&ctx->mb_bits);
1003 av_freep(&ctx->mb_qscale);
1004 av_freep(&ctx->mb_rc);
1005 av_freep(&ctx->mb_cmp);
1006 av_freep(&ctx->slice_size);
1007 av_freep(&ctx->slice_offs);
1009 av_freep(&ctx->qmatrix_c);
1010 av_freep(&ctx->qmatrix_l);
1011 av_freep(&ctx->qmatrix_c16);
1012 av_freep(&ctx->qmatrix_l16);
1014 for (i = 1; i < avctx->thread_count; i++)
1015 av_freep(&ctx->thread[i]);
1017 av_frame_free(&avctx->coded_frame);
1022 AVCodec ff_dnxhd_encoder = {
1024 .long_name = NULL_IF_CONFIG_SMALL("VC3/DNxHD"),
1025 .type = AVMEDIA_TYPE_VIDEO,
1026 .id = AV_CODEC_ID_DNXHD,
1027 .priv_data_size = sizeof(DNXHDEncContext),
1028 .init = dnxhd_encode_init,
1029 .encode2 = dnxhd_encode_picture,
1030 .close = dnxhd_encode_end,
1031 .capabilities = CODEC_CAP_SLICE_THREADS,
1032 .pix_fmts = (const enum AVPixelFormat[]){ AV_PIX_FMT_YUV422P,
1033 AV_PIX_FMT_YUV422P10,
1035 .priv_class = &class,