3 * Copyright (C) 2004 Mike Melanson <melanson@pcisys.net>
5 * This file is part of FFmpeg.
7 * FFmpeg is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
12 * FFmpeg is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
24 * Sorenson Vector Quantizer #1 (SVQ1) video codec.
25 * For more information of the SVQ1 algorithm, visit:
26 * http://www.pcisys.net/~melanson/codecs/
32 #include "mpegvideo.h"
37 #include "svq1enc_cb.h"
43 typedef struct SVQ1Context {
44 MpegEncContext m; // needed for motion estimation, should not be used for anything else, the idea is to make the motion estimation eventually independent of MpegEncContext, so this will be removed then (FIXME/XXX)
45 AVCodecContext *avctx;
48 AVFrame current_picture;
53 PutBitContext reorder_pb[6]; //why ooh why this sick breadth first order, everything is slower and more complex
58 /* Y plane block dimensions */
62 /* U & V plane (C planes) block dimensions */
68 int16_t (*motion_val8[3])[2];
69 int16_t (*motion_val16[3])[2];
76 static void svq1_write_header(SVQ1Context *s, int frame_type)
81 put_bits(&s->pb, 22, 0x20);
83 /* temporal reference (sure hope this is a "don't care") */
84 put_bits(&s->pb, 8, 0x00);
87 put_bits(&s->pb, 2, frame_type - 1);
89 if (frame_type == AV_PICTURE_TYPE_I) {
91 /* no checksum since frame code is 0x20 */
93 /* no embedded string either */
95 /* output 5 unknown bits (2 + 2 + 1) */
96 put_bits(&s->pb, 5, 2); /* 2 needed by quicktime decoder */
98 i= ff_match_2uint16(ff_svq1_frame_size_table, FF_ARRAY_ELEMS(ff_svq1_frame_size_table), s->frame_width, s->frame_height);
99 put_bits(&s->pb, 3, i);
103 put_bits(&s->pb, 12, s->frame_width);
104 put_bits(&s->pb, 12, s->frame_height);
108 /* no checksum or extra data (next 2 bits get 0) */
109 put_bits(&s->pb, 2, 0);
113 #define QUALITY_THRESHOLD 100
114 #define THRESHOLD_MULTIPLIER 0.6
116 static int encode_block(SVQ1Context *s, uint8_t *src, uint8_t *ref, uint8_t *decoded, int stride, int level, int threshold, int lambda, int intra){
117 int count, y, x, i, j, split, best_mean, best_score, best_count;
119 int block_sum[7]= {0, 0, 0, 0, 0, 0};
120 int w= 2<<((level+2)>>1);
121 int h= 2<<((level+1)>>1);
123 int16_t block[7][256];
124 const int8_t *codebook_sum, *codebook;
125 const uint16_t (*mean_vlc)[2];
126 const uint8_t (*multistage_vlc)[2];
129 //FIXME optimize, this doenst need to be done multiple times
131 codebook_sum= svq1_intra_codebook_sum[level];
132 codebook= ff_svq1_intra_codebooks[level];
133 mean_vlc= ff_svq1_intra_mean_vlc;
134 multistage_vlc= ff_svq1_intra_multistage_vlc[level];
137 int v= src[x + y*stride];
138 block[0][x + w*y]= v;
144 codebook_sum= svq1_inter_codebook_sum[level];
145 codebook= ff_svq1_inter_codebooks[level];
146 mean_vlc= ff_svq1_inter_mean_vlc + 256;
147 multistage_vlc= ff_svq1_inter_multistage_vlc[level];
150 int v= src[x + y*stride] - ref[x + y*stride];
151 block[0][x + w*y]= v;
159 best_score -= (int)(((unsigned)block_sum[0]*block_sum[0])>>(level+3));
160 best_mean= (block_sum[0] + (size>>1)) >> (level+3);
163 for(count=1; count<7; count++){
164 int best_vector_score= INT_MAX;
165 int best_vector_sum=-999, best_vector_mean=-999;
166 const int stage= count-1;
167 const int8_t *vector;
170 int sum= codebook_sum[stage*16 + i];
171 int sqr, diff, score;
173 vector = codebook + stage*size*16 + i*size;
174 sqr = s->dsp.ssd_int8_vs_int16(vector, block[stage], size);
175 diff= block_sum[stage] - sum;
176 score= sqr - ((diff*(int64_t)diff)>>(level+3)); //FIXME 64bit slooow
177 if(score < best_vector_score){
178 int mean= (diff + (size>>1)) >> (level+3);
179 assert(mean >-300 && mean<300);
180 mean= av_clip(mean, intra?0:-256, 255);
181 best_vector_score= score;
182 best_vector[stage]= i;
183 best_vector_sum= sum;
184 best_vector_mean= mean;
187 assert(best_vector_mean != -999);
188 vector= codebook + stage*size*16 + best_vector[stage]*size;
189 for(j=0; j<size; j++){
190 block[stage+1][j] = block[stage][j] - vector[j];
192 block_sum[stage+1]= block_sum[stage] - best_vector_sum;
194 lambda*(+ 1 + 4*count
195 + multistage_vlc[1+count][1]
196 + mean_vlc[best_vector_mean][1]);
198 if(best_vector_score < best_score){
199 best_score= best_vector_score;
201 best_mean= best_vector_mean;
207 if(best_score > threshold && level){
209 int offset= (level&1) ? stride*h/2 : w/2;
210 PutBitContext backup[6];
212 for(i=level-1; i>=0; i--){
213 backup[i]= s->reorder_pb[i];
215 score += encode_block(s, src , ref , decoded , stride, level-1, threshold>>1, lambda, intra);
216 score += encode_block(s, src + offset, ref + offset, decoded + offset, stride, level-1, threshold>>1, lambda, intra);
219 if(score < best_score){
223 for(i=level-1; i>=0; i--){
224 s->reorder_pb[i]= backup[i];
229 put_bits(&s->reorder_pb[level], 1, split);
232 assert((best_mean >= 0 && best_mean<256) || !intra);
233 assert(best_mean >= -256 && best_mean<256);
234 assert(best_count >=0 && best_count<7);
235 assert(level<4 || best_count==0);
237 /* output the encoding */
238 put_bits(&s->reorder_pb[level],
239 multistage_vlc[1 + best_count][1],
240 multistage_vlc[1 + best_count][0]);
241 put_bits(&s->reorder_pb[level], mean_vlc[best_mean][1],
242 mean_vlc[best_mean][0]);
244 for (i = 0; i < best_count; i++){
245 assert(best_vector[i]>=0 && best_vector[i]<16);
246 put_bits(&s->reorder_pb[level], 4, best_vector[i]);
251 decoded[x + y*stride]= src[x + y*stride] - block[best_count][x + w*y] + best_mean;
260 static int svq1_encode_plane(SVQ1Context *s, int plane, unsigned char *src_plane, unsigned char *ref_plane, unsigned char *decoded_plane,
261 int width, int height, int src_stride, int stride)
265 int block_width, block_height;
268 uint8_t *src = s->scratchbuf + stride * 16;
269 const int lambda= (s->picture.quality*s->picture.quality) >> (2*FF_LAMBDA_SHIFT);
271 /* figure out the acceptable level thresholds in advance */
272 threshold[5] = QUALITY_THRESHOLD;
273 for (level = 4; level >= 0; level--)
274 threshold[level] = threshold[level + 1] * THRESHOLD_MULTIPLIER;
276 block_width = (width + 15) / 16;
277 block_height = (height + 15) / 16;
279 if(s->picture.pict_type == AV_PICTURE_TYPE_P){
280 s->m.avctx= s->avctx;
281 s->m.current_picture_ptr= &s->m.current_picture;
282 s->m.last_picture_ptr = &s->m.last_picture;
283 s->m.last_picture.f.data[0] = ref_plane;
285 s->m.last_picture.f.linesize[0] =
286 s->m.new_picture.f.linesize[0] =
287 s->m.current_picture.f.linesize[0] = stride;
290 s->m.mb_width= block_width;
291 s->m.mb_height= block_height;
292 s->m.mb_stride= s->m.mb_width+1;
293 s->m.b8_stride= 2*s->m.mb_width+1;
295 s->m.pict_type= s->picture.pict_type;
296 s->m.me_method= s->avctx->me_method;
297 s->m.me.scene_change_score=0;
298 s->m.flags= s->avctx->flags;
299 // s->m.out_format = FMT_H263;
300 // s->m.unrestricted_mv= 1;
302 s->m.lambda= s->picture.quality;
303 s->m.qscale= (s->m.lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);
304 s->m.lambda2= (s->m.lambda*s->m.lambda + FF_LAMBDA_SCALE/2) >> FF_LAMBDA_SHIFT;
306 if(!s->motion_val8[plane]){
307 s->motion_val8 [plane]= av_mallocz((s->m.b8_stride*block_height*2 + 2)*2*sizeof(int16_t));
308 s->motion_val16[plane]= av_mallocz((s->m.mb_stride*(block_height + 2) + 1)*2*sizeof(int16_t));
311 s->m.mb_type= s->mb_type;
313 //dummies, to avoid segfaults
314 s->m.current_picture.mb_mean= (uint8_t *)s->dummy;
315 s->m.current_picture.mb_var= (uint16_t*)s->dummy;
316 s->m.current_picture.mc_mb_var= (uint16_t*)s->dummy;
317 s->m.current_picture.f.mb_type = s->dummy;
319 s->m.current_picture.f.motion_val[0] = s->motion_val8[plane] + 2;
320 s->m.p_mv_table= s->motion_val16[plane] + s->m.mb_stride + 1;
321 s->m.dsp= s->dsp; //move
324 s->m.me.dia_size= s->avctx->dia_size;
325 s->m.first_slice_line=1;
326 for (y = 0; y < block_height; y++) {
327 s->m.new_picture.f.data[0] = src - y*16*stride; //ugly
330 for(i=0; i<16 && i + 16*y<height; i++){
331 memcpy(&src[i*stride], &src_plane[(i+16*y)*src_stride], width);
332 for(x=width; x<16*block_width; x++)
333 src[i*stride+x]= src[i*stride+x-1];
335 for(; i<16 && i + 16*y<16*block_height; i++)
336 memcpy(&src[i*stride], &src[(i-1)*stride], 16*block_width);
338 for (x = 0; x < block_width; x++) {
340 ff_init_block_index(&s->m);
341 ff_update_block_index(&s->m);
343 ff_estimate_p_frame_motion(&s->m, x, y);
345 s->m.first_slice_line=0;
348 ff_fix_long_p_mvs(&s->m);
349 ff_fix_long_mvs(&s->m, NULL, 0, s->m.p_mv_table, s->m.f_code, CANDIDATE_MB_TYPE_INTER, 0);
352 s->m.first_slice_line=1;
353 for (y = 0; y < block_height; y++) {
354 for(i=0; i<16 && i + 16*y<height; i++){
355 memcpy(&src[i*stride], &src_plane[(i+16*y)*src_stride], width);
356 for(x=width; x<16*block_width; x++)
357 src[i*stride+x]= src[i*stride+x-1];
359 for(; i<16 && i + 16*y<16*block_height; i++)
360 memcpy(&src[i*stride], &src[(i-1)*stride], 16*block_width);
363 for (x = 0; x < block_width; x++) {
364 uint8_t reorder_buffer[3][6][7*32];
366 int offset = y * 16 * stride + x * 16;
367 uint8_t *decoded= decoded_plane + offset;
368 uint8_t *ref= ref_plane + offset;
369 int score[4]={0,0,0,0}, best;
370 uint8_t *temp = s->scratchbuf;
372 if(s->pb.buf_end - s->pb.buf - (put_bits_count(&s->pb)>>3) < 3000){ //FIXME check size
373 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
378 ff_init_block_index(&s->m);
379 ff_update_block_index(&s->m);
381 if(s->picture.pict_type == AV_PICTURE_TYPE_I || (s->m.mb_type[x + y*s->m.mb_stride]&CANDIDATE_MB_TYPE_INTRA)){
383 init_put_bits(&s->reorder_pb[i], reorder_buffer[0][i], 7*32);
385 if(s->picture.pict_type == AV_PICTURE_TYPE_P){
386 const uint8_t *vlc= ff_svq1_block_type_vlc[SVQ1_BLOCK_INTRA];
387 put_bits(&s->reorder_pb[5], vlc[1], vlc[0]);
388 score[0]= vlc[1]*lambda;
390 score[0]+= encode_block(s, src+16*x, NULL, temp, stride, 5, 64, lambda, 1);
392 count[0][i]= put_bits_count(&s->reorder_pb[i]);
393 flush_put_bits(&s->reorder_pb[i]);
400 if(s->picture.pict_type == AV_PICTURE_TYPE_P){
401 const uint8_t *vlc= ff_svq1_block_type_vlc[SVQ1_BLOCK_INTER];
402 int mx, my, pred_x, pred_y, dxy;
405 motion_ptr= h263_pred_motion(&s->m, 0, 0, &pred_x, &pred_y);
406 if(s->m.mb_type[x + y*s->m.mb_stride]&CANDIDATE_MB_TYPE_INTER){
408 init_put_bits(&s->reorder_pb[i], reorder_buffer[1][i], 7*32);
410 put_bits(&s->reorder_pb[5], vlc[1], vlc[0]);
412 s->m.pb= s->reorder_pb[5];
415 assert(mx>=-32 && mx<=31);
416 assert(my>=-32 && my<=31);
417 assert(pred_x>=-32 && pred_x<=31);
418 assert(pred_y>=-32 && pred_y<=31);
419 ff_h263_encode_motion(&s->m, mx - pred_x, 1);
420 ff_h263_encode_motion(&s->m, my - pred_y, 1);
421 s->reorder_pb[5]= s->m.pb;
422 score[1] += lambda*put_bits_count(&s->reorder_pb[5]);
424 dxy= (mx&1) + 2*(my&1);
426 s->dsp.put_pixels_tab[0][dxy](temp+16, ref + (mx>>1) + stride*(my>>1), stride, 16);
428 score[1]+= encode_block(s, src+16*x, temp+16, decoded, stride, 5, 64, lambda, 0);
429 best= score[1] <= score[0];
431 vlc= ff_svq1_block_type_vlc[SVQ1_BLOCK_SKIP];
432 score[2]= s->dsp.sse[0](NULL, src+16*x, ref, stride, 16);
433 score[2]+= vlc[1]*lambda;
434 if(score[2] < score[best] && mx==0 && my==0){
436 s->dsp.put_pixels_tab[0][0](decoded, ref, stride, 16);
440 put_bits(&s->pb, vlc[1], vlc[0]);
446 count[1][i]= put_bits_count(&s->reorder_pb[i]);
447 flush_put_bits(&s->reorder_pb[i]);
450 motion_ptr[0 ] = motion_ptr[1 ]=
451 motion_ptr[2 ] = motion_ptr[3 ]=
452 motion_ptr[0+2*s->m.b8_stride] = motion_ptr[1+2*s->m.b8_stride]=
453 motion_ptr[2+2*s->m.b8_stride] = motion_ptr[3+2*s->m.b8_stride]=0;
457 s->rd_total += score[best];
460 avpriv_copy_bits(&s->pb, reorder_buffer[best][i], count[best][i]);
463 s->dsp.put_pixels_tab[0][0](decoded, temp, stride, 16);
466 s->m.first_slice_line=0;
471 static av_cold int svq1_encode_init(AVCodecContext *avctx)
473 SVQ1Context * const s = avctx->priv_data;
475 dsputil_init(&s->dsp, avctx);
476 avctx->coded_frame= (AVFrame*)&s->picture;
478 s->frame_width = avctx->width;
479 s->frame_height = avctx->height;
481 s->y_block_width = (s->frame_width + 15) / 16;
482 s->y_block_height = (s->frame_height + 15) / 16;
484 s->c_block_width = (s->frame_width / 4 + 15) / 16;
485 s->c_block_height = (s->frame_height / 4 + 15) / 16;
490 s->m.me.scratchpad= av_mallocz((avctx->width+64)*2*16*2*sizeof(uint8_t));
491 s->m.me.map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
492 s->m.me.score_map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
493 s->mb_type = av_mallocz((s->y_block_width+1)*s->y_block_height*sizeof(int16_t));
494 s->dummy = av_mallocz((s->y_block_width+1)*s->y_block_height*sizeof(int32_t));
495 h263_encode_init(&s->m); //mv_penalty
500 static int svq1_encode_frame(AVCodecContext *avctx, unsigned char *buf,
501 int buf_size, void *data)
503 SVQ1Context * const s = avctx->priv_data;
504 AVFrame *pict = data;
505 AVFrame * const p= (AVFrame*)&s->picture;
509 if(avctx->pix_fmt != PIX_FMT_YUV410P){
510 av_log(avctx, AV_LOG_ERROR, "unsupported pixel format\n");
514 if(!s->current_picture.data[0]){
515 avctx->get_buffer(avctx, &s->current_picture);
516 avctx->get_buffer(avctx, &s->last_picture);
517 s->scratchbuf = av_malloc(s->current_picture.linesize[0] * 16 * 2);
520 temp= s->current_picture;
521 s->current_picture= s->last_picture;
522 s->last_picture= temp;
524 init_put_bits(&s->pb, buf, buf_size);
527 p->pict_type = avctx->gop_size && avctx->frame_number % avctx->gop_size ? AV_PICTURE_TYPE_P : AV_PICTURE_TYPE_I;
528 p->key_frame = p->pict_type == AV_PICTURE_TYPE_I;
530 svq1_write_header(s, p->pict_type);
532 if(svq1_encode_plane(s, i,
533 s->picture.data[i], s->last_picture.data[i], s->current_picture.data[i],
534 s->frame_width / (i?4:1), s->frame_height / (i?4:1),
535 s->picture.linesize[i], s->current_picture.linesize[i]) < 0)
539 // avpriv_align_put_bits(&s->pb);
540 while(put_bits_count(&s->pb) & 31)
541 put_bits(&s->pb, 1, 0);
543 flush_put_bits(&s->pb);
545 return put_bits_count(&s->pb) / 8;
548 static av_cold int svq1_encode_end(AVCodecContext *avctx)
550 SVQ1Context * const s = avctx->priv_data;
553 av_log(avctx, AV_LOG_DEBUG, "RD: %f\n", s->rd_total/(double)(avctx->width*avctx->height*avctx->frame_number));
555 av_freep(&s->m.me.scratchpad);
556 av_freep(&s->m.me.map);
557 av_freep(&s->m.me.score_map);
558 av_freep(&s->mb_type);
560 av_freep(&s->scratchbuf);
563 av_freep(&s->motion_val8[i]);
564 av_freep(&s->motion_val16[i]);
566 if(s->current_picture.data[0])
567 avctx->release_buffer(avctx, &s->current_picture);
568 if(s->last_picture.data[0])
569 avctx->release_buffer(avctx, &s->last_picture);
575 AVCodec ff_svq1_encoder = {
577 .type = AVMEDIA_TYPE_VIDEO,
579 .priv_data_size = sizeof(SVQ1Context),
580 .init = svq1_encode_init,
581 .encode = svq1_encode_frame,
582 .close = svq1_encode_end,
583 .pix_fmts= (const enum PixelFormat[]){PIX_FMT_YUV410P, PIX_FMT_NONE},
584 .long_name= NULL_IF_CONFIG_SMALL("Sorenson Vector Quantizer 1 / Sorenson Video 1 / SVQ1"),