1 /*****************************************************************************
2 * macroblock.c: h264 encoder library
3 *****************************************************************************
4 * Copyright (C) 2003-2008 x264 project
6 * Authors: Laurent Aimar <fenrir@via.ecp.fr>
7 * Loren Merritt <lorenm@u.washington.edu>
8 * Fiona Glaser <fiona@x264.com>
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.
23 *****************************************************************************/
26 #include "encoder/me.h"
28 void x264_mb_predict_mv( x264_t *h, int i_list, int idx, int i_width, int16_t mvp[2] )
30 const int i8 = x264_scan8[idx];
31 const int i_ref= h->mb.cache.ref[i_list][i8];
32 int i_refa = h->mb.cache.ref[i_list][i8 - 1];
33 int16_t *mv_a = h->mb.cache.mv[i_list][i8 - 1];
34 int i_refb = h->mb.cache.ref[i_list][i8 - 8];
35 int16_t *mv_b = h->mb.cache.mv[i_list][i8 - 8];
36 int i_refc = h->mb.cache.ref[i_list][i8 - 8 + i_width];
37 int16_t *mv_c = h->mb.cache.mv[i_list][i8 - 8 + i_width];
39 if( (idx&3) >= 2 + (i_width&1) || i_refc == -2 )
41 i_refc = h->mb.cache.ref[i_list][i8 - 8 - 1];
42 mv_c = h->mb.cache.mv[i_list][i8 - 8 - 1];
45 if( h->mb.i_partition == D_16x8 )
64 else if( h->mb.i_partition == D_8x16 )
84 int i_count = (i_refa == i_ref) + (i_refb == i_ref) + (i_refc == i_ref);
89 x264_median_mv( mvp, mv_a, mv_b, mv_c );
91 else if( i_count == 1 )
95 else if( i_refb == i_ref )
100 else if( i_refb == -2 && i_refc == -2 && i_refa != -2 )
106 void x264_mb_predict_mv_16x16( x264_t *h, int i_list, int i_ref, int16_t mvp[2] )
108 int i_refa = h->mb.cache.ref[i_list][X264_SCAN8_0 - 1];
109 int16_t *mv_a = h->mb.cache.mv[i_list][X264_SCAN8_0 - 1];
110 int i_refb = h->mb.cache.ref[i_list][X264_SCAN8_0 - 8];
111 int16_t *mv_b = h->mb.cache.mv[i_list][X264_SCAN8_0 - 8];
112 int i_refc = h->mb.cache.ref[i_list][X264_SCAN8_0 - 8 + 4];
113 int16_t *mv_c = h->mb.cache.mv[i_list][X264_SCAN8_0 - 8 + 4];
116 i_refc = h->mb.cache.ref[i_list][X264_SCAN8_0 - 8 - 1];
117 mv_c = h->mb.cache.mv[i_list][X264_SCAN8_0 - 8 - 1];
120 int i_count = (i_refa == i_ref) + (i_refb == i_ref) + (i_refc == i_ref);
125 x264_median_mv( mvp, mv_a, mv_b, mv_c );
127 else if( i_count == 1 )
129 if( i_refa == i_ref )
131 else if( i_refb == i_ref )
136 else if( i_refb == -2 && i_refc == -2 && i_refa != -2 )
143 void x264_mb_predict_mv_pskip( x264_t *h, int16_t mv[2] )
145 int i_refa = h->mb.cache.ref[0][X264_SCAN8_0 - 1];
146 int i_refb = h->mb.cache.ref[0][X264_SCAN8_0 - 8];
147 int16_t *mv_a = h->mb.cache.mv[0][X264_SCAN8_0 - 1];
148 int16_t *mv_b = h->mb.cache.mv[0][X264_SCAN8_0 - 8];
150 if( i_refa == -2 || i_refb == -2 ||
151 !( i_refa | M32( mv_a ) ) ||
152 !( i_refb | M32( mv_b ) ) )
158 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;
167 const int type_col = h->fref1[0]->mb_type[h->mb.i_mb_xy];
168 const int partition_col = h->fref1[0]->mb_partition[h->mb.i_mb_xy];
170 x264_macroblock_cache_ref( h, 0, 0, 4, 4, 1, 0 );
172 h->mb.i_partition = partition_col;
174 if( IS_INTRA( type_col ) )
176 x264_macroblock_cache_ref( h, 0, 0, 4, 4, 0, 0 );
177 x264_macroblock_cache_mv( h, 0, 0, 4, 4, 0, 0 );
178 x264_macroblock_cache_mv( h, 0, 0, 4, 4, 1, 0 );
182 /* Don't do any checks other than the ones we have to, based
183 * on the size of the colocated partitions.
184 * Depends on the enum order: D_8x8, D_16x8, D_8x16, D_16x16 */
185 int max_i8 = (D_16x16 - partition_col) + 1;
186 int step = (partition_col == D_16x8) + 1;
187 int width = 4 >> ((D_16x16 - partition_col)&1);
188 int height = 4 >> ((D_16x16 - partition_col)>>1);
190 for( i8 = 0; i8 < max_i8; i8 += step )
194 const int i_part_8x8 = i_mb_8x8 + x8 + y8 * h->mb.i_b8_stride;
195 const int i_ref1_ref = h->fref1[0]->ref[0][i_part_8x8];
196 const 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);
200 const int dist_scale_factor = h->mb.dist_scale_factor[i_ref][0];
201 const int16_t *mv_col = h->fref1[0]->mv[0][i_mb_4x4 + 3*x8 + 3*y8 * h->mb.i_b4_stride];
202 const int l0x = ( dist_scale_factor * mv_col[0] + 128 ) >> 8;
203 const int l0y = ( dist_scale_factor * mv_col[1] + 128 ) >> 8;
204 if( h->param.i_threads > 1 && (l0y > h->mb.mv_max_spel[1] || l0y-mv_col[1] > h->mb.mv_max_spel[1]) )
206 x264_macroblock_cache_ref( h, 2*x8, 2*y8, width, height, 0, i_ref );
207 x264_macroblock_cache_mv( h, 2*x8, 2*y8, width, height, 0, pack16to32_mask(l0x, l0y) );
208 x264_macroblock_cache_mv( h, 2*x8, 2*y8, width, height, 1, pack16to32_mask(l0x-mv_col[0], l0y-mv_col[1]) );
212 /* the collocated ref isn't in the current list0 */
213 /* FIXME: we might still be able to use direct_8x8 on some partitions */
214 /* FIXME: with B-pyramid + extensive ref list reordering
215 * (not currently used), we would also have to check
216 * l1mv1 like in spatial mode */
224 static int x264_mb_predict_mv_direct16x16_spatial( x264_t *h )
227 ALIGNED_ARRAY_8( int16_t, mv,[2],[2] );
228 int i_list, i8, i_ref;
229 const int8_t *l1ref0 = &h->fref1[0]->ref[0][h->mb.i_b8_xy];
230 const int8_t *l1ref1 = &h->fref1[0]->ref[1][h->mb.i_b8_xy];
231 const int16_t (*l1mv[2])[2] = { (const int16_t (*)[2]) &h->fref1[0]->mv[0][h->mb.i_b4_xy],
232 (const int16_t (*)[2]) &h->fref1[0]->mv[1][h->mb.i_b4_xy] };
233 const int type_col = h->fref1[0]->mb_type[h->mb.i_mb_xy];
234 const int partition_col = h->fref1[0]->mb_partition[h->mb.i_mb_xy];
236 h->mb.i_partition = partition_col;
238 for( i_list = 0; i_list < 2; i_list++ )
240 int i_refa = h->mb.cache.ref[i_list][X264_SCAN8_0 - 1];
241 int16_t *mv_a = h->mb.cache.mv[i_list][X264_SCAN8_0 - 1];
242 int i_refb = h->mb.cache.ref[i_list][X264_SCAN8_0 - 8];
243 int16_t *mv_b = h->mb.cache.mv[i_list][X264_SCAN8_0 - 8];
244 int i_refc = h->mb.cache.ref[i_list][X264_SCAN8_0 - 8 + 4];
245 int16_t *mv_c = h->mb.cache.mv[i_list][X264_SCAN8_0 - 8 + 4];
248 i_refc = h->mb.cache.ref[i_list][X264_SCAN8_0 - 8 - 1];
249 mv_c = h->mb.cache.mv[i_list][X264_SCAN8_0 - 8 - 1];
252 i_ref = X264_MIN3( (unsigned)i_refa, (unsigned)i_refb, (unsigned)i_refc );
256 M32( mv[i_list] ) = 0;
260 /* Same as x264_mb_predict_mv_16x16, but simplified to eliminate cases
261 * not relevant to spatial direct. */
262 int i_count = (i_refa == i_ref) + (i_refb == i_ref) + (i_refc == i_ref);
265 x264_median_mv( mv[i_list], mv_a, mv_b, mv_c );
268 if( i_refa == i_ref )
269 CP32( mv[i_list], mv_a );
270 else if( i_refb == i_ref )
271 CP32( mv[i_list], mv_b );
273 CP32( mv[i_list], mv_c );
277 x264_macroblock_cache_ref( h, 0, 0, 4, 4, i_list, i_ref );
278 x264_macroblock_cache_mv_ptr( h, 0, 0, 4, 4, i_list, mv[i_list] );
282 if( (M16( ref ) & 0x8080) == 0x8080 ) /* if( ref[0] < 0 && ref[1] < 0 ) */
284 x264_macroblock_cache_ref( h, 0, 0, 4, 4, 0, 0 );
285 x264_macroblock_cache_ref( h, 0, 0, 4, 4, 1, 0 );
289 if( !M64( mv ) || IS_INTRA( type_col ) || (ref[0]&&ref[1]) )
292 if( h->param.i_threads > 1
293 && ( mv[0][1] > h->mb.mv_max_spel[1]
294 || mv[1][1] > h->mb.mv_max_spel[1] ) )
297 fprintf(stderr, "direct_spatial: (%d,%d) (%d,%d) > %d \n",
298 mv[0][0], mv[0][1], mv[1][0], mv[1][1],
299 h->mb.mv_max_spel[1]);
304 /* Don't do any checks other than the ones we have to, based
305 * on the size of the colocated partitions.
306 * Depends on the enum order: D_8x8, D_16x8, D_8x16, D_16x16 */
307 int max_i8 = (D_16x16 - partition_col) + 1;
308 int step = (partition_col == D_16x8) + 1;
309 int width = 4 >> ((D_16x16 - partition_col)&1);
310 int height = 4 >> ((D_16x16 - partition_col)>>1);
313 for( i8 = 0; i8 < max_i8; i8 += step )
316 const int y8 = i8>>1;
317 const int o8 = x8 + y8 * h->mb.i_b8_stride;
318 const int o4 = 3*(x8 + y8 * h->mb.i_b4_stride);
320 if( l1ref0[o8] == 0 )
322 else if( l1ref0[o8] < 0 && l1ref1[o8] == 0 )
327 if( abs( l1mv[idx][o4][0] ) <= 1 && abs( l1mv[idx][o4][1] ) <= 1 )
329 if( ref[0] == 0 ) x264_macroblock_cache_mv( h, 2*x8, 2*y8, width, height, 0, 0 );
330 if( ref[1] == 0 ) x264_macroblock_cache_mv( h, 2*x8, 2*y8, width, height, 1, 0 );
337 int x264_mb_predict_mv_direct16x16( x264_t *h, int *b_changed )
340 if( h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_NONE )
342 else if( h->sh.b_direct_spatial_mv_pred )
343 b_available = x264_mb_predict_mv_direct16x16_spatial( h );
345 b_available = x264_mb_predict_mv_direct16x16_temporal( h );
347 if( b_changed != NULL && b_available )
351 changed = M32( h->mb.cache.direct_mv[0][0] ) ^ M32( h->mb.cache.mv[0][x264_scan8[0]] );
352 changed |= M32( h->mb.cache.direct_mv[1][0] ) ^ M32( h->mb.cache.mv[1][x264_scan8[0]] );
353 changed |= h->mb.cache.direct_ref[0][0] ^ h->mb.cache.ref[0][x264_scan8[0]];
354 changed |= h->mb.cache.direct_ref[1][0] ^ h->mb.cache.ref[1][x264_scan8[0]];
355 if( !changed && h->mb.i_partition != D_16x16 )
357 changed |= M32( h->mb.cache.direct_mv[0][3] ) ^ M32( h->mb.cache.mv[0][x264_scan8[12]] );
358 changed |= M32( h->mb.cache.direct_mv[1][3] ) ^ M32( h->mb.cache.mv[1][x264_scan8[12]] );
359 changed |= h->mb.cache.direct_ref[0][3] ^ h->mb.cache.ref[0][x264_scan8[12]];
360 changed |= h->mb.cache.direct_ref[1][3] ^ h->mb.cache.ref[1][x264_scan8[12]];
362 if( !changed && h->mb.i_partition == D_8x8 )
364 changed |= M32( h->mb.cache.direct_mv[0][1] ) ^ M32( h->mb.cache.mv[0][x264_scan8[4]] );
365 changed |= M32( h->mb.cache.direct_mv[1][1] ) ^ M32( h->mb.cache.mv[1][x264_scan8[4]] );
366 changed |= M32( h->mb.cache.direct_mv[0][2] ) ^ M32( h->mb.cache.mv[0][x264_scan8[8]] );
367 changed |= M32( h->mb.cache.direct_mv[1][2] ) ^ M32( h->mb.cache.mv[1][x264_scan8[8]] );
368 changed |= h->mb.cache.direct_ref[0][1] ^ h->mb.cache.ref[0][x264_scan8[4]];
369 changed |= h->mb.cache.direct_ref[1][1] ^ h->mb.cache.ref[1][x264_scan8[4]];
370 changed |= h->mb.cache.direct_ref[0][2] ^ h->mb.cache.ref[0][x264_scan8[8]];
371 changed |= h->mb.cache.direct_ref[1][2] ^ h->mb.cache.ref[1][x264_scan8[8]];
373 *b_changed = changed;
382 for( l = 0; l < 2; l++ )
384 CP32( h->mb.cache.direct_mv[l][0], h->mb.cache.mv[l][x264_scan8[ 0]] );
385 CP32( h->mb.cache.direct_mv[l][1], h->mb.cache.mv[l][x264_scan8[ 4]] );
386 CP32( h->mb.cache.direct_mv[l][2], h->mb.cache.mv[l][x264_scan8[ 8]] );
387 CP32( h->mb.cache.direct_mv[l][3], h->mb.cache.mv[l][x264_scan8[12]] );
388 h->mb.cache.direct_ref[l][0] = h->mb.cache.ref[l][x264_scan8[ 0]];
389 h->mb.cache.direct_ref[l][1] = h->mb.cache.ref[l][x264_scan8[ 4]];
390 h->mb.cache.direct_ref[l][2] = h->mb.cache.ref[l][x264_scan8[ 8]];
391 h->mb.cache.direct_ref[l][3] = h->mb.cache.ref[l][x264_scan8[12]];
392 h->mb.cache.direct_partition = h->mb.i_partition;
399 /* This just improves encoder performance, it's not part of the spec */
400 void x264_mb_predict_mv_ref16x16( x264_t *h, int i_list, int i_ref, int16_t mvc[9][2], int *i_mvc )
402 int16_t (*mvr)[2] = h->mb.mvr[i_list][i_ref];
405 #define SET_MVP(mvp) { \
406 CP32( mvc[i], mvp ); \
411 if( h->sh.i_type == SLICE_TYPE_B
412 && h->mb.cache.ref[i_list][x264_scan8[12]] == i_ref )
414 SET_MVP( h->mb.cache.mv[i_list][x264_scan8[12]] );
417 if( i_ref == 0 && h->frames.b_have_lowres )
419 int16_t (*lowres_mv)[2] = i_list ? h->fenc->lowres_mvs[1][h->fref1[0]->i_frame-h->fenc->i_frame-1]
420 : h->fenc->lowres_mvs[0][h->fenc->i_frame-h->fref0[0]->i_frame-1];
421 if( lowres_mv[0][0] != 0x7fff )
423 M32( mvc[i] ) = (M32( lowres_mv[h->mb.i_mb_xy] )*2)&0xfffeffff;
428 /* spatial predictors */
429 if( h->mb.i_neighbour & MB_LEFT )
431 int i_mb_l = h->mb.i_mb_xy - 1;
432 /* skip MBs didn't go through the whole search process, so mvr is undefined */
433 if( !IS_SKIP( h->mb.type[i_mb_l] ) )
434 SET_MVP( mvr[i_mb_l] );
436 if( h->mb.i_neighbour & MB_TOP )
438 int i_mb_t = h->mb.i_mb_top_xy;
439 if( !IS_SKIP( h->mb.type[i_mb_t] ) )
440 SET_MVP( mvr[i_mb_t] );
442 if( h->mb.i_neighbour & MB_TOPLEFT && !IS_SKIP( h->mb.type[i_mb_t - 1] ) )
443 SET_MVP( mvr[i_mb_t-1] );
444 if( h->mb.i_mb_x < h->mb.i_mb_stride - 1 && !IS_SKIP( h->mb.type[i_mb_t + 1] ) )
445 SET_MVP( mvr[i_mb_t+1] );
449 /* temporal predictors */
450 /* FIXME temporal scaling w/ interlace */
451 if( h->fref0[0]->i_ref[0] > 0 && !h->sh.b_mbaff )
453 x264_frame_t *l0 = h->fref0[0];
455 #define SET_TMVP(dx, dy) { \
456 int i_b4 = h->mb.i_b4_xy + dx*4 + dy*4*h->mb.i_b4_stride; \
457 int i_b8 = h->mb.i_b8_xy + dx*2 + dy*2*h->mb.i_b8_stride; \
458 int ref_col = l0->ref[0][i_b8]; \
461 int scale = (h->fdec->i_poc - h->fdec->ref_poc[0][i_ref]) * l0->inv_ref_poc[ref_col];\
462 mvc[i][0] = (l0->mv[0][i_b4][0]*scale + 128) >> 8;\
463 mvc[i][1] = (l0->mv[0][i_b4][1]*scale + 128) >> 8;\
469 if( h->mb.i_mb_x < h->sps->i_mb_width-1 )
471 if( h->mb.i_mb_y < h->sps->i_mb_height-1 )
479 /* Set up a lookup table for delta pocs to reduce an IDIV to an IMUL */
480 static void setup_inverse_delta_pocs( x264_t *h )
483 for( i = 0; i < h->i_ref0; i++ )
485 int delta = h->fdec->i_poc - h->fref0[i]->i_poc;
486 h->fdec->inv_ref_poc[i] = (256 + delta/2) / delta;
490 static NOINLINE void x264_mb_mc_0xywh( x264_t *h, int x, int y, int width, int height )
492 const int i8 = x264_scan8[0]+x+8*y;
493 const int i_ref = h->mb.cache.ref[0][i8];
494 const int mvx = x264_clip3( h->mb.cache.mv[0][i8][0], h->mb.mv_min[0], h->mb.mv_max[0] ) + 4*4*x;
495 int mvy = x264_clip3( h->mb.cache.mv[0][i8][1], h->mb.mv_min[1], h->mb.mv_max[1] ) + 4*4*y;
497 h->mc.mc_luma( &h->mb.pic.p_fdec[0][4*y*FDEC_STRIDE+4*x], FDEC_STRIDE,
498 h->mb.pic.p_fref[0][i_ref], h->mb.pic.i_stride[0],
499 mvx, mvy, 4*width, 4*height, &h->sh.weight[i_ref][0] );
501 // chroma is offset if MCing from a field of opposite parity
502 if( h->mb.b_interlaced & i_ref )
503 mvy += (h->mb.i_mb_y & 1)*4 - 2;
505 h->mc.mc_chroma( &h->mb.pic.p_fdec[1][2*y*FDEC_STRIDE+2*x], FDEC_STRIDE,
506 h->mb.pic.p_fref[0][i_ref][4], h->mb.pic.i_stride[1],
507 mvx, mvy, 2*width, 2*height );
509 if( h->sh.weight[i_ref][1].weightfn )
510 h->sh.weight[i_ref][1].weightfn[width>>1]( &h->mb.pic.p_fdec[1][2*y*FDEC_STRIDE+2*x], FDEC_STRIDE,
511 &h->mb.pic.p_fdec[1][2*y*FDEC_STRIDE+2*x], FDEC_STRIDE,
512 &h->sh.weight[i_ref][1], height*2 );
514 h->mc.mc_chroma( &h->mb.pic.p_fdec[2][2*y*FDEC_STRIDE+2*x], FDEC_STRIDE,
515 h->mb.pic.p_fref[0][i_ref][5], h->mb.pic.i_stride[2],
516 mvx, mvy, 2*width, 2*height );
518 if( h->sh.weight[i_ref][2].weightfn )
519 h->sh.weight[i_ref][2].weightfn[width>>1]( &h->mb.pic.p_fdec[2][2*y*FDEC_STRIDE+2*x], FDEC_STRIDE,
520 &h->mb.pic.p_fdec[2][2*y*FDEC_STRIDE+2*x], FDEC_STRIDE,
521 &h->sh.weight[i_ref][2],height*2 );
524 static NOINLINE void x264_mb_mc_1xywh( x264_t *h, int x, int y, int width, int height )
526 const int i8 = x264_scan8[0]+x+8*y;
527 const int i_ref = h->mb.cache.ref[1][i8];
528 const int mvx = x264_clip3( h->mb.cache.mv[1][i8][0], h->mb.mv_min[0], h->mb.mv_max[0] ) + 4*4*x;
529 int mvy = x264_clip3( h->mb.cache.mv[1][i8][1], h->mb.mv_min[1], h->mb.mv_max[1] ) + 4*4*y;
531 h->mc.mc_luma( &h->mb.pic.p_fdec[0][4*y*FDEC_STRIDE+4*x], FDEC_STRIDE,
532 h->mb.pic.p_fref[1][i_ref], h->mb.pic.i_stride[0],
533 mvx, mvy, 4*width, 4*height, weight_none );
535 if( h->mb.b_interlaced & i_ref )
536 mvy += (h->mb.i_mb_y & 1)*4 - 2;
538 h->mc.mc_chroma( &h->mb.pic.p_fdec[1][2*y*FDEC_STRIDE+2*x], FDEC_STRIDE,
539 h->mb.pic.p_fref[1][i_ref][4], h->mb.pic.i_stride[1],
540 mvx, mvy, 2*width, 2*height );
542 h->mc.mc_chroma( &h->mb.pic.p_fdec[2][2*y*FDEC_STRIDE+2*x], FDEC_STRIDE,
543 h->mb.pic.p_fref[1][i_ref][5], h->mb.pic.i_stride[2],
544 mvx, mvy, 2*width, 2*height );
547 static NOINLINE void x264_mb_mc_01xywh( x264_t *h, int x, int y, int width, int height )
549 const int i8 = x264_scan8[0]+x+8*y;
550 const int i_ref0 = h->mb.cache.ref[0][i8];
551 const int i_ref1 = h->mb.cache.ref[1][i8];
552 const int weight = h->mb.bipred_weight[i_ref0][i_ref1];
553 const int mvx0 = x264_clip3( h->mb.cache.mv[0][i8][0], h->mb.mv_min[0], h->mb.mv_max[0] ) + 4*4*x;
554 const int mvx1 = x264_clip3( h->mb.cache.mv[1][i8][0], h->mb.mv_min[0], h->mb.mv_max[0] ) + 4*4*x;
555 int mvy0 = x264_clip3( h->mb.cache.mv[0][i8][1], h->mb.mv_min[1], h->mb.mv_max[1] ) + 4*4*y;
556 int mvy1 = x264_clip3( h->mb.cache.mv[1][i8][1], h->mb.mv_min[1], h->mb.mv_max[1] ) + 4*4*y;
557 int i_mode = x264_size2pixel[height][width];
558 int i_stride0 = 16, i_stride1 = 16;
559 ALIGNED_ARRAY_16( uint8_t, tmp0,[16*16] );
560 ALIGNED_ARRAY_16( uint8_t, tmp1,[16*16] );
561 uint8_t *src0, *src1;
563 src0 = h->mc.get_ref( tmp0, &i_stride0, h->mb.pic.p_fref[0][i_ref0], h->mb.pic.i_stride[0],
564 mvx0, mvy0, 4*width, 4*height, weight_none );
565 src1 = h->mc.get_ref( tmp1, &i_stride1, h->mb.pic.p_fref[1][i_ref1], h->mb.pic.i_stride[0],
566 mvx1, mvy1, 4*width, 4*height, weight_none );
567 h->mc.avg[i_mode]( &h->mb.pic.p_fdec[0][4*y*FDEC_STRIDE+4*x], FDEC_STRIDE,
568 src0, i_stride0, src1, i_stride1, weight );
570 if( h->mb.b_interlaced & i_ref0 )
571 mvy0 += (h->mb.i_mb_y & 1)*4 - 2;
572 if( h->mb.b_interlaced & i_ref1 )
573 mvy1 += (h->mb.i_mb_y & 1)*4 - 2;
575 h->mc.mc_chroma( tmp0, 16, h->mb.pic.p_fref[0][i_ref0][4], h->mb.pic.i_stride[1],
576 mvx0, mvy0, 2*width, 2*height );
577 h->mc.mc_chroma( tmp1, 16, h->mb.pic.p_fref[1][i_ref1][4], h->mb.pic.i_stride[1],
578 mvx1, mvy1, 2*width, 2*height );
579 h->mc.avg[i_mode+3]( &h->mb.pic.p_fdec[1][2*y*FDEC_STRIDE+2*x], FDEC_STRIDE, tmp0, 16, tmp1, 16, weight );
580 h->mc.mc_chroma( tmp0, 16, h->mb.pic.p_fref[0][i_ref0][5], h->mb.pic.i_stride[2],
581 mvx0, mvy0, 2*width, 2*height );
582 h->mc.mc_chroma( tmp1, 16, h->mb.pic.p_fref[1][i_ref1][5], h->mb.pic.i_stride[2],
583 mvx1, mvy1, 2*width, 2*height );
584 h->mc.avg[i_mode+3]( &h->mb.pic.p_fdec[2][2*y*FDEC_STRIDE+2*x], FDEC_STRIDE, tmp0, 16, tmp1, 16, weight );
587 void x264_mb_mc_8x8( x264_t *h, int i8 )
589 const int x = 2*(i8&1);
590 const int y = 2*(i8>>1);
592 if( h->sh.i_type == SLICE_TYPE_P )
594 switch( h->mb.i_sub_partition[i8] )
597 x264_mb_mc_0xywh( h, x, y, 2, 2 );
600 x264_mb_mc_0xywh( h, x, y+0, 2, 1 );
601 x264_mb_mc_0xywh( h, x, y+1, 2, 1 );
604 x264_mb_mc_0xywh( h, x+0, y, 1, 2 );
605 x264_mb_mc_0xywh( h, x+1, y, 1, 2 );
608 x264_mb_mc_0xywh( h, x+0, y+0, 1, 1 );
609 x264_mb_mc_0xywh( h, x+1, y+0, 1, 1 );
610 x264_mb_mc_0xywh( h, x+0, y+1, 1, 1 );
611 x264_mb_mc_0xywh( h, x+1, y+1, 1, 1 );
617 const int i8 = x264_scan8[0] + x + 8*y;
619 if( h->mb.cache.ref[0][i8] >= 0 )
620 if( h->mb.cache.ref[1][i8] >= 0 )
621 x264_mb_mc_01xywh( h, x, y, 2, 2 );
623 x264_mb_mc_0xywh( h, x, y, 2, 2 );
625 x264_mb_mc_1xywh( h, x, y, 2, 2 );
629 void x264_mb_mc( x264_t *h )
631 if( h->mb.i_partition == D_8x8 )
634 for( i = 0; i < 4; i++ )
635 x264_mb_mc_8x8( h, i );
639 const int ref0a = h->mb.cache.ref[0][x264_scan8[ 0]];
640 const int ref0b = h->mb.cache.ref[0][x264_scan8[12]];
641 const int ref1a = h->mb.cache.ref[1][x264_scan8[ 0]];
642 const int ref1b = h->mb.cache.ref[1][x264_scan8[12]];
644 if( h->mb.i_partition == D_16x16 )
647 if( ref1a >= 0 ) x264_mb_mc_01xywh( h, 0, 0, 4, 4 );
648 else x264_mb_mc_0xywh ( h, 0, 0, 4, 4 );
649 else x264_mb_mc_1xywh ( h, 0, 0, 4, 4 );
651 else if( h->mb.i_partition == D_16x8 )
654 if( ref1a >= 0 ) x264_mb_mc_01xywh( h, 0, 0, 4, 2 );
655 else x264_mb_mc_0xywh ( h, 0, 0, 4, 2 );
656 else x264_mb_mc_1xywh ( h, 0, 0, 4, 2 );
659 if( ref1b >= 0 ) x264_mb_mc_01xywh( h, 0, 2, 4, 2 );
660 else x264_mb_mc_0xywh ( h, 0, 2, 4, 2 );
661 else x264_mb_mc_1xywh ( h, 0, 2, 4, 2 );
663 else if( h->mb.i_partition == D_8x16 )
666 if( ref1a >= 0 ) x264_mb_mc_01xywh( h, 0, 0, 2, 4 );
667 else x264_mb_mc_0xywh ( h, 0, 0, 2, 4 );
668 else x264_mb_mc_1xywh ( h, 0, 0, 2, 4 );
671 if( ref1b >= 0 ) x264_mb_mc_01xywh( h, 2, 0, 2, 4 );
672 else x264_mb_mc_0xywh ( h, 2, 0, 2, 4 );
673 else x264_mb_mc_1xywh ( h, 2, 0, 2, 4 );
678 int x264_macroblock_cache_init( x264_t *h )
681 int i_mb_count = h->mb.i_mb_count;
683 h->mb.i_mb_stride = h->sps->i_mb_width;
684 h->mb.i_b8_stride = h->sps->i_mb_width * 2;
685 h->mb.i_b4_stride = h->sps->i_mb_width * 4;
687 h->mb.b_interlaced = h->param.b_interlaced;
689 CHECKED_MALLOC( h->mb.qp, i_mb_count * sizeof(int8_t) );
690 CHECKED_MALLOC( h->mb.cbp, i_mb_count * sizeof(int16_t) );
691 CHECKED_MALLOC( h->mb.skipbp, i_mb_count * sizeof(int8_t) );
692 CHECKED_MALLOC( h->mb.mb_transform_size, i_mb_count * sizeof(int8_t) );
694 /* 0 -> 3 top(4), 4 -> 6 : left(3) */
695 CHECKED_MALLOC( h->mb.intra4x4_pred_mode, i_mb_count * 8 * sizeof(int8_t) );
698 CHECKED_MALLOC( h->mb.non_zero_count, i_mb_count * 24 * sizeof(uint8_t) );
700 if( h->param.b_cabac )
702 CHECKED_MALLOC( h->mb.chroma_pred_mode, i_mb_count * sizeof(int8_t) );
703 CHECKED_MALLOC( h->mb.mvd[0], 2*16 * i_mb_count * sizeof(int16_t) );
704 CHECKED_MALLOC( h->mb.mvd[1], 2*16 * i_mb_count * sizeof(int16_t) );
709 int i_refs = X264_MIN(16, (i ? 1 + !!h->param.i_bframe_pyramid : h->param.i_frame_reference) ) << h->param.b_interlaced;
710 if( h->param.analyse.i_weighted_pred == X264_WEIGHTP_SMART )
711 i_refs = X264_MIN(16, i_refs + 2); //smart weights add two duplicate frames
712 else if( h->param.analyse.i_weighted_pred == X264_WEIGHTP_BLIND )
713 i_refs = X264_MIN(16, i_refs + 1); //blind weights add one duplicate frame
715 for( j=0; j < i_refs; j++ )
716 CHECKED_MALLOC( h->mb.mvr[i][j], 2 * i_mb_count * sizeof(int16_t) );
719 if( h->param.analyse.i_weighted_pred )
721 int i_padv = PADV << h->param.b_interlaced;
722 #define ALIGN(x,a) (((x)+((a)-1))&~((a)-1))
723 int align = h->param.cpu&X264_CPU_CACHELINE_64 ? 64 : h->param.cpu&X264_CPU_CACHELINE_32 ? 32 : 16;
724 int i_stride, luma_plane_size;
727 if( h->param.analyse.i_weighted_pred == X264_WEIGHTP_FAKE )
729 // only need buffer for lookahead
730 if( !h->param.i_sync_lookahead || h == h->thread[h->param.i_threads] )
732 // Fake analysis only works on lowres
733 i_stride = ALIGN( h->sps->i_mb_width*8 + 2*PADH, align );
734 luma_plane_size = i_stride * (h->sps->i_mb_height*8+2*i_padv);
735 // Only need 1 buffer for analysis
743 i_stride = ALIGN( h->sps->i_mb_width*16 + 2*PADH, align );
744 luma_plane_size = i_stride * (h->sps->i_mb_height*16+2*i_padv);
746 if( h->param.analyse.i_weighted_pred == X264_WEIGHTP_SMART )
747 //SMART can weight one ref and one offset -1
750 //blind only has one weighted copy (offset -1)
754 for( i = 0; i < numweightbuf; i++ )
755 CHECKED_MALLOC( h->mb.p_weight_buf[i], luma_plane_size );
759 for( i=0; i<=h->param.b_interlaced; i++ )
762 /* shouldn't really be initialized, just silences a valgrind false-positive in predict_8x8_filter_mmx */
763 CHECKED_MALLOCZERO( h->mb.intra_border_backup[i][j], (h->sps->i_mb_width*16+32)>>!!j );
764 h->mb.intra_border_backup[i][j] += 8;
770 void x264_macroblock_cache_end( x264_t *h )
773 for( i=0; i<=h->param.b_interlaced; i++ )
775 x264_free( h->mb.intra_border_backup[i][j] - 8 );
777 for( j=0; j<32; j++ )
778 x264_free( h->mb.mvr[i][j] );
779 for( i=0; i<16; i++ )
780 x264_free( h->mb.p_weight_buf[i] );
782 if( h->param.b_cabac )
784 x264_free( h->mb.chroma_pred_mode );
785 x264_free( h->mb.mvd[0] );
786 x264_free( h->mb.mvd[1] );
788 x264_free( h->mb.intra4x4_pred_mode );
789 x264_free( h->mb.non_zero_count );
790 x264_free( h->mb.mb_transform_size );
791 x264_free( h->mb.skipbp );
792 x264_free( h->mb.cbp );
793 x264_free( h->mb.qp );
795 void x264_macroblock_slice_init( x264_t *h )
799 h->mb.mv[0] = h->fdec->mv[0];
800 h->mb.mv[1] = h->fdec->mv[1];
801 h->mb.ref[0] = h->fdec->ref[0];
802 h->mb.ref[1] = h->fdec->ref[1];
803 h->mb.type = h->fdec->mb_type;
804 h->mb.partition = h->fdec->mb_partition;
806 h->fdec->i_ref[0] = h->i_ref0;
807 h->fdec->i_ref[1] = h->i_ref1;
808 for( i = 0; i < h->i_ref0; i++ )
809 h->fdec->ref_poc[0][i] = h->fref0[i]->i_poc;
810 if( h->sh.i_type == SLICE_TYPE_B )
812 for( i = 0; i < h->i_ref1; i++ )
813 h->fdec->ref_poc[1][i] = h->fref1[i]->i_poc;
815 map_col_to_list0(-1) = -1;
816 map_col_to_list0(-2) = -2;
817 for( i = 0; i < h->fref1[0]->i_ref[0]; i++ )
819 int poc = h->fref1[0]->ref_poc[0][i];
820 map_col_to_list0(i) = -2;
821 for( j = 0; j < h->i_ref0; j++ )
822 if( h->fref0[j]->i_poc == poc )
824 map_col_to_list0(i) = j;
829 if( h->sh.i_type == SLICE_TYPE_P )
830 memset( h->mb.cache.skip, 0, X264_SCAN8_SIZE * sizeof( int8_t ) );
832 /* init with not available (for top right idx=7,15) */
833 memset( h->mb.cache.ref[0], -2, X264_SCAN8_SIZE * sizeof( int8_t ) );
834 memset( h->mb.cache.ref[1], -2, X264_SCAN8_SIZE * sizeof( int8_t ) );
836 setup_inverse_delta_pocs( h );
838 h->mb.i_neighbour4[6] =
839 h->mb.i_neighbour4[9] =
840 h->mb.i_neighbour4[12] =
841 h->mb.i_neighbour4[14] = MB_LEFT|MB_TOP|MB_TOPLEFT|MB_TOPRIGHT;
842 h->mb.i_neighbour4[3] =
843 h->mb.i_neighbour4[7] =
844 h->mb.i_neighbour4[11] =
845 h->mb.i_neighbour4[13] =
846 h->mb.i_neighbour4[15] =
847 h->mb.i_neighbour8[3] = MB_LEFT|MB_TOP|MB_TOPLEFT;
850 void x264_macroblock_thread_init( x264_t *h )
862 h->mb.pic.p_fenc[0] = h->mb.pic.fenc_buf;
863 h->mb.pic.p_fenc[1] = h->mb.pic.fenc_buf + 16*FENC_STRIDE;
864 h->mb.pic.p_fenc[2] = h->mb.pic.fenc_buf + 16*FENC_STRIDE + 8;
865 h->mb.pic.p_fdec[0] = h->mb.pic.fdec_buf + 2*FDEC_STRIDE;
866 h->mb.pic.p_fdec[1] = h->mb.pic.fdec_buf + 19*FDEC_STRIDE;
867 h->mb.pic.p_fdec[2] = h->mb.pic.fdec_buf + 19*FDEC_STRIDE + 16;
870 void x264_prefetch_fenc( x264_t *h, x264_frame_t *fenc, int i_mb_x, int i_mb_y )
872 int stride_y = fenc->i_stride[0];
873 int stride_uv = fenc->i_stride[1];
874 int off_y = 16 * (i_mb_x + i_mb_y * stride_y);
875 int off_uv = 8 * (i_mb_x + i_mb_y * stride_uv);
876 h->mc.prefetch_fenc( fenc->plane[0]+off_y, stride_y,
877 fenc->plane[1+(i_mb_x&1)]+off_uv, stride_uv, i_mb_x );
880 static NOINLINE void copy_column8( uint8_t *dst, uint8_t *src )
882 // input pointers are offset by 4 rows because that's faster (smaller instruction size on x86)
884 for( i = -4; i < 4; i++ )
885 dst[i*FDEC_STRIDE] = src[i*FDEC_STRIDE];
888 static void ALWAYS_INLINE x264_macroblock_load_pic_pointers( x264_t *h, int i_mb_x, int i_mb_y, int i)
890 const int w = (i == 0 ? 16 : 8);
891 const int i_stride = h->fdec->i_stride[!!i];
892 const int i_stride2 = i_stride << h->mb.b_interlaced;
893 const int i_pix_offset = h->mb.b_interlaced
894 ? w * (i_mb_x + (i_mb_y&~1) * i_stride) + (i_mb_y&1) * i_stride
895 : w * (i_mb_x + i_mb_y * i_stride);
896 const uint8_t *plane_fdec = &h->fdec->plane[i][i_pix_offset];
897 const uint8_t *intra_fdec = h->param.b_sliced_threads ? plane_fdec-i_stride2 :
898 &h->mb.intra_border_backup[i_mb_y & h->sh.b_mbaff][i][i_mb_x*16>>!!i];
899 int ref_pix_offset[2] = { i_pix_offset, i_pix_offset };
900 x264_frame_t **fref[2] = { h->fref0, h->fref1 };
902 if( h->mb.b_interlaced )
903 ref_pix_offset[1] += (1-2*(i_mb_y&1)) * i_stride;
904 h->mb.pic.i_stride[i] = i_stride2;
905 h->mb.pic.p_fenc_plane[i] = &h->fenc->plane[i][i_pix_offset];
906 h->mc.copy[i?PIXEL_8x8:PIXEL_16x16]( h->mb.pic.p_fenc[i], FENC_STRIDE,
907 h->mb.pic.p_fenc_plane[i], i_stride2, w );
909 memcpy( &h->mb.pic.p_fdec[i][-1-FDEC_STRIDE], intra_fdec-1, w*3/2+1 );
911 memset( &h->mb.pic.p_fdec[i][-1-FDEC_STRIDE], 0, w*3/2+1 );
912 if( h->mb.b_interlaced || h->mb.b_reencode_mb )
913 for( j = 0; j < w; j++ )
914 h->mb.pic.p_fdec[i][-1+j*FDEC_STRIDE] = plane_fdec[-1+j*i_stride2];
915 for( j = 0; j < h->mb.pic.i_fref[0]; j++ )
917 h->mb.pic.p_fref[0][j][i==0 ? 0:i+3] = &fref[0][j >> h->mb.b_interlaced]->plane[i][ref_pix_offset[j&1]];
920 for( k = 1; k < 4; k++ )
921 h->mb.pic.p_fref[0][j][k] = &fref[0][j >> h->mb.b_interlaced]->filtered[k][ref_pix_offset[j&1]];
922 if( h->sh.weight[j][0].weightfn )
923 h->mb.pic.p_fref_w[j] = &h->fenc->weighted[j >> h->mb.b_interlaced][ref_pix_offset[j&1]];
925 h->mb.pic.p_fref_w[j] = h->mb.pic.p_fref[0][j][0];
928 if( h->sh.i_type == SLICE_TYPE_B )
929 for( j = 0; j < h->mb.pic.i_fref[1]; j++ )
931 h->mb.pic.p_fref[1][j][i==0 ? 0:i+3] = &fref[1][j >> h->mb.b_interlaced]->plane[i][ref_pix_offset[j&1]];
933 for( k = 1; k < 4; k++ )
934 h->mb.pic.p_fref[1][j][k] = &fref[1][j >> h->mb.b_interlaced]->filtered[k][ref_pix_offset[j&1]];
938 void x264_macroblock_cache_load( x264_t *h, int i_mb_x, int i_mb_y )
940 int i_mb_xy = i_mb_y * h->mb.i_mb_stride + i_mb_x;
941 int i_mb_4x4 = 4*(i_mb_y * h->mb.i_b4_stride + i_mb_x);
942 int i_mb_8x8 = 2*(i_mb_y * h->mb.i_b8_stride + i_mb_x);
943 int i_top_y = i_mb_y - (1 << h->mb.b_interlaced);
944 int i_top_xy = i_top_y * h->mb.i_mb_stride + i_mb_x;
945 int i_top_4x4 = (4*i_top_y+3) * h->mb.i_b4_stride + 4*i_mb_x;
946 int i_top_8x8 = (2*i_top_y+1) * h->mb.i_b8_stride + 2*i_mb_x;
948 int i_top_type = -1; /* gcc warn */
954 h->mb.i_mb_x = i_mb_x;
955 h->mb.i_mb_y = i_mb_y;
956 h->mb.i_mb_xy = i_mb_xy;
957 h->mb.i_b8_xy = i_mb_8x8;
958 h->mb.i_b4_xy = i_mb_4x4;
959 h->mb.i_mb_top_xy = i_top_xy;
960 h->mb.i_neighbour = 0;
961 h->mb.i_neighbour_intra = 0;
964 if( i_top_xy >= h->sh.i_first_mb )
966 h->mb.i_mb_type_top =
967 i_top_type = h->mb.type[i_top_xy];
968 h->mb.cache.i_cbp_top = h->mb.cbp[i_top_xy];
970 h->mb.i_neighbour |= MB_TOP;
972 if( !h->param.b_constrained_intra || IS_INTRA( i_top_type ) )
973 h->mb.i_neighbour_intra |= MB_TOP;
976 CP32( &h->mb.cache.intra4x4_pred_mode[x264_scan8[0] - 8], &h->mb.intra4x4_pred_mode[i_top_xy][0] );
978 /* load non_zero_count */
979 CP32( &h->mb.cache.non_zero_count[x264_scan8[0] - 8], &h->mb.non_zero_count[i_top_xy][12] );
980 /* shift because x264_scan8[16] is misaligned */
981 M32( &h->mb.cache.non_zero_count[x264_scan8[16+0] - 9] ) = M16( &h->mb.non_zero_count[i_top_xy][18] ) << 8;
982 M32( &h->mb.cache.non_zero_count[x264_scan8[16+4] - 9] ) = M16( &h->mb.non_zero_count[i_top_xy][22] ) << 8;
986 h->mb.i_mb_type_top = -1;
987 h->mb.cache.i_cbp_top = -1;
990 M32( &h->mb.cache.intra4x4_pred_mode[x264_scan8[0] - 8] ) = 0xFFFFFFFFU;
992 /* load non_zero_count */
993 M32( &h->mb.cache.non_zero_count[x264_scan8[ 0] - 8] ) = 0x80808080U;
994 M32( &h->mb.cache.non_zero_count[x264_scan8[16+0] - 9] ) = 0x80808080U;
995 M32( &h->mb.cache.non_zero_count[x264_scan8[16+4] - 9] ) = 0x80808080U;
998 if( i_mb_x > 0 && i_mb_xy > h->sh.i_first_mb )
1000 i_left_xy = i_mb_xy - 1;
1001 h->mb.i_mb_type_left =
1002 i_left_type = h->mb.type[i_left_xy];
1003 h->mb.cache.i_cbp_left = h->mb.cbp[h->mb.i_mb_xy - 1];
1005 h->mb.i_neighbour |= MB_LEFT;
1007 if( !h->param.b_constrained_intra || IS_INTRA( i_left_type ) )
1008 h->mb.i_neighbour_intra |= MB_LEFT;
1011 h->mb.cache.intra4x4_pred_mode[x264_scan8[0 ] - 1] = h->mb.intra4x4_pred_mode[i_left_xy][4];
1012 h->mb.cache.intra4x4_pred_mode[x264_scan8[2 ] - 1] = h->mb.intra4x4_pred_mode[i_left_xy][5];
1013 h->mb.cache.intra4x4_pred_mode[x264_scan8[8 ] - 1] = h->mb.intra4x4_pred_mode[i_left_xy][6];
1014 h->mb.cache.intra4x4_pred_mode[x264_scan8[10] - 1] = h->mb.intra4x4_pred_mode[i_left_xy][3];
1016 /* load non_zero_count */
1017 h->mb.cache.non_zero_count[x264_scan8[0 ] - 1] = h->mb.non_zero_count[i_left_xy][3];
1018 h->mb.cache.non_zero_count[x264_scan8[2 ] - 1] = h->mb.non_zero_count[i_left_xy][7];
1019 h->mb.cache.non_zero_count[x264_scan8[8 ] - 1] = h->mb.non_zero_count[i_left_xy][11];
1020 h->mb.cache.non_zero_count[x264_scan8[10] - 1] = h->mb.non_zero_count[i_left_xy][15];
1022 h->mb.cache.non_zero_count[x264_scan8[16+0] - 1] = h->mb.non_zero_count[i_left_xy][16+1];
1023 h->mb.cache.non_zero_count[x264_scan8[16+2] - 1] = h->mb.non_zero_count[i_left_xy][16+3];
1025 h->mb.cache.non_zero_count[x264_scan8[16+4+0] - 1] = h->mb.non_zero_count[i_left_xy][16+4+1];
1026 h->mb.cache.non_zero_count[x264_scan8[16+4+2] - 1] = h->mb.non_zero_count[i_left_xy][16+4+3];
1030 h->mb.i_mb_type_left = -1;
1031 h->mb.cache.i_cbp_left = -1;
1033 h->mb.cache.intra4x4_pred_mode[x264_scan8[0 ] - 1] =
1034 h->mb.cache.intra4x4_pred_mode[x264_scan8[2 ] - 1] =
1035 h->mb.cache.intra4x4_pred_mode[x264_scan8[8 ] - 1] =
1036 h->mb.cache.intra4x4_pred_mode[x264_scan8[10] - 1] = -1;
1038 /* load non_zero_count */
1039 h->mb.cache.non_zero_count[x264_scan8[0 ] - 1] =
1040 h->mb.cache.non_zero_count[x264_scan8[2 ] - 1] =
1041 h->mb.cache.non_zero_count[x264_scan8[8 ] - 1] =
1042 h->mb.cache.non_zero_count[x264_scan8[10] - 1] =
1043 h->mb.cache.non_zero_count[x264_scan8[16+0] - 1] =
1044 h->mb.cache.non_zero_count[x264_scan8[16+2] - 1] =
1045 h->mb.cache.non_zero_count[x264_scan8[16+4+0] - 1] =
1046 h->mb.cache.non_zero_count[x264_scan8[16+4+2] - 1] = 0x80;
1049 if( i_mb_x < h->sps->i_mb_width - 1 && i_top_xy + 1 >= h->sh.i_first_mb )
1051 h->mb.i_neighbour |= MB_TOPRIGHT;
1052 h->mb.i_mb_type_topright = h->mb.type[ i_top_xy + 1 ];
1053 if( !h->param.b_constrained_intra || IS_INTRA( h->mb.i_mb_type_topright ) )
1054 h->mb.i_neighbour_intra |= MB_TOPRIGHT;
1057 h->mb.i_mb_type_topright = -1;
1058 if( i_mb_x > 0 && i_top_xy - 1 >= h->sh.i_first_mb )
1060 h->mb.i_neighbour |= MB_TOPLEFT;
1061 h->mb.i_mb_type_topleft = h->mb.type[ i_top_xy - 1 ];
1062 if( !h->param.b_constrained_intra || IS_INTRA( h->mb.i_mb_type_topleft ) )
1063 h->mb.i_neighbour_intra |= MB_TOPLEFT;
1066 h->mb.i_mb_type_topleft = -1;
1068 if( h->pps->b_transform_8x8_mode )
1070 h->mb.cache.i_neighbour_transform_size =
1071 ( i_left_type >= 0 && h->mb.mb_transform_size[i_left_xy] )
1072 + ( i_top_type >= 0 && h->mb.mb_transform_size[i_top_xy] );
1077 h->mb.pic.i_fref[0] = h->i_ref0 << h->mb.b_interlaced;
1078 h->mb.pic.i_fref[1] = h->i_ref1 << h->mb.b_interlaced;
1079 h->mb.cache.i_neighbour_interlaced =
1080 !!(h->mb.i_neighbour & MB_LEFT)
1081 + !!(h->mb.i_neighbour & MB_TOP);
1084 if( !h->mb.b_interlaced && !h->mb.b_reencode_mb )
1086 copy_column8( h->mb.pic.p_fdec[0]-1+ 4*FDEC_STRIDE, h->mb.pic.p_fdec[0]+15+ 4*FDEC_STRIDE );
1087 copy_column8( h->mb.pic.p_fdec[0]-1+12*FDEC_STRIDE, h->mb.pic.p_fdec[0]+15+12*FDEC_STRIDE );
1088 copy_column8( h->mb.pic.p_fdec[1]-1+ 4*FDEC_STRIDE, h->mb.pic.p_fdec[1]+ 7+ 4*FDEC_STRIDE );
1089 copy_column8( h->mb.pic.p_fdec[2]-1+ 4*FDEC_STRIDE, h->mb.pic.p_fdec[2]+ 7+ 4*FDEC_STRIDE );
1092 /* load picture pointers */
1093 x264_macroblock_load_pic_pointers( h, i_mb_x, i_mb_y, 0 );
1094 x264_macroblock_load_pic_pointers( h, i_mb_x, i_mb_y, 1 );
1095 x264_macroblock_load_pic_pointers( h, i_mb_x, i_mb_y, 2 );
1097 if( h->fdec->integral )
1099 assert( !h->mb.b_interlaced );
1100 for( i = 0; i < h->mb.pic.i_fref[0]; i++ )
1101 h->mb.pic.p_integral[0][i] = &h->fref0[i]->integral[ 16 * ( i_mb_x + i_mb_y * h->fdec->i_stride[0] )];
1102 for( i = 0; i < h->mb.pic.i_fref[1]; i++ )
1103 h->mb.pic.p_integral[1][i] = &h->fref1[i]->integral[ 16 * ( i_mb_x + i_mb_y * h->fdec->i_stride[0] )];
1106 x264_prefetch_fenc( h, h->fenc, i_mb_x, i_mb_y );
1108 /* load ref/mv/mvd */
1109 if( h->sh.i_type != SLICE_TYPE_I )
1111 const int s8x8 = h->mb.i_b8_stride;
1112 const int s4x4 = h->mb.i_b4_stride;
1116 for( i_list = 0; i_list < (h->sh.i_type == SLICE_TYPE_B ? 2 : 1 ); i_list++ )
1119 h->mb.cache.ref[i_list][x264_scan8[5 ]+1] =
1120 h->mb.cache.ref[i_list][x264_scan8[7 ]+1] =
1121 h->mb.cache.ref[i_list][x264_scan8[13]+1] = -2;
1124 if( h->mb.i_neighbour & MB_TOPLEFT )
1126 const int i8 = x264_scan8[0] - 1 - 1*8;
1127 const int ir = i_top_8x8 - 1;
1128 const int iv = i_top_4x4 - 1;
1129 h->mb.cache.ref[i_list][i8] = h->mb.ref[i_list][ir];
1130 CP32( h->mb.cache.mv[i_list][i8], h->mb.mv[i_list][iv] );
1134 const int i8 = x264_scan8[0] - 1 - 1*8;
1135 h->mb.cache.ref[i_list][i8] = -2;
1136 M32( h->mb.cache.mv[i_list][i8] ) = 0;
1139 if( h->mb.i_neighbour & MB_TOP )
1141 const int i8 = x264_scan8[0] - 8;
1142 const int ir = i_top_8x8;
1143 const int iv = i_top_4x4;
1144 h->mb.cache.ref[i_list][i8+0] =
1145 h->mb.cache.ref[i_list][i8+1] = h->mb.ref[i_list][ir + 0];
1146 h->mb.cache.ref[i_list][i8+2] =
1147 h->mb.cache.ref[i_list][i8+3] = h->mb.ref[i_list][ir + 1];
1148 CP64( h->mb.cache.mv[i_list][i8+0], h->mb.mv[i_list][iv+0] );
1149 CP64( h->mb.cache.mv[i_list][i8+2], h->mb.mv[i_list][iv+2] );
1153 const int i8 = x264_scan8[0] - 8;
1154 M64( h->mb.cache.mv[i_list][i8+0] ) = 0;
1155 M64( h->mb.cache.mv[i_list][i8+2] ) = 0;
1156 M32( &h->mb.cache.ref[i_list][i8] ) = (uint8_t)(-2) * 0x01010101U;
1159 if( h->mb.i_neighbour & MB_TOPRIGHT )
1161 const int i8 = x264_scan8[0] + 4 - 1*8;
1162 const int ir = i_top_8x8 + 2;
1163 const int iv = i_top_4x4 + 4;
1164 h->mb.cache.ref[i_list][i8] = h->mb.ref[i_list][ir];
1165 CP32( h->mb.cache.mv[i_list][i8], h->mb.mv[i_list][iv] );
1169 const int i8 = x264_scan8[0] + 4 - 1*8;
1170 h->mb.cache.ref[i_list][i8] = -2;
1171 M32( h->mb.cache.mv[i_list][i8] ) = 0;
1174 if( h->mb.i_neighbour & MB_LEFT )
1176 const int i8 = x264_scan8[0] - 1;
1177 const int ir = i_mb_8x8 - 1;
1178 const int iv = i_mb_4x4 - 1;
1179 h->mb.cache.ref[i_list][i8+0*8] =
1180 h->mb.cache.ref[i_list][i8+1*8] = h->mb.ref[i_list][ir + 0*s8x8];
1181 h->mb.cache.ref[i_list][i8+2*8] =
1182 h->mb.cache.ref[i_list][i8+3*8] = h->mb.ref[i_list][ir + 1*s8x8];
1184 CP32( h->mb.cache.mv[i_list][i8+0*8], h->mb.mv[i_list][iv + 0*s4x4] );
1185 CP32( h->mb.cache.mv[i_list][i8+1*8], h->mb.mv[i_list][iv + 1*s4x4] );
1186 CP32( h->mb.cache.mv[i_list][i8+2*8], h->mb.mv[i_list][iv + 2*s4x4] );
1187 CP32( h->mb.cache.mv[i_list][i8+3*8], h->mb.mv[i_list][iv + 3*s4x4] );
1191 const int i8 = x264_scan8[0] - 1;
1192 for( i = 0; i < 4; i++ )
1194 h->mb.cache.ref[i_list][i8+i*8] = -2;
1195 M32( h->mb.cache.mv[i_list][i8+i*8] ) = 0;
1199 if( h->param.b_cabac )
1201 if( i_top_type >= 0 )
1203 const int i8 = x264_scan8[0] - 8;
1204 const int iv = i_top_4x4;
1205 CP64( h->mb.cache.mvd[i_list][i8+0], h->mb.mvd[i_list][iv+0] );
1206 CP64( h->mb.cache.mvd[i_list][i8+2], h->mb.mvd[i_list][iv+2] );
1210 const int i8 = x264_scan8[0] - 8;
1211 M64( h->mb.cache.mvd[i_list][i8+0] ) = 0;
1212 M64( h->mb.cache.mvd[i_list][i8+2] ) = 0;
1215 if( i_left_type >= 0 )
1217 const int i8 = x264_scan8[0] - 1;
1218 const int iv = i_mb_4x4 - 1;
1219 CP32( h->mb.cache.mvd[i_list][i8+0*8], h->mb.mvd[i_list][iv + 0*s4x4] );
1220 CP32( h->mb.cache.mvd[i_list][i8+1*8], h->mb.mvd[i_list][iv + 1*s4x4] );
1221 CP32( h->mb.cache.mvd[i_list][i8+2*8], h->mb.mvd[i_list][iv + 2*s4x4] );
1222 CP32( h->mb.cache.mvd[i_list][i8+3*8], h->mb.mvd[i_list][iv + 3*s4x4] );
1226 const int i8 = x264_scan8[0] - 1;
1227 for( i = 0; i < 4; i++ )
1228 M32( h->mb.cache.mvd[i_list][i8+i*8] ) = 0;
1234 if( h->sh.i_type == SLICE_TYPE_B )
1236 h->mb.bipred_weight = h->mb.bipred_weight_buf[h->mb.b_interlaced&(i_mb_y&1)];
1237 h->mb.dist_scale_factor = h->mb.dist_scale_factor_buf[h->mb.b_interlaced&(i_mb_y&1)];
1238 if( h->param.b_cabac )
1241 x264_macroblock_cache_skip( h, 0, 0, 4, 4, 0 );
1242 skipbp = i_left_type >= 0 ? h->mb.skipbp[i_left_xy] : 0;
1243 h->mb.cache.skip[x264_scan8[0] - 1] = skipbp & 0x2;
1244 h->mb.cache.skip[x264_scan8[8] - 1] = skipbp & 0x8;
1245 skipbp = i_top_type >= 0 ? h->mb.skipbp[i_top_xy] : 0;
1246 h->mb.cache.skip[x264_scan8[0] - 8] = skipbp & 0x4;
1247 h->mb.cache.skip[x264_scan8[4] - 8] = skipbp & 0x8;
1251 if( h->sh.i_type == SLICE_TYPE_P )
1252 x264_mb_predict_mv_pskip( h, h->mb.cache.pskip_mv );
1255 h->mb.i_neighbour4[0] =
1256 h->mb.i_neighbour8[0] = (h->mb.i_neighbour_intra & (MB_TOP|MB_LEFT|MB_TOPLEFT))
1257 | ((h->mb.i_neighbour_intra & MB_TOP) ? MB_TOPRIGHT : 0);
1258 h->mb.i_neighbour4[4] =
1259 h->mb.i_neighbour4[1] = MB_LEFT | ((h->mb.i_neighbour_intra & MB_TOP) ? (MB_TOP|MB_TOPLEFT|MB_TOPRIGHT) : 0);
1260 h->mb.i_neighbour4[2] =
1261 h->mb.i_neighbour4[8] =
1262 h->mb.i_neighbour4[10] =
1263 h->mb.i_neighbour8[2] = MB_TOP|MB_TOPRIGHT | ((h->mb.i_neighbour_intra & MB_LEFT) ? (MB_LEFT|MB_TOPLEFT) : 0);
1264 h->mb.i_neighbour4[5] =
1265 h->mb.i_neighbour8[1] = MB_LEFT | (h->mb.i_neighbour_intra & MB_TOPRIGHT)
1266 | ((h->mb.i_neighbour_intra & MB_TOP) ? MB_TOP|MB_TOPLEFT : 0);
1269 static void ALWAYS_INLINE x264_macroblock_store_pic( x264_t *h, int i )
1272 int i_stride = h->fdec->i_stride[!!i];
1273 int i_stride2 = i_stride << h->mb.b_interlaced;
1274 int i_pix_offset = h->mb.b_interlaced
1275 ? w * (h->mb.i_mb_x + (h->mb.i_mb_y&~1) * i_stride) + (h->mb.i_mb_y&1) * i_stride
1276 : w * (h->mb.i_mb_x + h->mb.i_mb_y * i_stride);
1277 h->mc.copy[i?PIXEL_8x8:PIXEL_16x16](
1278 &h->fdec->plane[i][i_pix_offset], i_stride2,
1279 h->mb.pic.p_fdec[i], FDEC_STRIDE, w );
1282 void x264_macroblock_cache_save( x264_t *h )
1284 const int i_mb_xy = h->mb.i_mb_xy;
1285 const int i_mb_type = x264_mb_type_fix[h->mb.i_type];
1286 const int s8x8 = h->mb.i_b8_stride;
1287 const int s4x4 = h->mb.i_b4_stride;
1288 const int i_mb_4x4 = h->mb.i_b4_xy;
1289 const int i_mb_8x8 = h->mb.i_b8_xy;
1291 /* GCC pessimizes direct stores to heap-allocated 8-bit arrays due to aliasing.*/
1292 /* By only dereferencing them once, we avoid this issue. */
1293 int8_t *intra4x4_pred_mode = h->mb.intra4x4_pred_mode[i_mb_xy];
1294 uint8_t *non_zero_count = h->mb.non_zero_count[i_mb_xy];
1298 x264_macroblock_store_pic( h, 0 );
1299 x264_macroblock_store_pic( h, 1 );
1300 x264_macroblock_store_pic( h, 2 );
1302 x264_prefetch_fenc( h, h->fdec, h->mb.i_mb_x, h->mb.i_mb_y );
1304 h->mb.type[i_mb_xy] = i_mb_type;
1305 h->mb.partition[i_mb_xy] = IS_INTRA( i_mb_type ) ? D_16x16 : h->mb.i_partition;
1306 h->mb.i_mb_prev_xy = i_mb_xy;
1309 if( i_mb_type == I_4x4 )
1311 CP32( &intra4x4_pred_mode[0], &h->mb.cache.intra4x4_pred_mode[x264_scan8[10]] );
1312 M32( &intra4x4_pred_mode[4] ) = pack8to32(h->mb.cache.intra4x4_pred_mode[x264_scan8[5] ],
1313 h->mb.cache.intra4x4_pred_mode[x264_scan8[7] ],
1314 h->mb.cache.intra4x4_pred_mode[x264_scan8[13] ], 0);
1316 else if( !h->param.b_constrained_intra || IS_INTRA(i_mb_type) )
1317 M64( intra4x4_pred_mode ) = I_PRED_4x4_DC * 0x0101010101010101ULL;
1319 M64( intra4x4_pred_mode ) = (uint8_t)(-1) * 0x0101010101010101ULL;
1322 if( i_mb_type == I_PCM )
1324 h->mb.qp[i_mb_xy] = 0;
1325 h->mb.i_last_dqp = 0;
1326 h->mb.i_cbp_chroma = 2;
1327 h->mb.i_cbp_luma = 0xf;
1328 h->mb.cbp[i_mb_xy] = 0x72f; /* all set */
1329 h->mb.b_transform_8x8 = 0;
1330 memset( non_zero_count, 16, 24 );
1334 /* save non zero count */
1335 CP32( &non_zero_count[0*4], &h->mb.cache.non_zero_count[x264_scan8[0]+0*8] );
1336 CP32( &non_zero_count[1*4], &h->mb.cache.non_zero_count[x264_scan8[0]+1*8] );
1337 CP32( &non_zero_count[2*4], &h->mb.cache.non_zero_count[x264_scan8[0]+2*8] );
1338 CP32( &non_zero_count[3*4], &h->mb.cache.non_zero_count[x264_scan8[0]+3*8] );
1339 M16( &non_zero_count[16+0*2] ) = M32( &h->mb.cache.non_zero_count[x264_scan8[16+0*2]-1] ) >> 8;
1340 M16( &non_zero_count[16+1*2] ) = M32( &h->mb.cache.non_zero_count[x264_scan8[16+1*2]-1] ) >> 8;
1341 M16( &non_zero_count[16+2*2] ) = M32( &h->mb.cache.non_zero_count[x264_scan8[16+2*2]-1] ) >> 8;
1342 M16( &non_zero_count[16+3*2] ) = M32( &h->mb.cache.non_zero_count[x264_scan8[16+3*2]-1] ) >> 8;
1344 if( h->mb.i_type != I_16x16 && h->mb.i_cbp_luma == 0 && h->mb.i_cbp_chroma == 0 )
1345 h->mb.i_qp = h->mb.i_last_qp;
1346 h->mb.qp[i_mb_xy] = h->mb.i_qp;
1347 h->mb.i_last_dqp = h->mb.i_qp - h->mb.i_last_qp;
1348 h->mb.i_last_qp = h->mb.i_qp;
1351 if( h->mb.i_cbp_luma == 0 && h->mb.i_type != I_8x8 )
1352 h->mb.b_transform_8x8 = 0;
1353 h->mb.mb_transform_size[i_mb_xy] = h->mb.b_transform_8x8;
1355 if( h->sh.i_type != SLICE_TYPE_I )
1357 if( !IS_INTRA( i_mb_type ) )
1359 h->mb.ref[0][i_mb_8x8+0+0*s8x8] = h->mb.cache.ref[0][x264_scan8[0]];
1360 h->mb.ref[0][i_mb_8x8+1+0*s8x8] = h->mb.cache.ref[0][x264_scan8[4]];
1361 h->mb.ref[0][i_mb_8x8+0+1*s8x8] = h->mb.cache.ref[0][x264_scan8[8]];
1362 h->mb.ref[0][i_mb_8x8+1+1*s8x8] = h->mb.cache.ref[0][x264_scan8[12]];
1363 for( y = 0; y < 4; y++ )
1365 CP64( h->mb.mv[0][i_mb_4x4+y*s4x4+0], h->mb.cache.mv[0][x264_scan8[0]+8*y+0] );
1366 CP64( h->mb.mv[0][i_mb_4x4+y*s4x4+2], h->mb.cache.mv[0][x264_scan8[0]+8*y+2] );
1368 if( h->sh.i_type == SLICE_TYPE_B )
1370 h->mb.ref[1][i_mb_8x8+0+0*s8x8] = h->mb.cache.ref[1][x264_scan8[0]];
1371 h->mb.ref[1][i_mb_8x8+1+0*s8x8] = h->mb.cache.ref[1][x264_scan8[4]];
1372 h->mb.ref[1][i_mb_8x8+0+1*s8x8] = h->mb.cache.ref[1][x264_scan8[8]];
1373 h->mb.ref[1][i_mb_8x8+1+1*s8x8] = h->mb.cache.ref[1][x264_scan8[12]];
1374 for( y = 0; y < 4; y++ )
1376 CP64( h->mb.mv[1][i_mb_4x4+y*s4x4+0], h->mb.cache.mv[1][x264_scan8[0]+8*y+0] );
1377 CP64( h->mb.mv[1][i_mb_4x4+y*s4x4+2], h->mb.cache.mv[1][x264_scan8[0]+8*y+2] );
1384 for( i_list = 0; i_list < (h->sh.i_type == SLICE_TYPE_B ? 2 : 1 ); i_list++ )
1386 M16( &h->mb.ref[i_list][i_mb_8x8+0*s8x8] ) = (uint8_t)(-1) * 0x0101;
1387 M16( &h->mb.ref[i_list][i_mb_8x8+1*s8x8] ) = (uint8_t)(-1) * 0x0101;
1388 for( y = 0; y < 4; y++ )
1390 M64( h->mb.mv[i_list][i_mb_4x4+y*s4x4+0] ) = 0;
1391 M64( h->mb.mv[i_list][i_mb_4x4+y*s4x4+2] ) = 0;
1397 if( h->param.b_cabac )
1399 if( IS_INTRA(i_mb_type) && i_mb_type != I_PCM )
1400 h->mb.chroma_pred_mode[i_mb_xy] = x264_mb_pred_mode8x8c_fix[ h->mb.i_chroma_pred_mode ];
1402 h->mb.chroma_pred_mode[i_mb_xy] = I_PRED_CHROMA_DC;
1404 if( !IS_INTRA( i_mb_type ) && !IS_SKIP( i_mb_type ) && !IS_DIRECT( i_mb_type ) )
1406 for( y = 0; y < 4; y++ )
1408 CP64( h->mb.mvd[0][i_mb_4x4+y*s4x4+0], h->mb.cache.mvd[0][x264_scan8[0]+8*y+0] );
1409 CP64( h->mb.mvd[0][i_mb_4x4+y*s4x4+2], h->mb.cache.mvd[0][x264_scan8[0]+8*y+2] );
1411 if( h->sh.i_type == SLICE_TYPE_B )
1412 for( y = 0; y < 4; y++ )
1414 CP64( h->mb.mvd[1][i_mb_4x4+y*s4x4+0], h->mb.cache.mvd[1][x264_scan8[0]+8*y+0] );
1415 CP64( h->mb.mvd[1][i_mb_4x4+y*s4x4+2], h->mb.cache.mvd[1][x264_scan8[0]+8*y+2] );
1420 for( y = 0; y < 4; y++ )
1422 M64( h->mb.mvd[0][i_mb_4x4+y*s4x4+0] ) = 0;
1423 M64( h->mb.mvd[0][i_mb_4x4+y*s4x4+2] ) = 0;
1425 if( h->sh.i_type == SLICE_TYPE_B )
1426 for( y = 0; y < 4; y++ )
1428 M64( h->mb.mvd[1][i_mb_4x4+y*s4x4+0] ) = 0;
1429 M64( h->mb.mvd[1][i_mb_4x4+y*s4x4+2] ) = 0;
1433 if( h->sh.i_type == SLICE_TYPE_B )
1435 if( i_mb_type == B_SKIP || i_mb_type == B_DIRECT )
1436 h->mb.skipbp[i_mb_xy] = 0xf;
1437 else if( i_mb_type == B_8x8 )
1439 int skipbp = ( h->mb.i_sub_partition[0] == D_DIRECT_8x8 ) << 0;
1440 skipbp |= ( h->mb.i_sub_partition[1] == D_DIRECT_8x8 ) << 1;
1441 skipbp |= ( h->mb.i_sub_partition[2] == D_DIRECT_8x8 ) << 2;
1442 skipbp |= ( h->mb.i_sub_partition[3] == D_DIRECT_8x8 ) << 3;
1443 h->mb.skipbp[i_mb_xy] = skipbp;
1446 h->mb.skipbp[i_mb_xy] = 0;
1452 void x264_macroblock_bipred_init( x264_t *h )
1454 int i_ref0, i_ref1, field;
1455 for( field = 0; field <= h->sh.b_mbaff; field++ )
1456 for( i_ref0 = 0; i_ref0 < (h->i_ref0<<h->sh.b_mbaff); i_ref0++ )
1458 int poc0 = h->fref0[i_ref0>>h->sh.b_mbaff]->i_poc;
1459 if( h->sh.b_mbaff && field^(i_ref0&1) )
1460 poc0 += h->sh.i_delta_poc_bottom;
1461 for( i_ref1 = 0; i_ref1 < (h->i_ref1<<h->sh.b_mbaff); i_ref1++ )
1463 int dist_scale_factor;
1464 int poc1 = h->fref1[i_ref1>>h->sh.b_mbaff]->i_poc;
1465 if( h->sh.b_mbaff && field^(i_ref1&1) )
1466 poc1 += h->sh.i_delta_poc_bottom;
1467 int cur_poc = h->fdec->i_poc + field*h->sh.i_delta_poc_bottom;
1468 int td = x264_clip3( poc1 - poc0, -128, 127 );
1469 if( td == 0 /* || pic0 is a long-term ref */ )
1470 dist_scale_factor = 256;
1473 int tb = x264_clip3( cur_poc - poc0, -128, 127 );
1474 int tx = (16384 + (abs(td) >> 1)) / td;
1475 dist_scale_factor = x264_clip3( (tb * tx + 32) >> 6, -1024, 1023 );
1478 h->mb.dist_scale_factor_buf[field][i_ref0][i_ref1] = dist_scale_factor;
1480 dist_scale_factor >>= 2;
1481 if( h->param.analyse.b_weighted_bipred
1482 && dist_scale_factor >= -64
1483 && dist_scale_factor <= 128 )
1485 h->mb.bipred_weight_buf[field][i_ref0][i_ref1] = 64 - dist_scale_factor;
1486 // ssse3 implementation of biweight doesn't support the extrema.
1487 // if we ever generate them, we'll have to drop that optimization.
1488 assert( dist_scale_factor >= -63 && dist_scale_factor <= 127 );
1491 h->mb.bipred_weight_buf[field][i_ref0][i_ref1] = 32;