1 /*****************************************************************************
2 * mvpred.c: motion vector prediction
3 *****************************************************************************
4 * Copyright (C) 2003-2010 x264 project
6 * Authors: Loren Merritt <lorenm@u.washington.edu>
7 * Fiona Glaser <fiona@x264.com>
8 * Laurent Aimar <fenrir@via.ecp.fr>
10 * This program is free software; you can redistribute it and/or modify
11 * it under the terms of the GNU General Public License as published by
12 * the Free Software Foundation; either version 2 of the License, or
13 * (at your option) any later version.
15 * This program 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
18 * GNU General Public License for more details.
20 * You should have received a copy of the GNU General Public License
21 * along with this program; if not, write to the Free Software
22 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02111, USA.
24 * This program is also available under a commercial proprietary license.
25 * For more information, contact us at licensing@x264.com.
26 *****************************************************************************/
30 void x264_mb_predict_mv( x264_t *h, int i_list, int idx, int i_width, int16_t mvp[2] )
32 const int i8 = x264_scan8[idx];
33 const int i_ref= h->mb.cache.ref[i_list][i8];
34 int i_refa = h->mb.cache.ref[i_list][i8 - 1];
35 int16_t *mv_a = h->mb.cache.mv[i_list][i8 - 1];
36 int i_refb = h->mb.cache.ref[i_list][i8 - 8];
37 int16_t *mv_b = h->mb.cache.mv[i_list][i8 - 8];
38 int i_refc = h->mb.cache.ref[i_list][i8 - 8 + i_width];
39 int16_t *mv_c = h->mb.cache.mv[i_list][i8 - 8 + i_width];
41 if( (idx&3) >= 2 + (i_width&1) || i_refc == -2 )
43 i_refc = h->mb.cache.ref[i_list][i8 - 8 - 1];
44 mv_c = h->mb.cache.mv[i_list][i8 - 8 - 1];
47 if( h->mb.i_partition == D_16x8 )
66 else if( h->mb.i_partition == D_8x16 )
86 int i_count = (i_refa == i_ref) + (i_refb == i_ref) + (i_refc == i_ref);
91 x264_median_mv( mvp, mv_a, mv_b, mv_c );
93 else if( i_count == 1 )
97 else if( i_refb == i_ref )
102 else if( i_refb == -2 && i_refc == -2 && i_refa != -2 )
108 void x264_mb_predict_mv_16x16( x264_t *h, int i_list, int i_ref, int16_t mvp[2] )
110 int i_refa = h->mb.cache.ref[i_list][X264_SCAN8_0 - 1];
111 int16_t *mv_a = h->mb.cache.mv[i_list][X264_SCAN8_0 - 1];
112 int i_refb = h->mb.cache.ref[i_list][X264_SCAN8_0 - 8];
113 int16_t *mv_b = h->mb.cache.mv[i_list][X264_SCAN8_0 - 8];
114 int i_refc = h->mb.cache.ref[i_list][X264_SCAN8_0 - 8 + 4];
115 int16_t *mv_c = h->mb.cache.mv[i_list][X264_SCAN8_0 - 8 + 4];
118 i_refc = h->mb.cache.ref[i_list][X264_SCAN8_0 - 8 - 1];
119 mv_c = h->mb.cache.mv[i_list][X264_SCAN8_0 - 8 - 1];
122 int i_count = (i_refa == i_ref) + (i_refb == i_ref) + (i_refc == i_ref);
127 x264_median_mv( mvp, mv_a, mv_b, mv_c );
129 else if( i_count == 1 )
131 if( i_refa == i_ref )
133 else if( i_refb == i_ref )
138 else if( i_refb == -2 && i_refc == -2 && i_refa != -2 )
145 void x264_mb_predict_mv_pskip( x264_t *h, int16_t mv[2] )
147 int i_refa = h->mb.cache.ref[0][X264_SCAN8_0 - 1];
148 int i_refb = h->mb.cache.ref[0][X264_SCAN8_0 - 8];
149 int16_t *mv_a = h->mb.cache.mv[0][X264_SCAN8_0 - 1];
150 int16_t *mv_b = h->mb.cache.mv[0][X264_SCAN8_0 - 8];
152 if( i_refa == -2 || i_refb == -2 ||
153 !( i_refa | M32( mv_a ) ) ||
154 !( i_refb | M32( mv_b ) ) )
159 x264_mb_predict_mv_16x16( h, 0, 0, mv );
162 static int x264_mb_predict_mv_direct16x16_temporal( x264_t *h )
164 int i_mb_4x4 = 16 * h->mb.i_mb_stride * h->mb.i_mb_y + 4 * h->mb.i_mb_x;
165 int i_mb_8x8 = 4 * h->mb.i_mb_stride * h->mb.i_mb_y + 2 * h->mb.i_mb_x;
166 const int type_col = h->fref1[0]->mb_type[h->mb.i_mb_xy];
167 const int partition_col = h->fref1[0]->mb_partition[h->mb.i_mb_xy];
169 x264_macroblock_cache_ref( h, 0, 0, 4, 4, 1, 0 );
171 h->mb.i_partition = partition_col;
173 if( IS_INTRA( type_col ) )
175 x264_macroblock_cache_ref( h, 0, 0, 4, 4, 0, 0 );
176 x264_macroblock_cache_mv( h, 0, 0, 4, 4, 0, 0 );
177 x264_macroblock_cache_mv( h, 0, 0, 4, 4, 1, 0 );
181 /* Don't do any checks other than the ones we have to, based
182 * on the size of the colocated partitions.
183 * Depends on the enum order: D_8x8, D_16x8, D_8x16, D_16x16 */
184 int max_i8 = (D_16x16 - partition_col) + 1;
185 int step = (partition_col == D_16x8) + 1;
186 int width = 4 >> ((D_16x16 - partition_col)&1);
187 int height = 4 >> ((D_16x16 - partition_col)>>1);
189 for( int i8 = 0; i8 < max_i8; i8 += step )
193 int i_part_8x8 = i_mb_8x8 + x8 + y8 * h->mb.i_b8_stride;
194 int i_ref1_ref = h->fref1[0]->ref[0][i_part_8x8];
195 int i_ref = (map_col_to_list0(i_ref1_ref>>h->sh.b_mbaff) << h->sh.b_mbaff) + (i_ref1_ref&h->sh.b_mbaff);
199 int dist_scale_factor = h->mb.dist_scale_factor[i_ref][0];
200 int16_t *mv_col = h->fref1[0]->mv[0][i_mb_4x4 + 3*x8 + 3*y8 * h->mb.i_b4_stride];
201 int l0x = ( dist_scale_factor * mv_col[0] + 128 ) >> 8;
202 int l0y = ( dist_scale_factor * mv_col[1] + 128 ) >> 8;
203 if( h->param.i_threads > 1 && (l0y > h->mb.mv_max_spel[1] || l0y-mv_col[1] > h->mb.mv_max_spel[1]) )
205 x264_macroblock_cache_ref( h, 2*x8, 2*y8, width, height, 0, i_ref );
206 x264_macroblock_cache_mv( h, 2*x8, 2*y8, width, height, 0, pack16to32_mask(l0x, l0y) );
207 x264_macroblock_cache_mv( h, 2*x8, 2*y8, width, height, 1, pack16to32_mask(l0x-mv_col[0], l0y-mv_col[1]) );
211 /* the collocated ref isn't in the current list0 */
212 /* FIXME: we might still be able to use direct_8x8 on some partitions */
213 /* FIXME: with B-pyramid + extensive ref list reordering
214 * (not currently used), we would also have to check
215 * l1mv1 like in spatial mode */
223 static int x264_mb_predict_mv_direct16x16_spatial( x264_t *h )
226 ALIGNED_ARRAY_8( int16_t, mv,[2],[2] );
227 const int8_t *l1ref0 = &h->fref1[0]->ref[0][h->mb.i_b8_xy];
228 const int8_t *l1ref1 = &h->fref1[0]->ref[1][h->mb.i_b8_xy];
229 const int16_t (*l1mv[2])[2] = { (const int16_t (*)[2]) &h->fref1[0]->mv[0][h->mb.i_b4_xy],
230 (const int16_t (*)[2]) &h->fref1[0]->mv[1][h->mb.i_b4_xy] };
231 const int type_col = h->fref1[0]->mb_type[h->mb.i_mb_xy];
232 const int partition_col = h->fref1[0]->mb_partition[h->mb.i_mb_xy];
234 h->mb.i_partition = partition_col;
236 for( int i_list = 0; i_list < 2; i_list++ )
238 int i_refa = h->mb.cache.ref[i_list][X264_SCAN8_0 - 1];
239 int16_t *mv_a = h->mb.cache.mv[i_list][X264_SCAN8_0 - 1];
240 int i_refb = h->mb.cache.ref[i_list][X264_SCAN8_0 - 8];
241 int16_t *mv_b = h->mb.cache.mv[i_list][X264_SCAN8_0 - 8];
242 int i_refc = h->mb.cache.ref[i_list][X264_SCAN8_0 - 8 + 4];
243 int16_t *mv_c = h->mb.cache.mv[i_list][X264_SCAN8_0 - 8 + 4];
246 i_refc = h->mb.cache.ref[i_list][X264_SCAN8_0 - 8 - 1];
247 mv_c = h->mb.cache.mv[i_list][X264_SCAN8_0 - 8 - 1];
250 int i_ref = X264_MIN3( (unsigned)i_refa, (unsigned)i_refb, (unsigned)i_refc );
254 M32( mv[i_list] ) = 0;
258 /* Same as x264_mb_predict_mv_16x16, but simplified to eliminate cases
259 * not relevant to spatial direct. */
260 int i_count = (i_refa == i_ref) + (i_refb == i_ref) + (i_refc == i_ref);
263 x264_median_mv( mv[i_list], mv_a, mv_b, mv_c );
266 if( i_refa == i_ref )
267 CP32( mv[i_list], mv_a );
268 else if( i_refb == i_ref )
269 CP32( mv[i_list], mv_b );
271 CP32( mv[i_list], mv_c );
275 x264_macroblock_cache_ref( h, 0, 0, 4, 4, i_list, i_ref );
276 x264_macroblock_cache_mv_ptr( h, 0, 0, 4, 4, i_list, mv[i_list] );
280 if( (M16( ref ) & 0x8080) == 0x8080 ) /* if( ref[0] < 0 && ref[1] < 0 ) */
282 x264_macroblock_cache_ref( h, 0, 0, 4, 4, 0, 0 );
283 x264_macroblock_cache_ref( h, 0, 0, 4, 4, 1, 0 );
287 if( h->param.i_threads > 1
288 && ( mv[0][1] > h->mb.mv_max_spel[1]
289 || mv[1][1] > h->mb.mv_max_spel[1] ) )
292 fprintf(stderr, "direct_spatial: (%d,%d) (%d,%d) > %d \n",
293 mv[0][0], mv[0][1], mv[1][0], mv[1][1],
294 h->mb.mv_max_spel[1]);
299 if( !M64( mv ) || IS_INTRA( type_col ) || (ref[0]&&ref[1]) )
302 /* Don't do any checks other than the ones we have to, based
303 * on the size of the colocated partitions.
304 * Depends on the enum order: D_8x8, D_16x8, D_8x16, D_16x16 */
305 int max_i8 = (D_16x16 - partition_col) + 1;
306 int step = (partition_col == D_16x8) + 1;
307 int width = 4 >> ((D_16x16 - partition_col)&1);
308 int height = 4 >> ((D_16x16 - partition_col)>>1);
311 for( int i8 = 0; i8 < max_i8; i8 += step )
314 const int y8 = i8>>1;
315 const int o8 = x8 + y8 * h->mb.i_b8_stride;
316 const int o4 = 3*(x8 + y8 * h->mb.i_b4_stride);
318 if( l1ref0[o8] == 0 )
320 else if( l1ref0[o8] < 0 && l1ref1[o8] == 0 )
325 if( abs( l1mv[idx][o4][0] ) <= 1 && abs( l1mv[idx][o4][1] ) <= 1 )
327 if( ref[0] == 0 ) x264_macroblock_cache_mv( h, 2*x8, 2*y8, width, height, 0, 0 );
328 if( ref[1] == 0 ) x264_macroblock_cache_mv( h, 2*x8, 2*y8, width, height, 1, 0 );
335 int x264_mb_predict_mv_direct16x16( x264_t *h, int *b_changed )
338 if( h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_NONE )
340 else if( h->sh.b_direct_spatial_mv_pred )
341 b_available = x264_mb_predict_mv_direct16x16_spatial( h );
343 b_available = x264_mb_predict_mv_direct16x16_temporal( h );
345 if( b_changed != NULL && b_available )
349 changed = M32( h->mb.cache.direct_mv[0][0] ) ^ M32( h->mb.cache.mv[0][x264_scan8[0]] );
350 changed |= M32( h->mb.cache.direct_mv[1][0] ) ^ M32( h->mb.cache.mv[1][x264_scan8[0]] );
351 changed |= h->mb.cache.direct_ref[0][0] ^ h->mb.cache.ref[0][x264_scan8[0]];
352 changed |= h->mb.cache.direct_ref[1][0] ^ h->mb.cache.ref[1][x264_scan8[0]];
353 if( !changed && h->mb.i_partition != D_16x16 )
355 changed |= M32( h->mb.cache.direct_mv[0][3] ) ^ M32( h->mb.cache.mv[0][x264_scan8[12]] );
356 changed |= M32( h->mb.cache.direct_mv[1][3] ) ^ M32( h->mb.cache.mv[1][x264_scan8[12]] );
357 changed |= h->mb.cache.direct_ref[0][3] ^ h->mb.cache.ref[0][x264_scan8[12]];
358 changed |= h->mb.cache.direct_ref[1][3] ^ h->mb.cache.ref[1][x264_scan8[12]];
360 if( !changed && h->mb.i_partition == D_8x8 )
362 changed |= M32( h->mb.cache.direct_mv[0][1] ) ^ M32( h->mb.cache.mv[0][x264_scan8[4]] );
363 changed |= M32( h->mb.cache.direct_mv[1][1] ) ^ M32( h->mb.cache.mv[1][x264_scan8[4]] );
364 changed |= M32( h->mb.cache.direct_mv[0][2] ) ^ M32( h->mb.cache.mv[0][x264_scan8[8]] );
365 changed |= M32( h->mb.cache.direct_mv[1][2] ) ^ M32( h->mb.cache.mv[1][x264_scan8[8]] );
366 changed |= h->mb.cache.direct_ref[0][1] ^ h->mb.cache.ref[0][x264_scan8[4]];
367 changed |= h->mb.cache.direct_ref[1][1] ^ h->mb.cache.ref[1][x264_scan8[4]];
368 changed |= h->mb.cache.direct_ref[0][2] ^ h->mb.cache.ref[0][x264_scan8[8]];
369 changed |= h->mb.cache.direct_ref[1][2] ^ h->mb.cache.ref[1][x264_scan8[8]];
371 *b_changed = changed;
378 for( int l = 0; l < 2; l++ )
380 CP32( h->mb.cache.direct_mv[l][0], h->mb.cache.mv[l][x264_scan8[ 0]] );
381 CP32( h->mb.cache.direct_mv[l][1], h->mb.cache.mv[l][x264_scan8[ 4]] );
382 CP32( h->mb.cache.direct_mv[l][2], h->mb.cache.mv[l][x264_scan8[ 8]] );
383 CP32( h->mb.cache.direct_mv[l][3], h->mb.cache.mv[l][x264_scan8[12]] );
384 h->mb.cache.direct_ref[l][0] = h->mb.cache.ref[l][x264_scan8[ 0]];
385 h->mb.cache.direct_ref[l][1] = h->mb.cache.ref[l][x264_scan8[ 4]];
386 h->mb.cache.direct_ref[l][2] = h->mb.cache.ref[l][x264_scan8[ 8]];
387 h->mb.cache.direct_ref[l][3] = h->mb.cache.ref[l][x264_scan8[12]];
388 h->mb.cache.direct_partition = h->mb.i_partition;
394 /* This just improves encoder performance, it's not part of the spec */
395 void x264_mb_predict_mv_ref16x16( x264_t *h, int i_list, int i_ref, int16_t mvc[9][2], int *i_mvc )
397 int16_t (*mvr)[2] = h->mb.mvr[i_list][i_ref];
400 #define SET_MVP(mvp) \
402 CP32( mvc[i], mvp ); \
407 if( h->sh.i_type == SLICE_TYPE_B
408 && h->mb.cache.ref[i_list][x264_scan8[12]] == i_ref )
410 SET_MVP( h->mb.cache.mv[i_list][x264_scan8[12]] );
413 if( i_ref == 0 && h->frames.b_have_lowres )
415 int idx = i_list ? h->fref1[0]->i_frame-h->fenc->i_frame-1
416 : h->fenc->i_frame-h->fref0[0]->i_frame-1;
417 if( idx <= h->param.i_bframe )
419 int16_t (*lowres_mv)[2] = h->fenc->lowres_mvs[i_list][idx];
420 if( lowres_mv[0][0] != 0x7fff )
422 M32( mvc[i] ) = (M32( lowres_mv[h->mb.i_mb_xy] )*2)&0xfffeffff;
428 /* spatial predictors */
429 SET_MVP( mvr[h->mb.i_mb_left_xy] );
430 SET_MVP( mvr[h->mb.i_mb_top_xy] );
431 SET_MVP( mvr[h->mb.i_mb_topleft_xy] );
432 SET_MVP( mvr[h->mb.i_mb_topright_xy] );
435 /* temporal predictors */
436 if( h->fref0[0]->i_ref[0] > 0 )
438 x264_frame_t *l0 = h->fref0[0];
439 x264_frame_t **fref = i_list ? h->fref1 : h->fref0;
440 int field = h->mb.i_mb_y&1;
441 int curpoc = h->fdec->i_poc + field*h->sh.i_delta_poc_bottom;
442 int refpoc = fref[i_ref>>h->sh.b_mbaff]->i_poc;
443 if( h->sh.b_mbaff && field^(i_ref&1) )
444 refpoc += h->sh.i_delta_poc_bottom;
446 #define SET_TMVP( dx, dy ) \
448 int mb_index = h->mb.i_mb_xy + dx + dy*h->mb.i_mb_stride; \
449 int scale = (curpoc - refpoc) * l0->inv_ref_poc[h->mb.b_interlaced&field]; \
450 mvc[i][0] = (l0->mv16x16[mb_index][0]*scale + 128) >> 8; \
451 mvc[i][1] = (l0->mv16x16[mb_index][1]*scale + 128) >> 8; \
456 if( h->mb.i_mb_x < h->mb.i_mb_width-1 )
458 if( h->mb.i_mb_y < h->mb.i_mb_height-1 )