4 * Copyright (C) 2007 Vitor Sessak <vitor1001@gmail.com>
5 * Copyright (C) 2004-2007 Eric Lasota
6 * Based on RoQ specs (C) 2001 Tim Ferguson
8 * This file is part of Libav.
10 * Libav is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Lesser General Public
12 * License as published by the Free Software Foundation; either
13 * version 2.1 of the License, or (at your option) any later version.
15 * Libav is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * Lesser General Public License for more details.
20 * You should have received a copy of the GNU Lesser General Public
21 * License along with Libav; if not, write to the Free Software
22 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
27 * id RoQ encoder by Vitor. Based on the Switchblade3 library and the
28 * Switchblade3 Libav glue by Eric Lasota.
36 * CODEBOOK - 2 + 8 bits
37 * SUBDIVIDE - 2 + combined subcel cost
42 * CODEBOOK - 2 + 8 bits
43 * SUBDIVIDE - 2 + 4*8 bits
45 * Maximum cost: 138 bits per cel
47 * Proper evaluation requires LCD fraction comparison, which requires
48 * Squared Error (SE) loss * savings increase
50 * Maximum savings increase: 136 bits
51 * Maximum SE loss without overflow: 31580641
52 * Components in 8x8 supercel: 192
53 * Maximum SE precision per component: 164482
54 * >65025, so no truncation is needed (phew)
59 #include "libavutil/attributes.h"
61 #include "bytestream.h"
69 * Maximum number of generated 4x4 codebooks. Can't be 256 to workaround a
72 #define MAX_CBS_4x4 255
74 #define MAX_CBS_2x2 256 ///< Maximum number of 2x2 codebooks.
76 /* The cast is useful when multiplying it by INT_MAX */
77 #define ROQ_LAMBDA_SCALE ((uint64_t) FF_LAMBDA_SCALE)
79 /* Macroblock support functions */
80 static void unpack_roq_cell(roq_cell *cell, uint8_t u[4*3])
82 memcpy(u , cell->y, 4);
83 memset(u+4, cell->u, 4);
84 memset(u+8, cell->v, 4);
87 static void unpack_roq_qcell(uint8_t cb2[], roq_qcell *qcell, uint8_t u[4*4*3])
90 static const int offsets[4] = {0, 2, 8, 10};
92 for (cp=0; cp<3; cp++)
94 u[4*4*cp + offsets[i] ] = cb2[qcell->idx[i]*2*2*3 + 4*cp ];
95 u[4*4*cp + offsets[i]+1] = cb2[qcell->idx[i]*2*2*3 + 4*cp+1];
96 u[4*4*cp + offsets[i]+4] = cb2[qcell->idx[i]*2*2*3 + 4*cp+2];
97 u[4*4*cp + offsets[i]+5] = cb2[qcell->idx[i]*2*2*3 + 4*cp+3];
102 static void enlarge_roq_mb4(uint8_t base[3*16], uint8_t u[3*64])
106 for(cp=0; cp<3; cp++)
109 *u++ = base[(y/2)*4 + (x/2) + 16*cp];
112 static inline int square(int x)
117 static inline int eval_sse(const uint8_t *a, const uint8_t *b, int count)
122 diff += square(*b++ - *a++);
127 // FIXME Could use DSPContext.sse, but it is not so speed critical (used
128 // just for motion estimation).
129 static int block_sse(uint8_t * const *buf1, uint8_t * const *buf2, int x1, int y1,
130 int x2, int y2, const int *stride1, const int *stride2, int size)
135 for (k=0; k<3; k++) {
136 int bias = (k ? CHROMA_BIAS : 4);
137 for (i=0; i<size; i++)
138 sse += bias*eval_sse(buf1[k] + (y1+i)*stride1[k] + x1,
139 buf2[k] + (y2+i)*stride2[k] + x2, size);
145 static int eval_motion_dist(RoqContext *enc, int x, int y, motion_vect vect,
151 if (mx < -7 || mx > 7)
154 if (my < -7 || my > 7)
160 if ((unsigned) mx > enc->width-size || (unsigned) my > enc->height-size)
163 return block_sse(enc->frame_to_enc->data, enc->last_frame->data, x, y,
165 enc->frame_to_enc->linesize, enc->last_frame->linesize,
170 * @return distortion between two macroblocks
172 static inline int squared_diff_macroblock(uint8_t a[], uint8_t b[], int size)
176 for(cp=0;cp<3;cp++) {
177 int bias = (cp ? CHROMA_BIAS : 4);
178 sdiff += bias*eval_sse(a, b, size*size);
202 SubcelEvaluation subCels[4];
207 int sourceX, sourceY;
214 int usedCB2[MAX_CBS_2x2];
215 int usedCB4[MAX_CBS_4x4];
216 uint8_t unpacked_cb2[MAX_CBS_2x2*2*2*3];
217 uint8_t unpacked_cb4[MAX_CBS_4x4*4*4*3];
218 uint8_t unpacked_cb4_enlarged[MAX_CBS_4x4*8*8*3];
224 typedef struct RoqTempData
226 CelEvaluation *cel_evals;
228 int f2i4[MAX_CBS_4x4];
229 int i2f4[MAX_CBS_4x4];
230 int f2i2[MAX_CBS_2x2];
231 int i2f2[MAX_CBS_2x2];
238 RoqCodebooks codebooks;
245 * Initialize cel evaluators and set their source coordinates
247 static int create_cel_evals(RoqContext *enc, RoqTempdata *tempData)
251 tempData->cel_evals = av_malloc(enc->width*enc->height/64 * sizeof(CelEvaluation));
252 if (!tempData->cel_evals)
253 return AVERROR(ENOMEM);
255 /* Map to the ROQ quadtree order */
256 for (y=0; y<enc->height; y+=16)
257 for (x=0; x<enc->width; x+=16)
259 tempData->cel_evals[n ].sourceX = x + (i&1)*8;
260 tempData->cel_evals[n++].sourceY = y + (i&2)*4;
267 * Get macroblocks from parts of the image
269 static void get_frame_mb(const AVFrame *frame, int x, int y, uint8_t mb[], int dim)
273 for (cp=0; cp<3; cp++) {
274 int stride = frame->linesize[cp];
275 for (i=0; i<dim; i++)
276 for (j=0; j<dim; j++)
277 *mb++ = frame->data[cp][(y+i)*stride + x + j];
282 * Find the codebook with the lowest distortion from an image
284 static int index_mb(uint8_t cluster[], uint8_t cb[], int numCB,
285 int *outIndex, int dim)
287 int i, lDiff = INT_MAX, pick=0;
289 /* Diff against the others */
290 for (i=0; i<numCB; i++) {
291 int diff = squared_diff_macroblock(cluster, cb + i*dim*dim*3, dim);
302 #define EVAL_MOTION(MOTION) \
304 diff = eval_motion_dist(enc, j, i, MOTION, blocksize); \
306 if (diff < lowestdiff) { \
312 static void motion_search(RoqContext *enc, int blocksize)
314 static const motion_vect offsets[8] = {
325 int diff, lowestdiff, oldbest;
327 motion_vect bestpick = {{0,0}};
330 motion_vect *last_motion;
331 motion_vect *this_motion;
332 motion_vect vect, vect2;
334 int max=(enc->width/blocksize)*enc->height/blocksize;
336 if (blocksize == 4) {
337 last_motion = enc->last_motion4;
338 this_motion = enc->this_motion4;
340 last_motion = enc->last_motion8;
341 this_motion = enc->this_motion8;
344 for (i=0; i<enc->height; i+=blocksize)
345 for (j=0; j<enc->width; j+=blocksize) {
346 lowestdiff = eval_motion_dist(enc, j, i, (motion_vect) {{0,0}},
352 EVAL_MOTION(enc->this_motion8[(i/8)*(enc->width/8) + j/8]);
354 offset = (i/blocksize)*enc->width/blocksize + j/blocksize;
355 if (offset < max && offset >= 0)
356 EVAL_MOTION(last_motion[offset]);
359 if (offset < max && offset >= 0)
360 EVAL_MOTION(last_motion[offset]);
362 offset = (i/blocksize + 1)*enc->width/blocksize + j/blocksize;
363 if (offset < max && offset >= 0)
364 EVAL_MOTION(last_motion[offset]);
366 off[0]= (i/blocksize)*enc->width/blocksize + j/blocksize - 1;
367 off[1]= off[0] - enc->width/blocksize + 1;
373 vect.d[k]= mid_pred(this_motion[off[0]].d[k],
374 this_motion[off[1]].d[k],
375 this_motion[off[2]].d[k]);
379 EVAL_MOTION(this_motion[off[k]]);
381 EVAL_MOTION(this_motion[off[0]]);
386 while (oldbest != lowestdiff) {
387 oldbest = lowestdiff;
388 for (k=0; k<8; k++) {
390 vect2.d[0] += offsets[k].d[0];
391 vect2.d[1] += offsets[k].d[1];
396 offset = (i/blocksize)*enc->width/blocksize + j/blocksize;
397 this_motion[offset] = bestpick;
402 * Get distortion for all options available to a subcel
404 static void gather_data_for_subcel(SubcelEvaluation *subcel, int x,
405 int y, RoqContext *enc, RoqTempdata *tempData)
412 static const int bitsUsed[4] = {2, 10, 10, 34};
414 if (enc->framesSinceKeyframe >= 1) {
415 subcel->motion = enc->this_motion4[y*enc->width/16 + x/4];
417 subcel->eval_dist[RoQ_ID_FCC] =
418 eval_motion_dist(enc, x, y,
419 enc->this_motion4[y*enc->width/16 + x/4], 4);
421 subcel->eval_dist[RoQ_ID_FCC] = INT_MAX;
423 if (enc->framesSinceKeyframe >= 2)
424 subcel->eval_dist[RoQ_ID_MOT] = block_sse(enc->frame_to_enc->data,
425 enc->current_frame->data, x,
427 enc->frame_to_enc->linesize,
428 enc->current_frame->linesize,
431 subcel->eval_dist[RoQ_ID_MOT] = INT_MAX;
433 cluster_index = y*enc->width/16 + x/4;
435 get_frame_mb(enc->frame_to_enc, x, y, mb4, 4);
437 subcel->eval_dist[RoQ_ID_SLD] = index_mb(mb4,
438 tempData->codebooks.unpacked_cb4,
439 tempData->codebooks.numCB4,
440 &subcel->cbEntry, 4);
442 subcel->eval_dist[RoQ_ID_CCC] = 0;
445 subcel->subCels[i] = tempData->closest_cb2[cluster_index*4+i];
447 get_frame_mb(enc->frame_to_enc, x+2*(i&1),
450 subcel->eval_dist[RoQ_ID_CCC] +=
451 squared_diff_macroblock(tempData->codebooks.unpacked_cb2 + subcel->subCels[i]*2*2*3, mb2, 2);
456 if (ROQ_LAMBDA_SCALE*subcel->eval_dist[i] + enc->lambda*bitsUsed[i] <
458 subcel->best_coding = i;
459 subcel->best_bit_use = bitsUsed[i];
460 best_dist = ROQ_LAMBDA_SCALE*subcel->eval_dist[i] +
461 enc->lambda*bitsUsed[i];
466 * Get distortion for all options available to a cel
468 static void gather_data_for_cel(CelEvaluation *cel, RoqContext *enc,
469 RoqTempdata *tempData)
472 int index = cel->sourceY*enc->width/64 + cel->sourceX/8;
473 int i, j, best_dist, divide_bit_use;
475 int bitsUsed[4] = {2, 10, 10, 0};
477 if (enc->framesSinceKeyframe >= 1) {
478 cel->motion = enc->this_motion8[index];
480 cel->eval_dist[RoQ_ID_FCC] =
481 eval_motion_dist(enc, cel->sourceX, cel->sourceY,
482 enc->this_motion8[index], 8);
484 cel->eval_dist[RoQ_ID_FCC] = INT_MAX;
486 if (enc->framesSinceKeyframe >= 2)
487 cel->eval_dist[RoQ_ID_MOT] = block_sse(enc->frame_to_enc->data,
488 enc->current_frame->data,
489 cel->sourceX, cel->sourceY,
490 cel->sourceX, cel->sourceY,
491 enc->frame_to_enc->linesize,
492 enc->current_frame->linesize,8);
494 cel->eval_dist[RoQ_ID_MOT] = INT_MAX;
496 get_frame_mb(enc->frame_to_enc, cel->sourceX, cel->sourceY, mb8, 8);
498 cel->eval_dist[RoQ_ID_SLD] =
499 index_mb(mb8, tempData->codebooks.unpacked_cb4_enlarged,
500 tempData->codebooks.numCB4, &cel->cbEntry, 8);
502 gather_data_for_subcel(cel->subCels + 0, cel->sourceX+0, cel->sourceY+0, enc, tempData);
503 gather_data_for_subcel(cel->subCels + 1, cel->sourceX+4, cel->sourceY+0, enc, tempData);
504 gather_data_for_subcel(cel->subCels + 2, cel->sourceX+0, cel->sourceY+4, enc, tempData);
505 gather_data_for_subcel(cel->subCels + 3, cel->sourceX+4, cel->sourceY+4, enc, tempData);
507 cel->eval_dist[RoQ_ID_CCC] = 0;
509 for (i=0; i<4; i++) {
510 cel->eval_dist[RoQ_ID_CCC] +=
511 cel->subCels[i].eval_dist[cel->subCels[i].best_coding];
512 divide_bit_use += cel->subCels[i].best_bit_use;
516 bitsUsed[3] = 2 + divide_bit_use;
519 if (ROQ_LAMBDA_SCALE*cel->eval_dist[i] + enc->lambda*bitsUsed[i] <
521 cel->best_coding = i;
522 best_dist = ROQ_LAMBDA_SCALE*cel->eval_dist[i] +
523 enc->lambda*bitsUsed[i];
526 tempData->used_option[cel->best_coding]++;
527 tempData->mainChunkSize += bitsUsed[cel->best_coding];
529 if (cel->best_coding == RoQ_ID_SLD)
530 tempData->codebooks.usedCB4[cel->cbEntry]++;
532 if (cel->best_coding == RoQ_ID_CCC)
533 for (i=0; i<4; i++) {
534 if (cel->subCels[i].best_coding == RoQ_ID_SLD)
535 tempData->codebooks.usedCB4[cel->subCels[i].cbEntry]++;
536 else if (cel->subCels[i].best_coding == RoQ_ID_CCC)
538 tempData->codebooks.usedCB2[cel->subCels[i].subCels[j]]++;
542 static void remap_codebooks(RoqContext *enc, RoqTempdata *tempData)
546 /* Make remaps for the final codebook usage */
547 for (i=0; i<MAX_CBS_4x4; i++) {
548 if (tempData->codebooks.usedCB4[i]) {
549 tempData->i2f4[i] = idx;
550 tempData->f2i4[idx] = i;
552 tempData->codebooks.usedCB2[enc->cb4x4[i].idx[j]]++;
557 tempData->numCB4 = idx;
560 for (i=0; i<MAX_CBS_2x2; i++) {
561 if (tempData->codebooks.usedCB2[i]) {
562 tempData->i2f2[i] = idx;
563 tempData->f2i2[idx] = i;
567 tempData->numCB2 = idx;
572 * Write codebook chunk
574 static void write_codebooks(RoqContext *enc, RoqTempdata *tempData)
577 uint8_t **outp= &enc->out_buf;
579 if (tempData->numCB2) {
580 bytestream_put_le16(outp, RoQ_QUAD_CODEBOOK);
581 bytestream_put_le32(outp, tempData->numCB2*6 + tempData->numCB4*4);
582 bytestream_put_byte(outp, tempData->numCB4);
583 bytestream_put_byte(outp, tempData->numCB2);
585 for (i=0; i<tempData->numCB2; i++) {
586 bytestream_put_buffer(outp, enc->cb2x2[tempData->f2i2[i]].y, 4);
587 bytestream_put_byte(outp, enc->cb2x2[tempData->f2i2[i]].u);
588 bytestream_put_byte(outp, enc->cb2x2[tempData->f2i2[i]].v);
591 for (i=0; i<tempData->numCB4; i++)
593 bytestream_put_byte(outp, tempData->i2f2[enc->cb4x4[tempData->f2i4[i]].idx[j]]);
598 static inline uint8_t motion_arg(motion_vect mot)
600 uint8_t ax = 8 - ((uint8_t) mot.d[0]);
601 uint8_t ay = 8 - ((uint8_t) mot.d[1]);
602 return ((ax&15)<<4) | (ay&15);
609 uint8_t argumentSpool[64];
614 /* NOTE: Typecodes must be spooled AFTER arguments!! */
615 static void write_typecode(CodingSpool *s, uint8_t type)
617 s->typeSpool |= (type & 3) << (14 - s->typeSpoolLength);
618 s->typeSpoolLength += 2;
619 if (s->typeSpoolLength == 16) {
620 bytestream_put_le16(s->pout, s->typeSpool);
621 bytestream_put_buffer(s->pout, s->argumentSpool,
622 s->args - s->argumentSpool);
623 s->typeSpoolLength = 0;
625 s->args = s->argumentSpool;
629 static void reconstruct_and_encode_image(RoqContext *enc, RoqTempdata *tempData, int w, int h, int numBlocks)
642 spool.typeSpoolLength=0;
643 spool.args = spool.argumentSpool;
644 spool.pout = &enc->out_buf;
646 if (tempData->used_option[RoQ_ID_CCC]%2)
647 tempData->mainChunkSize+=8; //FIXME
649 /* Write the video chunk header */
650 bytestream_put_le16(&enc->out_buf, RoQ_QUAD_VQ);
651 bytestream_put_le32(&enc->out_buf, tempData->mainChunkSize/8);
652 bytestream_put_byte(&enc->out_buf, 0x0);
653 bytestream_put_byte(&enc->out_buf, 0x0);
655 for (i=0; i<numBlocks; i++) {
656 eval = tempData->cel_evals + i;
660 dist += eval->eval_dist[eval->best_coding];
662 switch (eval->best_coding) {
664 write_typecode(&spool, RoQ_ID_MOT);
668 bytestream_put_byte(&spool.args, motion_arg(eval->motion));
670 write_typecode(&spool, RoQ_ID_FCC);
671 ff_apply_motion_8x8(enc, x, y,
672 eval->motion.d[0], eval->motion.d[1]);
676 bytestream_put_byte(&spool.args, tempData->i2f4[eval->cbEntry]);
677 write_typecode(&spool, RoQ_ID_SLD);
679 qcell = enc->cb4x4 + eval->cbEntry;
680 ff_apply_vector_4x4(enc, x , y , enc->cb2x2 + qcell->idx[0]);
681 ff_apply_vector_4x4(enc, x+4, y , enc->cb2x2 + qcell->idx[1]);
682 ff_apply_vector_4x4(enc, x , y+4, enc->cb2x2 + qcell->idx[2]);
683 ff_apply_vector_4x4(enc, x+4, y+4, enc->cb2x2 + qcell->idx[3]);
687 write_typecode(&spool, RoQ_ID_CCC);
689 for (j=0; j<4; j++) {
693 switch(eval->subCels[j].best_coding) {
698 bytestream_put_byte(&spool.args,
699 motion_arg(eval->subCels[j].motion));
701 ff_apply_motion_4x4(enc, subX, subY,
702 eval->subCels[j].motion.d[0],
703 eval->subCels[j].motion.d[1]);
707 bytestream_put_byte(&spool.args,
708 tempData->i2f4[eval->subCels[j].cbEntry]);
710 qcell = enc->cb4x4 + eval->subCels[j].cbEntry;
712 ff_apply_vector_2x2(enc, subX , subY ,
713 enc->cb2x2 + qcell->idx[0]);
714 ff_apply_vector_2x2(enc, subX+2, subY ,
715 enc->cb2x2 + qcell->idx[1]);
716 ff_apply_vector_2x2(enc, subX , subY+2,
717 enc->cb2x2 + qcell->idx[2]);
718 ff_apply_vector_2x2(enc, subX+2, subY+2,
719 enc->cb2x2 + qcell->idx[3]);
723 for (k=0; k<4; k++) {
724 int cb_idx = eval->subCels[j].subCels[k];
725 bytestream_put_byte(&spool.args,
726 tempData->i2f2[cb_idx]);
728 ff_apply_vector_2x2(enc, subX + 2*(k&1), subY + (k&2),
729 enc->cb2x2 + cb_idx);
733 write_typecode(&spool, eval->subCels[j].best_coding);
739 /* Flush the remainder of the argument/type spool */
740 while (spool.typeSpoolLength)
741 write_typecode(&spool, 0x0);
744 uint8_t *fdata[3] = {enc->frame_to_enc->data[0],
745 enc->frame_to_enc->data[1],
746 enc->frame_to_enc->data[2]};
747 uint8_t *cdata[3] = {enc->current_frame->data[0],
748 enc->current_frame->data[1],
749 enc->current_frame->data[2]};
750 av_log(enc->avctx, AV_LOG_ERROR, "Expected distortion: %i Actual: %i\n",
752 block_sse(fdata, cdata, 0, 0, 0, 0,
753 enc->frame_to_enc->linesize,
754 enc->current_frame->linesize,
755 enc->width)); //WARNING: Square dimensions implied...
761 * Create a single YUV cell from a 2x2 section of the image
763 static inline void frame_block_to_cell(uint8_t *block, uint8_t * const *data,
764 int top, int left, const int *stride)
769 for (j=0; j<2; j++) {
770 int x = (top+i)*stride[0] + left + j;
771 *block++ = data[0][x];
772 x = (top+i)*stride[1] + left + j;
782 * Create YUV clusters for the entire image
784 static void create_clusters(const AVFrame *frame, int w, int h, uint8_t *yuvClusters)
789 for (j=0; j<w; j+=4) {
790 for (k=0; k < 2; k++)
791 for (l=0; l < 2; l++)
792 frame_block_to_cell(yuvClusters + (l + 2*k)*6, frame->data,
793 i+2*k, j+2*l, frame->linesize);
798 static int generate_codebook(RoqContext *enc, RoqTempdata *tempdata,
799 int *points, int inputCount, roq_cell *results,
800 int size, int cbsize)
802 int i, j, k, ret = 0;
803 int c_size = size*size/4;
805 int *codebook = av_malloc(6*c_size*cbsize*sizeof(int));
809 return AVERROR(ENOMEM);
812 closest_cb = av_malloc(6*c_size*inputCount*sizeof(int));
814 ret = AVERROR(ENOMEM);
818 closest_cb = tempdata->closest_cb2;
820 ret = ff_init_elbg(points, 6 * c_size, inputCount, codebook,
821 cbsize, 1, closest_cb, &enc->randctx);
824 ret = ff_do_elbg(points, 6 * c_size, inputCount, codebook,
825 cbsize, 1, closest_cb, &enc->randctx);
830 for (i=0; i<cbsize; i++)
831 for (k=0; k<c_size; k++) {
833 results->y[j] = *buf++;
835 results->u = (*buf++ + CHROMA_BIAS/2)/CHROMA_BIAS;
836 results->v = (*buf++ + CHROMA_BIAS/2)/CHROMA_BIAS;
846 static int generate_new_codebooks(RoqContext *enc, RoqTempdata *tempData)
849 RoqCodebooks *codebooks = &tempData->codebooks;
850 int max = enc->width*enc->height/16;
852 roq_cell *results4 = av_malloc(sizeof(roq_cell)*MAX_CBS_4x4*4);
853 uint8_t *yuvClusters=av_malloc(sizeof(int)*max*6*4);
854 int *points = av_malloc(max*6*4*sizeof(int));
857 if (!results4 || !yuvClusters || !points) {
858 ret = AVERROR(ENOMEM);
862 /* Subsample YUV data */
863 create_clusters(enc->frame_to_enc, enc->width, enc->height, yuvClusters);
865 /* Cast to integer and apply chroma bias */
866 for (i=0; i<max*24; i++) {
867 bias = ((i%6)<4) ? 1 : CHROMA_BIAS;
868 points[i] = bias*yuvClusters[i];
871 /* Create 4x4 codebooks */
872 if ((ret = generate_codebook(enc, tempData, points, max,
873 results4, 4, MAX_CBS_4x4)) < 0)
876 codebooks->numCB4 = MAX_CBS_4x4;
878 tempData->closest_cb2 = av_malloc(max*4*sizeof(int));
879 if (!tempData->closest_cb2) {
880 ret = AVERROR(ENOMEM);
884 /* Create 2x2 codebooks */
885 if ((ret = generate_codebook(enc, tempData, points, max * 4,
886 enc->cb2x2, 2, MAX_CBS_2x2)) < 0)
889 codebooks->numCB2 = MAX_CBS_2x2;
891 /* Unpack 2x2 codebook clusters */
892 for (i=0; i<codebooks->numCB2; i++)
893 unpack_roq_cell(enc->cb2x2 + i, codebooks->unpacked_cb2 + i*2*2*3);
895 /* Index all 4x4 entries to the 2x2 entries, unpack, and enlarge */
896 for (i=0; i<codebooks->numCB4; i++) {
897 for (j=0; j<4; j++) {
898 unpack_roq_cell(&results4[4*i + j], mb2);
899 index_mb(mb2, codebooks->unpacked_cb2, codebooks->numCB2,
900 &enc->cb4x4[i].idx[j], 2);
902 unpack_roq_qcell(codebooks->unpacked_cb2, enc->cb4x4 + i,
903 codebooks->unpacked_cb4 + i*4*4*3);
904 enlarge_roq_mb4(codebooks->unpacked_cb4 + i*4*4*3,
905 codebooks->unpacked_cb4_enlarged + i*8*8*3);
908 av_free(yuvClusters);
914 static int roq_encode_video(RoqContext *enc)
916 RoqTempdata *tempData = enc->tmpData;
919 memset(tempData, 0, sizeof(*tempData));
921 ret = create_cel_evals(enc, tempData);
925 ret = generate_new_codebooks(enc, tempData);
929 if (enc->framesSinceKeyframe >= 1) {
930 motion_search(enc, 8);
931 motion_search(enc, 4);
935 for (i=0; i<enc->width*enc->height/64; i++)
936 gather_data_for_cel(tempData->cel_evals + i, enc, tempData);
938 /* Quake 3 can't handle chunks bigger than 65535 bytes */
939 if (tempData->mainChunkSize/8 > 65535) {
940 av_log(enc->avctx, AV_LOG_ERROR,
941 "Warning, generated a frame too big (%d > 65535), "
942 "try using a smaller qscale value.\n",
943 tempData->mainChunkSize/8);
945 tempData->mainChunkSize = 0;
946 memset(tempData->used_option, 0, sizeof(tempData->used_option));
947 memset(tempData->codebooks.usedCB4, 0,
948 sizeof(tempData->codebooks.usedCB4));
949 memset(tempData->codebooks.usedCB2, 0,
950 sizeof(tempData->codebooks.usedCB2));
955 remap_codebooks(enc, tempData);
957 write_codebooks(enc, tempData);
959 reconstruct_and_encode_image(enc, tempData, enc->width, enc->height,
960 enc->width*enc->height/64);
962 enc->avctx->coded_frame = enc->current_frame;
964 /* Rotate frame history */
965 FFSWAP(AVFrame *, enc->current_frame, enc->last_frame);
966 FFSWAP(motion_vect *, enc->last_motion4, enc->this_motion4);
967 FFSWAP(motion_vect *, enc->last_motion8, enc->this_motion8);
969 av_free(tempData->cel_evals);
970 av_free(tempData->closest_cb2);
972 enc->framesSinceKeyframe++;
977 static av_cold int roq_encode_end(AVCodecContext *avctx)
979 RoqContext *enc = avctx->priv_data;
981 av_frame_free(&enc->current_frame);
982 av_frame_free(&enc->last_frame);
984 av_free(enc->tmpData);
985 av_free(enc->this_motion4);
986 av_free(enc->last_motion4);
987 av_free(enc->this_motion8);
988 av_free(enc->last_motion8);
993 static av_cold int roq_encode_init(AVCodecContext *avctx)
995 RoqContext *enc = avctx->priv_data;
997 av_lfg_init(&enc->randctx, 1);
999 enc->framesSinceKeyframe = 0;
1000 if ((avctx->width & 0xf) || (avctx->height & 0xf)) {
1001 av_log(avctx, AV_LOG_ERROR, "Dimensions must be divisible by 16\n");
1005 if (((avctx->width)&(avctx->width-1))||((avctx->height)&(avctx->height-1)))
1006 av_log(avctx, AV_LOG_ERROR, "Warning: dimensions not power of two\n");
1008 enc->width = avctx->width;
1009 enc->height = avctx->height;
1011 enc->framesSinceKeyframe = 0;
1012 enc->first_frame = 1;
1014 enc->last_frame = av_frame_alloc();
1015 enc->current_frame = av_frame_alloc();
1016 if (!enc->last_frame || !enc->current_frame) {
1017 roq_encode_end(avctx);
1018 return AVERROR(ENOMEM);
1021 enc->tmpData = av_malloc(sizeof(RoqTempdata));
1024 av_mallocz((enc->width*enc->height/16)*sizeof(motion_vect));
1027 av_malloc ((enc->width*enc->height/16)*sizeof(motion_vect));
1030 av_mallocz((enc->width*enc->height/64)*sizeof(motion_vect));
1033 av_malloc ((enc->width*enc->height/64)*sizeof(motion_vect));
1038 static void roq_write_video_info_chunk(RoqContext *enc)
1040 /* ROQ info chunk */
1041 bytestream_put_le16(&enc->out_buf, RoQ_INFO);
1044 bytestream_put_le32(&enc->out_buf, 8);
1046 /* Unused argument */
1047 bytestream_put_byte(&enc->out_buf, 0x00);
1048 bytestream_put_byte(&enc->out_buf, 0x00);
1051 bytestream_put_le16(&enc->out_buf, enc->width);
1054 bytestream_put_le16(&enc->out_buf, enc->height);
1056 /* Unused in Quake 3, mimics the output of the real encoder */
1057 bytestream_put_byte(&enc->out_buf, 0x08);
1058 bytestream_put_byte(&enc->out_buf, 0x00);
1059 bytestream_put_byte(&enc->out_buf, 0x04);
1060 bytestream_put_byte(&enc->out_buf, 0x00);
1063 static int roq_encode_frame(AVCodecContext *avctx, AVPacket *pkt,
1064 const AVFrame *frame, int *got_packet)
1066 RoqContext *enc = avctx->priv_data;
1071 enc->frame_to_enc = frame;
1074 enc->lambda = frame->quality - 1;
1076 enc->lambda = 2*ROQ_LAMBDA_SCALE;
1078 /* 138 bits max per 8x8 block +
1079 * 256 codebooks*(6 bytes 2x2 + 4 bytes 4x4) + 8 bytes frame header */
1080 size = ((enc->width * enc->height / 64) * 138 + 7) / 8 + 256 * (6 + 4) + 8;
1081 if ((ret = ff_alloc_packet(pkt, size)) < 0) {
1082 av_log(avctx, AV_LOG_ERROR, "Error getting output packet with size %d.\n", size);
1085 enc->out_buf = pkt->data;
1087 /* Check for I frame */
1088 if (enc->framesSinceKeyframe == avctx->gop_size)
1089 enc->framesSinceKeyframe = 0;
1091 if (enc->first_frame) {
1092 /* Alloc memory for the reconstruction data (we must know the stride
1094 if (ff_get_buffer(avctx, enc->current_frame, 0) ||
1095 ff_get_buffer(avctx, enc->last_frame, 0)) {
1096 av_log(avctx, AV_LOG_ERROR, " RoQ: get_buffer() failed\n");
1100 /* Before the first video frame, write a "video info" chunk */
1101 roq_write_video_info_chunk(enc);
1103 enc->first_frame = 0;
1106 /* Encode the actual frame */
1107 ret = roq_encode_video(enc);
1111 pkt->size = enc->out_buf - pkt->data;
1112 if (enc->framesSinceKeyframe == 1)
1113 pkt->flags |= AV_PKT_FLAG_KEY;
1119 AVCodec ff_roq_encoder = {
1121 .long_name = NULL_IF_CONFIG_SMALL("id RoQ video"),
1122 .type = AVMEDIA_TYPE_VIDEO,
1123 .id = AV_CODEC_ID_ROQ,
1124 .priv_data_size = sizeof(RoqContext),
1125 .init = roq_encode_init,
1126 .encode2 = roq_encode_frame,
1127 .close = roq_encode_end,
1128 .supported_framerates = (const AVRational[]){ {30,1}, {0,0} },
1129 .pix_fmts = (const enum AVPixelFormat[]){ AV_PIX_FMT_YUV444P,