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
23 * @file libavcodec/svq1enc.c
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 == FF_I_TYPE) {
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
120 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){
121 int count, y, x, i, j, split, best_mean, best_score, best_count;
123 int block_sum[7]= {0, 0, 0, 0, 0, 0};
124 int w= 2<<((level+2)>>1);
125 int h= 2<<((level+1)>>1);
127 int16_t block[7][256];
128 const int8_t *codebook_sum, *codebook;
129 const uint16_t (*mean_vlc)[2];
130 const uint8_t (*multistage_vlc)[2];
133 //FIXME optimize, this doenst need to be done multiple times
135 codebook_sum= svq1_intra_codebook_sum[level];
136 codebook= ff_svq1_intra_codebooks[level];
137 mean_vlc= ff_svq1_intra_mean_vlc;
138 multistage_vlc= ff_svq1_intra_multistage_vlc[level];
141 int v= src[x + y*stride];
142 block[0][x + w*y]= v;
148 codebook_sum= svq1_inter_codebook_sum[level];
149 codebook= ff_svq1_inter_codebooks[level];
150 mean_vlc= ff_svq1_inter_mean_vlc + 256;
151 multistage_vlc= ff_svq1_inter_multistage_vlc[level];
154 int v= src[x + y*stride] - ref[x + y*stride];
155 block[0][x + w*y]= v;
163 best_score -= ((block_sum[0]*block_sum[0])>>(level+3));
164 best_mean= (block_sum[0] + (size>>1)) >> (level+3);
167 for(count=1; count<7; count++){
168 int best_vector_score= INT_MAX;
169 int best_vector_sum=-999, best_vector_mean=-999;
170 const int stage= count-1;
171 const int8_t *vector;
174 int sum= codebook_sum[stage*16 + i];
175 int sqr, diff, score;
177 vector = codebook + stage*size*16 + i*size;
178 sqr = s->dsp.ssd_int8_vs_int16(vector, block[stage], size);
179 diff= block_sum[stage] - sum;
180 score= sqr - ((diff*(int64_t)diff)>>(level+3)); //FIXME 64bit slooow
181 if(score < best_vector_score){
182 int mean= (diff + (size>>1)) >> (level+3);
183 assert(mean >-300 && mean<300);
184 mean= av_clip(mean, intra?0:-256, 255);
185 best_vector_score= score;
186 best_vector[stage]= i;
187 best_vector_sum= sum;
188 best_vector_mean= mean;
191 assert(best_vector_mean != -999);
192 vector= codebook + stage*size*16 + best_vector[stage]*size;
193 for(j=0; j<size; j++){
194 block[stage+1][j] = block[stage][j] - vector[j];
196 block_sum[stage+1]= block_sum[stage] - best_vector_sum;
198 lambda*(+ 1 + 4*count
199 + multistage_vlc[1+count][1]
200 + mean_vlc[best_vector_mean][1]);
202 if(best_vector_score < best_score){
203 best_score= best_vector_score;
205 best_mean= best_vector_mean;
211 if(best_score > threshold && level){
213 int offset= (level&1) ? stride*h/2 : w/2;
214 PutBitContext backup[6];
216 for(i=level-1; i>=0; i--){
217 backup[i]= s->reorder_pb[i];
219 score += encode_block(s, src , ref , decoded , stride, level-1, threshold>>1, lambda, intra);
220 score += encode_block(s, src + offset, ref + offset, decoded + offset, stride, level-1, threshold>>1, lambda, intra);
223 if(score < best_score){
227 for(i=level-1; i>=0; i--){
228 s->reorder_pb[i]= backup[i];
233 put_bits(&s->reorder_pb[level], 1, split);
236 assert((best_mean >= 0 && best_mean<256) || !intra);
237 assert(best_mean >= -256 && best_mean<256);
238 assert(best_count >=0 && best_count<7);
239 assert(level<4 || best_count==0);
241 /* output the encoding */
242 put_bits(&s->reorder_pb[level],
243 multistage_vlc[1 + best_count][1],
244 multistage_vlc[1 + best_count][0]);
245 put_bits(&s->reorder_pb[level], mean_vlc[best_mean][1],
246 mean_vlc[best_mean][0]);
248 for (i = 0; i < best_count; i++){
249 assert(best_vector[i]>=0 && best_vector[i]<16);
250 put_bits(&s->reorder_pb[level], 4, best_vector[i]);
255 decoded[x + y*stride]= src[x + y*stride] - block[best_count][x + w*y] + best_mean;
264 static int svq1_encode_plane(SVQ1Context *s, int plane, unsigned char *src_plane, unsigned char *ref_plane, unsigned char *decoded_plane,
265 int width, int height, int src_stride, int stride)
269 int block_width, block_height;
272 uint8_t *src = s->scratchbuf + stride * 16;
273 const int lambda= (s->picture.quality*s->picture.quality) >> (2*FF_LAMBDA_SHIFT);
275 /* figure out the acceptable level thresholds in advance */
276 threshold[5] = QUALITY_THRESHOLD;
277 for (level = 4; level >= 0; level--)
278 threshold[level] = threshold[level + 1] * THRESHOLD_MULTIPLIER;
280 block_width = (width + 15) / 16;
281 block_height = (height + 15) / 16;
283 if(s->picture.pict_type == FF_P_TYPE){
284 s->m.avctx= s->avctx;
285 s->m.current_picture_ptr= &s->m.current_picture;
286 s->m.last_picture_ptr = &s->m.last_picture;
287 s->m.last_picture.data[0]= ref_plane;
289 s->m.last_picture.linesize[0]=
290 s->m.new_picture.linesize[0]=
291 s->m.current_picture.linesize[0]= stride;
294 s->m.mb_width= block_width;
295 s->m.mb_height= block_height;
296 s->m.mb_stride= s->m.mb_width+1;
297 s->m.b8_stride= 2*s->m.mb_width+1;
299 s->m.pict_type= s->picture.pict_type;
300 s->m.me_method= s->avctx->me_method;
301 s->m.me.scene_change_score=0;
302 s->m.flags= s->avctx->flags;
303 // s->m.out_format = FMT_H263;
304 // s->m.unrestricted_mv= 1;
306 s->m.lambda= s->picture.quality;
307 s->m.qscale= (s->m.lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);
308 s->m.lambda2= (s->m.lambda*s->m.lambda + FF_LAMBDA_SCALE/2) >> FF_LAMBDA_SHIFT;
310 if(!s->motion_val8[plane]){
311 s->motion_val8 [plane]= av_mallocz((s->m.b8_stride*block_height*2 + 2)*2*sizeof(int16_t));
312 s->motion_val16[plane]= av_mallocz((s->m.mb_stride*(block_height + 2) + 1)*2*sizeof(int16_t));
315 s->m.mb_type= s->mb_type;
317 //dummies, to avoid segfaults
318 s->m.current_picture.mb_mean= (uint8_t *)s->dummy;
319 s->m.current_picture.mb_var= (uint16_t*)s->dummy;
320 s->m.current_picture.mc_mb_var= (uint16_t*)s->dummy;
321 s->m.current_picture.mb_type= s->dummy;
323 s->m.current_picture.motion_val[0]= s->motion_val8[plane] + 2;
324 s->m.p_mv_table= s->motion_val16[plane] + s->m.mb_stride + 1;
325 s->m.dsp= s->dsp; //move
328 s->m.me.dia_size= s->avctx->dia_size;
329 s->m.first_slice_line=1;
330 for (y = 0; y < block_height; y++) {
331 s->m.new_picture.data[0]= src - y*16*stride; //ugly
334 for(i=0; i<16 && i + 16*y<height; i++){
335 memcpy(&src[i*stride], &src_plane[(i+16*y)*src_stride], width);
336 for(x=width; x<16*block_width; x++)
337 src[i*stride+x]= src[i*stride+x-1];
339 for(; i<16 && i + 16*y<16*block_height; i++)
340 memcpy(&src[i*stride], &src[(i-1)*stride], 16*block_width);
342 for (x = 0; x < block_width; x++) {
344 ff_init_block_index(&s->m);
345 ff_update_block_index(&s->m);
347 ff_estimate_p_frame_motion(&s->m, x, y);
349 s->m.first_slice_line=0;
352 ff_fix_long_p_mvs(&s->m);
353 ff_fix_long_mvs(&s->m, NULL, 0, s->m.p_mv_table, s->m.f_code, CANDIDATE_MB_TYPE_INTER, 0);
356 s->m.first_slice_line=1;
357 for (y = 0; y < block_height; y++) {
358 for(i=0; i<16 && i + 16*y<height; i++){
359 memcpy(&src[i*stride], &src_plane[(i+16*y)*src_stride], width);
360 for(x=width; x<16*block_width; x++)
361 src[i*stride+x]= src[i*stride+x-1];
363 for(; i<16 && i + 16*y<16*block_height; i++)
364 memcpy(&src[i*stride], &src[(i-1)*stride], 16*block_width);
367 for (x = 0; x < block_width; x++) {
368 uint8_t reorder_buffer[3][6][7*32];
370 int offset = y * 16 * stride + x * 16;
371 uint8_t *decoded= decoded_plane + offset;
372 uint8_t *ref= ref_plane + offset;
373 int score[4]={0,0,0,0}, best;
374 uint8_t *temp = s->scratchbuf;
376 if(s->pb.buf_end - s->pb.buf - (put_bits_count(&s->pb)>>3) < 3000){ //FIXME check size
377 av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
382 ff_init_block_index(&s->m);
383 ff_update_block_index(&s->m);
385 if(s->picture.pict_type == FF_I_TYPE || (s->m.mb_type[x + y*s->m.mb_stride]&CANDIDATE_MB_TYPE_INTRA)){
387 init_put_bits(&s->reorder_pb[i], reorder_buffer[0][i], 7*32);
389 if(s->picture.pict_type == FF_P_TYPE){
390 const uint8_t *vlc= ff_svq1_block_type_vlc[SVQ1_BLOCK_INTRA];
391 put_bits(&s->reorder_pb[5], vlc[1], vlc[0]);
392 score[0]= vlc[1]*lambda;
394 score[0]+= encode_block(s, src+16*x, NULL, temp, stride, 5, 64, lambda, 1);
396 count[0][i]= put_bits_count(&s->reorder_pb[i]);
397 flush_put_bits(&s->reorder_pb[i]);
404 if(s->picture.pict_type == FF_P_TYPE){
405 const uint8_t *vlc= ff_svq1_block_type_vlc[SVQ1_BLOCK_INTER];
406 int mx, my, pred_x, pred_y, dxy;
409 motion_ptr= h263_pred_motion(&s->m, 0, 0, &pred_x, &pred_y);
410 if(s->m.mb_type[x + y*s->m.mb_stride]&CANDIDATE_MB_TYPE_INTER){
412 init_put_bits(&s->reorder_pb[i], reorder_buffer[1][i], 7*32);
414 put_bits(&s->reorder_pb[5], vlc[1], vlc[0]);
416 s->m.pb= s->reorder_pb[5];
419 assert(mx>=-32 && mx<=31);
420 assert(my>=-32 && my<=31);
421 assert(pred_x>=-32 && pred_x<=31);
422 assert(pred_y>=-32 && pred_y<=31);
423 ff_h263_encode_motion(&s->m, mx - pred_x, 1);
424 ff_h263_encode_motion(&s->m, my - pred_y, 1);
425 s->reorder_pb[5]= s->m.pb;
426 score[1] += lambda*put_bits_count(&s->reorder_pb[5]);
428 dxy= (mx&1) + 2*(my&1);
430 s->dsp.put_pixels_tab[0][dxy](temp+16, ref + (mx>>1) + stride*(my>>1), stride, 16);
432 score[1]+= encode_block(s, src+16*x, temp+16, decoded, stride, 5, 64, lambda, 0);
433 best= score[1] <= score[0];
435 vlc= ff_svq1_block_type_vlc[SVQ1_BLOCK_SKIP];
436 score[2]= s->dsp.sse[0](NULL, src+16*x, ref, stride, 16);
437 score[2]+= vlc[1]*lambda;
438 if(score[2] < score[best] && mx==0 && my==0){
440 s->dsp.put_pixels_tab[0][0](decoded, ref, stride, 16);
444 put_bits(&s->pb, vlc[1], vlc[0]);
450 count[1][i]= put_bits_count(&s->reorder_pb[i]);
451 flush_put_bits(&s->reorder_pb[i]);
454 motion_ptr[0 ] = motion_ptr[1 ]=
455 motion_ptr[2 ] = motion_ptr[3 ]=
456 motion_ptr[0+2*s->m.b8_stride] = motion_ptr[1+2*s->m.b8_stride]=
457 motion_ptr[2+2*s->m.b8_stride] = motion_ptr[3+2*s->m.b8_stride]=0;
461 s->rd_total += score[best];
464 ff_copy_bits(&s->pb, reorder_buffer[best][i], count[best][i]);
467 s->dsp.put_pixels_tab[0][0](decoded, temp, stride, 16);
470 s->m.first_slice_line=0;
475 static av_cold int svq1_encode_init(AVCodecContext *avctx)
477 SVQ1Context * const s = avctx->priv_data;
479 dsputil_init(&s->dsp, avctx);
480 avctx->coded_frame= (AVFrame*)&s->picture;
482 s->frame_width = avctx->width;
483 s->frame_height = avctx->height;
485 s->y_block_width = (s->frame_width + 15) / 16;
486 s->y_block_height = (s->frame_height + 15) / 16;
488 s->c_block_width = (s->frame_width / 4 + 15) / 16;
489 s->c_block_height = (s->frame_height / 4 + 15) / 16;
494 s->m.me.scratchpad= av_mallocz((avctx->width+64)*2*16*2*sizeof(uint8_t));
495 s->m.me.map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
496 s->m.me.score_map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
497 s->mb_type = av_mallocz((s->y_block_width+1)*s->y_block_height*sizeof(int16_t));
498 s->dummy = av_mallocz((s->y_block_width+1)*s->y_block_height*sizeof(int32_t));
499 h263_encode_init(&s->m); //mv_penalty
504 static int svq1_encode_frame(AVCodecContext *avctx, unsigned char *buf,
505 int buf_size, void *data)
507 SVQ1Context * const s = avctx->priv_data;
508 AVFrame *pict = data;
509 AVFrame * const p= (AVFrame*)&s->picture;
513 if(avctx->pix_fmt != PIX_FMT_YUV410P){
514 av_log(avctx, AV_LOG_ERROR, "unsupported pixel format\n");
518 if(!s->current_picture.data[0]){
519 avctx->get_buffer(avctx, &s->current_picture);
520 avctx->get_buffer(avctx, &s->last_picture);
521 s->scratchbuf = av_malloc(s->current_picture.linesize[0] * 16 * 2);
524 temp= s->current_picture;
525 s->current_picture= s->last_picture;
526 s->last_picture= temp;
528 init_put_bits(&s->pb, buf, buf_size);
531 p->pict_type = avctx->gop_size && avctx->frame_number % avctx->gop_size ? FF_P_TYPE : FF_I_TYPE;
532 p->key_frame = p->pict_type == FF_I_TYPE;
534 svq1_write_header(s, p->pict_type);
536 if(svq1_encode_plane(s, i,
537 s->picture.data[i], s->last_picture.data[i], s->current_picture.data[i],
538 s->frame_width / (i?4:1), s->frame_height / (i?4:1),
539 s->picture.linesize[i], s->current_picture.linesize[i]) < 0)
543 // align_put_bits(&s->pb);
544 while(put_bits_count(&s->pb) & 31)
545 put_bits(&s->pb, 1, 0);
547 flush_put_bits(&s->pb);
549 return put_bits_count(&s->pb) / 8;
552 static av_cold int svq1_encode_end(AVCodecContext *avctx)
554 SVQ1Context * const s = avctx->priv_data;
557 av_log(avctx, AV_LOG_DEBUG, "RD: %f\n", s->rd_total/(double)(avctx->width*avctx->height*avctx->frame_number));
559 av_freep(&s->m.me.scratchpad);
560 av_freep(&s->m.me.map);
561 av_freep(&s->m.me.score_map);
562 av_freep(&s->mb_type);
564 av_freep(&s->scratchbuf);
567 av_freep(&s->motion_val8[i]);
568 av_freep(&s->motion_val16[i]);
575 AVCodec svq1_encoder = {
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"),