827
882
MPV_common_end(s);
885
#endif /* CONFIG_DECODERS */
887
#ifdef CONFIG_ENCODERS
888
static void svq1_write_header(SVQ1Context *s, int frame_type)
893
put_bits(&s->pb, 22, 0x20);
895
/* temporal reference (sure hope this is a "don't care") */
896
put_bits(&s->pb, 8, 0x00);
899
put_bits(&s->pb, 2, frame_type - 1);
901
if (frame_type == I_TYPE) {
903
/* no checksum since frame code is 0x20 */
905
/* no embedded string either */
907
/* output 5 unknown bits (2 + 2 + 1) */
908
put_bits(&s->pb, 5, 2); /* 2 needed by quicktime decoder */
910
for (i = 0; i < 7; i++)
912
if ((svq1_frame_size_table[i].width == s->frame_width) &&
913
(svq1_frame_size_table[i].height == s->frame_height))
915
put_bits(&s->pb, 3, i);
922
put_bits(&s->pb, 3, 7);
923
put_bits(&s->pb, 12, s->frame_width);
924
put_bits(&s->pb, 12, s->frame_height);
928
/* no checksum or extra data (next 2 bits get 0) */
929
put_bits(&s->pb, 2, 0);
933
#define QUALITY_THRESHOLD 100
934
#define THRESHOLD_MULTIPLIER 0.6
936
#if defined(HAVE_ALTIVEC)
940
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){
941
int count, y, x, i, j, split, best_mean, best_score, best_count;
943
int block_sum[7]= {0, 0, 0, 0, 0, 0};
944
int w= 2<<((level+2)>>1);
945
int h= 2<<((level+1)>>1);
947
int16_t block[7][256];
948
const int8_t *codebook_sum, *codebook;
949
const uint16_t (*mean_vlc)[2];
950
const uint8_t (*multistage_vlc)[2];
953
//FIXME optimize, this doenst need to be done multiple times
955
codebook_sum= svq1_intra_codebook_sum[level];
956
codebook= svq1_intra_codebooks[level];
957
mean_vlc= svq1_intra_mean_vlc;
958
multistage_vlc= svq1_intra_multistage_vlc[level];
961
int v= src[x + y*stride];
962
block[0][x + w*y]= v;
968
codebook_sum= svq1_inter_codebook_sum[level];
969
codebook= svq1_inter_codebooks[level];
970
mean_vlc= svq1_inter_mean_vlc + 256;
971
multistage_vlc= svq1_inter_multistage_vlc[level];
974
int v= src[x + y*stride] - ref[x + y*stride];
975
block[0][x + w*y]= v;
983
best_score -= ((block_sum[0]*block_sum[0])>>(level+3));
984
best_mean= (block_sum[0] + (size>>1)) >> (level+3);
987
for(count=1; count<7; count++){
988
int best_vector_score= INT_MAX;
989
int best_vector_sum=-999, best_vector_mean=-999;
990
const int stage= count-1;
991
const int8_t *vector;
994
int sum= codebook_sum[stage*16 + i];
996
int diff, mean, score;
998
vector = codebook + stage*size*16 + i*size;
1000
for(j=0; j<size; j++){
1002
sqr += (v - block[stage][j])*(v - block[stage][j]);
1004
diff= block_sum[stage] - sum;
1005
mean= (diff + (size>>1)) >> (level+3);
1006
assert(mean >-300 && mean<300);
1007
if(intra) mean= av_clip(mean, 0, 255);
1008
else mean= av_clip(mean, -256, 255);
1009
score= sqr - ((diff*(int64_t)diff)>>(level+3)); //FIXME 64bit slooow
1010
if(score < best_vector_score){
1011
best_vector_score= score;
1012
best_vector[stage]= i;
1013
best_vector_sum= sum;
1014
best_vector_mean= mean;
1017
assert(best_vector_mean != -999);
1018
vector= codebook + stage*size*16 + best_vector[stage]*size;
1019
for(j=0; j<size; j++){
1020
block[stage+1][j] = block[stage][j] - vector[j];
1022
block_sum[stage+1]= block_sum[stage] - best_vector_sum;
1023
best_vector_score +=
1024
lambda*(+ 1 + 4*count
1025
+ multistage_vlc[1+count][1]
1026
+ mean_vlc[best_vector_mean][1]);
1028
if(best_vector_score < best_score){
1029
best_score= best_vector_score;
1031
best_mean= best_vector_mean;
1037
if(best_score > threshold && level){
1039
int offset= (level&1) ? stride*h/2 : w/2;
1040
PutBitContext backup[6];
1042
for(i=level-1; i>=0; i--){
1043
backup[i]= s->reorder_pb[i];
1045
score += encode_block(s, src , ref , decoded , stride, level-1, threshold>>1, lambda, intra);
1046
score += encode_block(s, src + offset, ref + offset, decoded + offset, stride, level-1, threshold>>1, lambda, intra);
1049
if(score < best_score){
1053
for(i=level-1; i>=0; i--){
1054
s->reorder_pb[i]= backup[i];
1059
put_bits(&s->reorder_pb[level], 1, split);
1062
assert((best_mean >= 0 && best_mean<256) || !intra);
1063
assert(best_mean >= -256 && best_mean<256);
1064
assert(best_count >=0 && best_count<7);
1065
assert(level<4 || best_count==0);
1067
/* output the encoding */
1068
put_bits(&s->reorder_pb[level],
1069
multistage_vlc[1 + best_count][1],
1070
multistage_vlc[1 + best_count][0]);
1071
put_bits(&s->reorder_pb[level], mean_vlc[best_mean][1],
1072
mean_vlc[best_mean][0]);
1074
for (i = 0; i < best_count; i++){
1075
assert(best_vector[i]>=0 && best_vector[i]<16);
1076
put_bits(&s->reorder_pb[level], 4, best_vector[i]);
1081
decoded[x + y*stride]= src[x + y*stride] - block[best_count][x + w*y] + best_mean;
1090
static int svq1_encode_plane(SVQ1Context *s, int plane, unsigned char *src_plane, unsigned char *ref_plane, unsigned char *decoded_plane,
1091
int width, int height, int src_stride, int stride)
1095
int block_width, block_height;
1098
const int lambda= (s->picture.quality*s->picture.quality) >> (2*FF_LAMBDA_SHIFT);
1100
/* figure out the acceptable level thresholds in advance */
1101
threshold[5] = QUALITY_THRESHOLD;
1102
for (level = 4; level >= 0; level--)
1103
threshold[level] = threshold[level + 1] * THRESHOLD_MULTIPLIER;
1105
block_width = (width + 15) / 16;
1106
block_height = (height + 15) / 16;
1108
if(s->picture.pict_type == P_TYPE){
1109
s->m.avctx= s->avctx;
1110
s->m.current_picture_ptr= &s->m.current_picture;
1111
s->m.last_picture_ptr = &s->m.last_picture;
1112
s->m.last_picture.data[0]= ref_plane;
1114
s->m.last_picture.linesize[0]=
1115
s->m.new_picture.linesize[0]=
1116
s->m.current_picture.linesize[0]= stride;
1118
s->m.height= height;
1119
s->m.mb_width= block_width;
1120
s->m.mb_height= block_height;
1121
s->m.mb_stride= s->m.mb_width+1;
1122
s->m.b8_stride= 2*s->m.mb_width+1;
1124
s->m.pict_type= s->picture.pict_type;
1125
s->m.me_method= s->avctx->me_method;
1126
s->m.me.scene_change_score=0;
1127
s->m.flags= s->avctx->flags;
1128
// s->m.out_format = FMT_H263;
1129
// s->m.unrestricted_mv= 1;
1131
s->m.lambda= s->picture.quality;
1132
s->m.qscale= (s->m.lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7);
1133
s->m.lambda2= (s->m.lambda*s->m.lambda + FF_LAMBDA_SCALE/2) >> FF_LAMBDA_SHIFT;
1135
if(!s->motion_val8[plane]){
1136
s->motion_val8 [plane]= av_mallocz((s->m.b8_stride*block_height*2 + 2)*2*sizeof(int16_t));
1137
s->motion_val16[plane]= av_mallocz((s->m.mb_stride*(block_height + 2) + 1)*2*sizeof(int16_t));
1140
s->m.mb_type= s->mb_type;
1142
//dummies, to avoid segfaults
1143
s->m.current_picture.mb_mean= (uint8_t *)s->dummy;
1144
s->m.current_picture.mb_var= (uint16_t*)s->dummy;
1145
s->m.current_picture.mc_mb_var= (uint16_t*)s->dummy;
1146
s->m.current_picture.mb_type= s->dummy;
1148
s->m.current_picture.motion_val[0]= s->motion_val8[plane] + 2;
1149
s->m.p_mv_table= s->motion_val16[plane] + s->m.mb_stride + 1;
1150
s->m.dsp= s->dsp; //move
1153
s->m.me.dia_size= s->avctx->dia_size;
1154
s->m.first_slice_line=1;
1155
for (y = 0; y < block_height; y++) {
1156
uint8_t src[stride*16];
1158
s->m.new_picture.data[0]= src - y*16*stride; //ugly
1161
for(i=0; i<16 && i + 16*y<height; i++){
1162
memcpy(&src[i*stride], &src_plane[(i+16*y)*src_stride], width);
1163
for(x=width; x<16*block_width; x++)
1164
src[i*stride+x]= src[i*stride+x-1];
1166
for(; i<16 && i + 16*y<16*block_height; i++)
1167
memcpy(&src[i*stride], &src[(i-1)*stride], 16*block_width);
1169
for (x = 0; x < block_width; x++) {
1171
ff_init_block_index(&s->m);
1172
ff_update_block_index(&s->m);
1174
ff_estimate_p_frame_motion(&s->m, x, y);
1176
s->m.first_slice_line=0;
1179
ff_fix_long_p_mvs(&s->m);
1180
ff_fix_long_mvs(&s->m, NULL, 0, s->m.p_mv_table, s->m.f_code, CANDIDATE_MB_TYPE_INTER, 0);
1183
s->m.first_slice_line=1;
1184
for (y = 0; y < block_height; y++) {
1185
uint8_t src[stride*16];
1187
for(i=0; i<16 && i + 16*y<height; i++){
1188
memcpy(&src[i*stride], &src_plane[(i+16*y)*src_stride], width);
1189
for(x=width; x<16*block_width; x++)
1190
src[i*stride+x]= src[i*stride+x-1];
1192
for(; i<16 && i + 16*y<16*block_height; i++)
1193
memcpy(&src[i*stride], &src[(i-1)*stride], 16*block_width);
1196
for (x = 0; x < block_width; x++) {
1197
uint8_t reorder_buffer[3][6][7*32];
1199
int offset = y * 16 * stride + x * 16;
1200
uint8_t *decoded= decoded_plane + offset;
1201
uint8_t *ref= ref_plane + offset;
1202
int score[4]={0,0,0,0}, best;
1203
uint8_t temp[16*stride];
1205
if(s->pb.buf_end - s->pb.buf - (put_bits_count(&s->pb)>>3) < 3000){ //FIXME check size
1206
av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
1211
ff_init_block_index(&s->m);
1212
ff_update_block_index(&s->m);
1214
if(s->picture.pict_type == I_TYPE || (s->m.mb_type[x + y*s->m.mb_stride]&CANDIDATE_MB_TYPE_INTRA)){
1216
init_put_bits(&s->reorder_pb[i], reorder_buffer[0][i], 7*32);
1218
if(s->picture.pict_type == P_TYPE){
1219
const uint8_t *vlc= svq1_block_type_vlc[SVQ1_BLOCK_INTRA];
1220
put_bits(&s->reorder_pb[5], vlc[1], vlc[0]);
1221
score[0]= vlc[1]*lambda;
1223
score[0]+= encode_block(s, src+16*x, NULL, temp, stride, 5, 64, lambda, 1);
1225
count[0][i]= put_bits_count(&s->reorder_pb[i]);
1226
flush_put_bits(&s->reorder_pb[i]);
1233
if(s->picture.pict_type == P_TYPE){
1234
const uint8_t *vlc= svq1_block_type_vlc[SVQ1_BLOCK_INTER];
1235
int mx, my, pred_x, pred_y, dxy;
1236
int16_t *motion_ptr;
1238
motion_ptr= h263_pred_motion(&s->m, 0, 0, &pred_x, &pred_y);
1239
if(s->m.mb_type[x + y*s->m.mb_stride]&CANDIDATE_MB_TYPE_INTER){
1241
init_put_bits(&s->reorder_pb[i], reorder_buffer[1][i], 7*32);
1243
put_bits(&s->reorder_pb[5], vlc[1], vlc[0]);
1245
s->m.pb= s->reorder_pb[5];
1248
assert(mx>=-32 && mx<=31);
1249
assert(my>=-32 && my<=31);
1250
assert(pred_x>=-32 && pred_x<=31);
1251
assert(pred_y>=-32 && pred_y<=31);
1252
ff_h263_encode_motion(&s->m, mx - pred_x, 1);
1253
ff_h263_encode_motion(&s->m, my - pred_y, 1);
1254
s->reorder_pb[5]= s->m.pb;
1255
score[1] += lambda*put_bits_count(&s->reorder_pb[5]);
1257
dxy= (mx&1) + 2*(my&1);
1259
s->dsp.put_pixels_tab[0][dxy](temp+16, ref + (mx>>1) + stride*(my>>1), stride, 16);
1261
score[1]+= encode_block(s, src+16*x, temp+16, decoded, stride, 5, 64, lambda, 0);
1262
best= score[1] <= score[0];
1264
vlc= svq1_block_type_vlc[SVQ1_BLOCK_SKIP];
1265
score[2]= s->dsp.sse[0](NULL, src+16*x, ref, stride, 16);
1266
score[2]+= vlc[1]*lambda;
1267
if(score[2] < score[best] && mx==0 && my==0){
1269
s->dsp.put_pixels_tab[0][0](decoded, ref, stride, 16);
1273
put_bits(&s->pb, vlc[1], vlc[0]);
1279
count[1][i]= put_bits_count(&s->reorder_pb[i]);
1280
flush_put_bits(&s->reorder_pb[i]);
1283
motion_ptr[0 ] = motion_ptr[1 ]=
1284
motion_ptr[2 ] = motion_ptr[3 ]=
1285
motion_ptr[0+2*s->m.b8_stride] = motion_ptr[1+2*s->m.b8_stride]=
1286
motion_ptr[2+2*s->m.b8_stride] = motion_ptr[3+2*s->m.b8_stride]=0;
1290
s->rd_total += score[best];
1292
for(i=5; i>=0; i--){
1293
ff_copy_bits(&s->pb, reorder_buffer[best][i], count[best][i]);
1296
s->dsp.put_pixels_tab[0][0](decoded, temp, stride, 16);
1299
s->m.first_slice_line=0;
1304
static int svq1_encode_init(AVCodecContext *avctx)
1306
SVQ1Context * const s = avctx->priv_data;
1308
dsputil_init(&s->dsp, avctx);
1309
avctx->coded_frame= (AVFrame*)&s->picture;
1311
s->frame_width = avctx->width;
1312
s->frame_height = avctx->height;
1314
s->y_block_width = (s->frame_width + 15) / 16;
1315
s->y_block_height = (s->frame_height + 15) / 16;
1317
s->c_block_width = (s->frame_width / 4 + 15) / 16;
1318
s->c_block_height = (s->frame_height / 4 + 15) / 16;
1322
s->m.me.scratchpad= av_mallocz((avctx->width+64)*2*16*2*sizeof(uint8_t));
1323
s->m.me.map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
1324
s->m.me.score_map = av_mallocz(ME_MAP_SIZE*sizeof(uint32_t));
1325
s->mb_type = av_mallocz((s->y_block_width+1)*s->y_block_height*sizeof(int16_t));
1326
s->dummy = av_mallocz((s->y_block_width+1)*s->y_block_height*sizeof(int32_t));
1327
h263_encode_init(&s->m); //mv_penalty
1332
static int svq1_encode_frame(AVCodecContext *avctx, unsigned char *buf,
1333
int buf_size, void *data)
1335
SVQ1Context * const s = avctx->priv_data;
1336
AVFrame *pict = data;
1337
AVFrame * const p= (AVFrame*)&s->picture;
1341
if(avctx->pix_fmt != PIX_FMT_YUV410P){
1342
av_log(avctx, AV_LOG_ERROR, "unsupported pixel format\n");
1346
if(!s->current_picture.data[0]){
1347
avctx->get_buffer(avctx, &s->current_picture);
1348
avctx->get_buffer(avctx, &s->last_picture);
1351
temp= s->current_picture;
1352
s->current_picture= s->last_picture;
1353
s->last_picture= temp;
1355
init_put_bits(&s->pb, buf, buf_size);
1358
p->pict_type = avctx->gop_size && avctx->frame_number % avctx->gop_size ? P_TYPE : I_TYPE;
1359
p->key_frame = p->pict_type == I_TYPE;
1361
svq1_write_header(s, p->pict_type);
1363
if(svq1_encode_plane(s, i,
1364
s->picture.data[i], s->last_picture.data[i], s->current_picture.data[i],
1365
s->frame_width / (i?4:1), s->frame_height / (i?4:1),
1366
s->picture.linesize[i], s->current_picture.linesize[i]) < 0)
1370
// align_put_bits(&s->pb);
1371
while(put_bits_count(&s->pb) & 31)
1372
put_bits(&s->pb, 1, 0);
1374
flush_put_bits(&s->pb);
1376
return (put_bits_count(&s->pb) / 8);
1379
static int svq1_encode_end(AVCodecContext *avctx)
1381
SVQ1Context * const s = avctx->priv_data;
1384
av_log(avctx, AV_LOG_DEBUG, "RD: %f\n", s->rd_total/(double)(avctx->width*avctx->height*avctx->frame_number));
1386
av_freep(&s->m.me.scratchpad);
1387
av_freep(&s->m.me.map);
1388
av_freep(&s->m.me.score_map);
1389
av_freep(&s->mb_type);
1390
av_freep(&s->dummy);
1393
av_freep(&s->motion_val8[i]);
1394
av_freep(&s->motion_val16[i]);
1400
#endif //CONFIG_ENCODERS
1402
#ifdef CONFIG_DECODERS
831
1403
AVCodec svq1_decoder = {
833
1405
CODEC_TYPE_VIDEO,