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];
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 if( i_refa == i_ref ) i_count++;
87 if( i_refb == i_ref ) i_count++;
88 if( i_refc == i_ref ) i_count++;
93 x264_median_mv( mvp, mv_a, mv_b, mv_c );
95 else if( i_count == 1 )
99 else if( i_refb == i_ref )
104 else if( i_refb == -2 && i_refc == -2 && i_refa != -2 )
110 void x264_mb_predict_mv_16x16( x264_t *h, int i_list, int i_ref, int16_t mvp[2] )
112 int i_refa = h->mb.cache.ref[i_list][X264_SCAN8_0 - 1];
113 int16_t *mv_a = h->mb.cache.mv[i_list][X264_SCAN8_0 - 1];
114 int i_refb = h->mb.cache.ref[i_list][X264_SCAN8_0 - 8];
115 int16_t *mv_b = h->mb.cache.mv[i_list][X264_SCAN8_0 - 8];
116 int i_refc = h->mb.cache.ref[i_list][X264_SCAN8_0 - 8 + 4];
117 int16_t *mv_c = h->mb.cache.mv[i_list][X264_SCAN8_0 - 8 + 4];
123 i_refc = h->mb.cache.ref[i_list][X264_SCAN8_0 - 8 - 1];
124 mv_c = h->mb.cache.mv[i_list][X264_SCAN8_0 - 8 - 1];
127 if( i_refa == i_ref ) i_count++;
128 if( i_refb == i_ref ) i_count++;
129 if( i_refc == i_ref ) i_count++;
134 x264_median_mv( mvp, mv_a, mv_b, mv_c );
136 else if( i_count == 1 )
138 if( i_refa == i_ref )
140 else if( i_refb == i_ref )
145 else if( i_refb == -2 && i_refc == -2 && i_refa != -2 )
152 void x264_mb_predict_mv_pskip( x264_t *h, int16_t mv[2] )
154 int i_refa = h->mb.cache.ref[0][X264_SCAN8_0 - 1];
155 int i_refb = h->mb.cache.ref[0][X264_SCAN8_0 - 8];
156 int16_t *mv_a = h->mb.cache.mv[0][X264_SCAN8_0 - 1];
157 int16_t *mv_b = h->mb.cache.mv[0][X264_SCAN8_0 - 8];
159 if( i_refa == -2 || i_refb == -2 ||
160 !( i_refa | M32( mv_a ) ) ||
161 !( i_refb | M32( mv_b ) ) )
167 x264_mb_predict_mv_16x16( h, 0, 0, mv );
171 static int x264_mb_predict_mv_direct16x16_temporal( x264_t *h )
173 int i_mb_4x4 = 16 * h->mb.i_mb_stride * h->mb.i_mb_y + 4 * h->mb.i_mb_x;
174 int i_mb_8x8 = 4 * h->mb.i_mb_stride * h->mb.i_mb_y + 2 * h->mb.i_mb_x;
176 const int type_col = h->fref1[0]->mb_type[h->mb.i_mb_xy];
178 x264_macroblock_cache_ref( h, 0, 0, 4, 4, 1, 0 );
180 if( IS_INTRA( type_col ) )
182 x264_macroblock_cache_ref( h, 0, 0, 4, 4, 0, 0 );
183 x264_macroblock_cache_mv( h, 0, 0, 4, 4, 0, 0 );
184 x264_macroblock_cache_mv( h, 0, 0, 4, 4, 1, 0 );
188 for( i8 = 0; i8 < 4; i8++ )
192 const int i_part_8x8 = i_mb_8x8 + x8 + y8 * h->mb.i_b8_stride;
193 const int i_ref1_ref = h->fref1[0]->ref[0][i_part_8x8];
194 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);
198 const int dist_scale_factor = h->mb.dist_scale_factor[i_ref][0];
199 const int16_t *mv_col = h->fref1[0]->mv[0][ i_mb_4x4 + 3*x8 + 3*y8 * h->mb.i_b4_stride];
200 const int l0x = ( dist_scale_factor * mv_col[0] + 128 ) >> 8;
201 const int l0y = ( dist_scale_factor * mv_col[1] + 128 ) >> 8;
202 if( h->param.i_threads > 1 && (l0y > h->mb.mv_max_spel[1] || l0y-mv_col[1] > h->mb.mv_max_spel[1]) )
204 x264_macroblock_cache_ref( h, 2*x8, 2*y8, 2, 2, 0, i_ref );
205 x264_macroblock_cache_mv( h, 2*x8, 2*y8, 2, 2, 0, pack16to32_mask(l0x, l0y) );
206 x264_macroblock_cache_mv( h, 2*x8, 2*y8, 2, 2, 1, pack16to32_mask(l0x-mv_col[0], l0y-mv_col[1]) );
210 /* the collocated ref isn't in the current list0 */
211 /* FIXME: we might still be able to use direct_8x8 on some partitions */
212 /* FIXME: with B-pyramid + extensive ref list reordering
213 * (not currently used), we would also have to check
214 * l1mv1 like in spatial mode */
222 static int x264_mb_predict_mv_direct16x16_spatial( x264_t *h )
225 ALIGNED_ARRAY_8( int16_t, mv,[2],[2] );
228 const int8_t *l1ref0 = &h->fref1[0]->ref[0][ h->mb.i_b8_xy ];
229 const int8_t *l1ref1 = &h->fref1[0]->ref[1][ h->mb.i_b8_xy ];
230 const int16_t (*l1mv0)[2] = (const int16_t (*)[2]) &h->fref1[0]->mv[0][ h->mb.i_b4_xy ];
231 const int16_t (*l1mv1)[2] = (const int16_t (*)[2]) &h->fref1[0]->mv[1][ h->mb.i_b4_xy ];
232 const int type_col = h->fref1[0]->mb_type[ h->mb.i_mb_xy ];
234 for( i_list=0; i_list<2; i_list++ )
236 int i_refa = h->mb.cache.ref[i_list][X264_SCAN8_0 - 1];
237 int i_refb = h->mb.cache.ref[i_list][X264_SCAN8_0 - 8];
238 int i_refc = h->mb.cache.ref[i_list][X264_SCAN8_0 - 8 + 4];
240 i_refc = h->mb.cache.ref[i_list][X264_SCAN8_0 - 8 - 1];
242 ref[i_list] = i_refa;
243 if( ref[i_list] < 0 || ( i_refb < ref[i_list] && i_refb >= 0 ))
244 ref[i_list] = i_refb;
245 if( ref[i_list] < 0 || ( i_refc < ref[i_list] && i_refc >= 0 ))
246 ref[i_list] = i_refc;
247 if( ref[i_list] < 0 )
251 if( ref[0] < 0 && ref[1] < 0 )
253 x264_macroblock_cache_ref( h, 0, 0, 4, 4, 0, 0 );
254 x264_macroblock_cache_ref( h, 0, 0, 4, 4, 1, 0 );
255 x264_macroblock_cache_mv( h, 0, 0, 4, 4, 0, 0 );
256 x264_macroblock_cache_mv( h, 0, 0, 4, 4, 1, 0 );
261 x264_mb_predict_mv_16x16( h, 0, ref[0], mv[0] );
266 x264_mb_predict_mv_16x16( h, 1, ref[1], mv[1] );
270 x264_macroblock_cache_ref( h, 0, 0, 4, 4, 0, ref[0] );
271 x264_macroblock_cache_ref( h, 0, 0, 4, 4, 1, ref[1] );
272 x264_macroblock_cache_mv_ptr( h, 0, 0, 4, 4, 0, mv[0] );
273 x264_macroblock_cache_mv_ptr( h, 0, 0, 4, 4, 1, mv[1] );
278 if( h->param.i_threads > 1
279 && ( mv[0][1] > h->mb.mv_max_spel[1]
280 || mv[1][1] > h->mb.mv_max_spel[1] ) )
283 fprintf(stderr, "direct_spatial: (%d,%d) (%d,%d) > %d \n",
284 mv[0][0], mv[0][1], mv[1][0], mv[1][1],
285 h->mb.mv_max_spel[1]);
290 if( IS_INTRA( type_col ) || (ref[0]&&ref[1]) )
294 for( i8=0; i8<4; i8++ )
298 const int o8 = x8 + y8 * h->mb.i_b8_stride;
299 const int o4 = 3*(x8 + y8 * h->mb.i_b4_stride);
300 if( l1ref0[o8] == 0 )
302 if( abs( l1mv0[o4][0] ) <= 1 && abs( l1mv0[o4][1] ) <= 1 )
304 if( ref[0] == 0 ) x264_macroblock_cache_mv( h, 2*x8, 2*y8, 2, 2, 0, 0 );
305 if( ref[1] == 0 ) x264_macroblock_cache_mv( h, 2*x8, 2*y8, 2, 2, 1, 0 );
308 else if( l1ref0[o8] < 0 && l1ref1[o8] == 0 )
310 if( abs( l1mv1[o4][0] ) <= 1 && abs( l1mv1[o4][1] ) <= 1 )
312 if( ref[0] == 0 ) x264_macroblock_cache_mv( h, 2*x8, 2*y8, 2, 2, 0, 0 );
313 if( ref[1] == 0 ) x264_macroblock_cache_mv( h, 2*x8, 2*y8, 2, 2, 1, 0 );
321 int x264_mb_predict_mv_direct16x16( x264_t *h, int *b_changed )
324 if( h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_NONE )
326 else if( h->sh.b_direct_spatial_mv_pred )
327 b_available = x264_mb_predict_mv_direct16x16_spatial( h );
329 b_available = x264_mb_predict_mv_direct16x16_temporal( h );
331 if( b_changed != NULL && b_available )
333 int type_col = h->fref1[0]->mb_type[h->mb.i_mb_xy];
336 if( IS_INTRA( type_col ) || type_col == P_SKIP )
338 changed |= M32( h->mb.cache.direct_mv[0][0] ) ^ M32( h->mb.cache.mv[0][X264_SCAN8_0] );
339 changed |= M32( h->mb.cache.direct_mv[1][0] ) ^ M32( h->mb.cache.mv[1][X264_SCAN8_0] );
340 changed |= h->mb.cache.direct_ref[0][0] ^ h->mb.cache.ref[0][X264_SCAN8_0];
341 changed |= h->mb.cache.direct_ref[1][0] ^ h->mb.cache.ref[1][X264_SCAN8_0];
346 for( l = 0; l < 2; l++ )
348 changed |= M32( h->mb.cache.direct_mv[l][0] ) ^ M32( h->mb.cache.mv[l][x264_scan8[ 0]] );
350 changed |= M32( h->mb.cache.direct_mv[l][1] ) ^ M32( h->mb.cache.mv[l][x264_scan8[ 4]] );
351 changed |= M32( h->mb.cache.direct_mv[l][2] ) ^ M32( h->mb.cache.mv[l][x264_scan8[ 8]] );
352 changed |= M32( h->mb.cache.direct_mv[l][3] ) ^ M32( h->mb.cache.mv[l][x264_scan8[12]] );
354 changed |= h->mb.cache.direct_ref[l][0] ^ h->mb.cache.ref[l][x264_scan8[ 0]];
355 changed |= h->mb.cache.direct_ref[l][1] ^ h->mb.cache.ref[l][x264_scan8[ 4]];
356 changed |= h->mb.cache.direct_ref[l][2] ^ h->mb.cache.ref[l][x264_scan8[ 8]];
357 changed |= h->mb.cache.direct_ref[l][3] ^ h->mb.cache.ref[l][x264_scan8[12]];
360 *b_changed = changed;
369 for( l = 0; l < 2; l++ )
371 CP32( h->mb.cache.direct_mv[l][0], h->mb.cache.mv[l][x264_scan8[ 0]] );
372 CP32( h->mb.cache.direct_mv[l][1], h->mb.cache.mv[l][x264_scan8[ 4]] );
373 CP32( h->mb.cache.direct_mv[l][2], h->mb.cache.mv[l][x264_scan8[ 8]] );
374 CP32( h->mb.cache.direct_mv[l][3], h->mb.cache.mv[l][x264_scan8[12]] );
375 h->mb.cache.direct_ref[l][0] = h->mb.cache.ref[l][x264_scan8[ 0]];
376 h->mb.cache.direct_ref[l][1] = h->mb.cache.ref[l][x264_scan8[ 4]];
377 h->mb.cache.direct_ref[l][2] = h->mb.cache.ref[l][x264_scan8[ 8]];
378 h->mb.cache.direct_ref[l][3] = h->mb.cache.ref[l][x264_scan8[12]];
385 /* This just improves encoder performance, it's not part of the spec */
386 void x264_mb_predict_mv_ref16x16( x264_t *h, int i_list, int i_ref, int16_t mvc[9][2], int *i_mvc )
388 int16_t (*mvr)[2] = h->mb.mvr[i_list][i_ref];
391 #define SET_MVP(mvp) { \
392 CP32( mvc[i], mvp ); \
397 if( h->sh.i_type == SLICE_TYPE_B
398 && h->mb.cache.ref[i_list][x264_scan8[12]] == i_ref )
400 SET_MVP( h->mb.cache.mv[i_list][x264_scan8[12]] );
403 if( i_ref == 0 && h->frames.b_have_lowres )
405 int16_t (*lowres_mv)[2] = i_list ? h->fenc->lowres_mvs[1][h->fref1[0]->i_frame-h->fenc->i_frame-1]
406 : h->fenc->lowres_mvs[0][h->fenc->i_frame-h->fref0[0]->i_frame-1];
407 if( lowres_mv[0][0] != 0x7fff )
409 M32( mvc[i] ) = (M32( lowres_mv[h->mb.i_mb_xy] )*2)&0xfffeffff;
414 /* spatial predictors */
415 if( h->mb.i_neighbour & MB_LEFT )
417 int i_mb_l = h->mb.i_mb_xy - 1;
418 /* skip MBs didn't go through the whole search process, so mvr is undefined */
419 if( !IS_SKIP( h->mb.type[i_mb_l] ) )
420 SET_MVP( mvr[i_mb_l] );
422 if( h->mb.i_neighbour & MB_TOP )
424 int i_mb_t = h->mb.i_mb_top_xy;
425 if( !IS_SKIP( h->mb.type[i_mb_t] ) )
426 SET_MVP( mvr[i_mb_t] );
428 if( h->mb.i_neighbour & MB_TOPLEFT && !IS_SKIP( h->mb.type[i_mb_t - 1] ) )
429 SET_MVP( mvr[i_mb_t-1] );
430 if( h->mb.i_mb_x < h->mb.i_mb_stride - 1 && !IS_SKIP( h->mb.type[i_mb_t + 1] ) )
431 SET_MVP( mvr[i_mb_t+1] );
435 /* temporal predictors */
436 /* FIXME temporal scaling w/ interlace */
437 if( h->fref0[0]->i_ref[0] > 0 && !h->sh.b_mbaff )
439 x264_frame_t *l0 = h->fref0[0];
441 #define SET_TMVP(dx, dy) { \
442 int i_b4 = h->mb.i_b4_xy + dx*4 + dy*4*h->mb.i_b4_stride; \
443 int i_b8 = h->mb.i_b8_xy + dx*2 + dy*2*h->mb.i_b8_stride; \
444 int ref_col = l0->ref[0][i_b8]; \
447 int scale = (h->fdec->i_poc - h->fdec->ref_poc[0][i_ref]) * l0->inv_ref_poc[ref_col];\
448 mvc[i][0] = (l0->mv[0][i_b4][0]*scale + 128) >> 8;\
449 mvc[i][1] = (l0->mv[0][i_b4][1]*scale + 128) >> 8;\
455 if( h->mb.i_mb_x < h->sps->i_mb_width-1 )
457 if( h->mb.i_mb_y < h->sps->i_mb_height-1 )
465 /* Set up a lookup table for delta pocs to reduce an IDIV to an IMUL */
466 static void setup_inverse_delta_pocs( x264_t *h )
469 for( i = 0; i < h->i_ref0; i++ )
471 int delta = h->fdec->i_poc - h->fref0[i]->i_poc;
472 h->fdec->inv_ref_poc[i] = (256 + delta/2) / delta;
476 static NOINLINE void x264_mb_mc_0xywh( x264_t *h, int x, int y, int width, int height )
478 const int i8 = x264_scan8[0]+x+8*y;
479 const int i_ref = h->mb.cache.ref[0][i8];
480 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;
481 int mvy = x264_clip3( h->mb.cache.mv[0][i8][1], h->mb.mv_min[1], h->mb.mv_max[1] ) + 4*4*y;
483 h->mc.mc_luma( &h->mb.pic.p_fdec[0][4*y*FDEC_STRIDE+4*x], FDEC_STRIDE,
484 h->mb.pic.p_fref[0][i_ref], h->mb.pic.i_stride[0],
485 mvx, mvy, 4*width, 4*height, &h->sh.weight[i_ref][0] );
487 // chroma is offset if MCing from a field of opposite parity
488 if( h->mb.b_interlaced & i_ref )
489 mvy += (h->mb.i_mb_y & 1)*4 - 2;
491 h->mc.mc_chroma( &h->mb.pic.p_fdec[1][2*y*FDEC_STRIDE+2*x], FDEC_STRIDE,
492 h->mb.pic.p_fref[0][i_ref][4], h->mb.pic.i_stride[1],
493 mvx, mvy, 2*width, 2*height );
495 if( h->sh.weight[i_ref][1].weightfn )
496 h->sh.weight[i_ref][1].weightfn[width>>1]( &h->mb.pic.p_fdec[1][2*y*FDEC_STRIDE+2*x], FDEC_STRIDE,
497 &h->mb.pic.p_fdec[1][2*y*FDEC_STRIDE+2*x], FDEC_STRIDE,
498 &h->sh.weight[i_ref][1], height*2 );
500 h->mc.mc_chroma( &h->mb.pic.p_fdec[2][2*y*FDEC_STRIDE+2*x], FDEC_STRIDE,
501 h->mb.pic.p_fref[0][i_ref][5], h->mb.pic.i_stride[2],
502 mvx, mvy, 2*width, 2*height );
504 if( h->sh.weight[i_ref][2].weightfn )
505 h->sh.weight[i_ref][2].weightfn[width>>1]( &h->mb.pic.p_fdec[2][2*y*FDEC_STRIDE+2*x], FDEC_STRIDE,
506 &h->mb.pic.p_fdec[2][2*y*FDEC_STRIDE+2*x], FDEC_STRIDE,
507 &h->sh.weight[i_ref][2],height*2 );
510 static NOINLINE void x264_mb_mc_1xywh( x264_t *h, int x, int y, int width, int height )
512 const int i8 = x264_scan8[0]+x+8*y;
513 const int i_ref = h->mb.cache.ref[1][i8];
514 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;
515 int mvy = x264_clip3( h->mb.cache.mv[1][i8][1], h->mb.mv_min[1], h->mb.mv_max[1] ) + 4*4*y;
517 h->mc.mc_luma( &h->mb.pic.p_fdec[0][4*y*FDEC_STRIDE+4*x], FDEC_STRIDE,
518 h->mb.pic.p_fref[1][i_ref], h->mb.pic.i_stride[0],
519 mvx, mvy, 4*width, 4*height, weight_none );
521 if( h->mb.b_interlaced & i_ref )
522 mvy += (h->mb.i_mb_y & 1)*4 - 2;
524 h->mc.mc_chroma( &h->mb.pic.p_fdec[1][2*y*FDEC_STRIDE+2*x], FDEC_STRIDE,
525 h->mb.pic.p_fref[1][i_ref][4], h->mb.pic.i_stride[1],
526 mvx, mvy, 2*width, 2*height );
528 h->mc.mc_chroma( &h->mb.pic.p_fdec[2][2*y*FDEC_STRIDE+2*x], FDEC_STRIDE,
529 h->mb.pic.p_fref[1][i_ref][5], h->mb.pic.i_stride[2],
530 mvx, mvy, 2*width, 2*height );
533 static NOINLINE void x264_mb_mc_01xywh( x264_t *h, int x, int y, int width, int height )
535 const int i8 = x264_scan8[0]+x+8*y;
536 const int i_ref0 = h->mb.cache.ref[0][i8];
537 const int i_ref1 = h->mb.cache.ref[1][i8];
538 const int weight = h->mb.bipred_weight[i_ref0][i_ref1];
539 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;
540 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;
541 int mvy0 = x264_clip3( h->mb.cache.mv[0][i8][1], h->mb.mv_min[1], h->mb.mv_max[1] ) + 4*4*y;
542 int mvy1 = x264_clip3( h->mb.cache.mv[1][i8][1], h->mb.mv_min[1], h->mb.mv_max[1] ) + 4*4*y;
543 int i_mode = x264_size2pixel[height][width];
544 int i_stride0 = 16, i_stride1 = 16;
545 ALIGNED_ARRAY_16( uint8_t, tmp0,[16*16] );
546 ALIGNED_ARRAY_16( uint8_t, tmp1,[16*16] );
547 uint8_t *src0, *src1;
549 src0 = h->mc.get_ref( tmp0, &i_stride0, h->mb.pic.p_fref[0][i_ref0], h->mb.pic.i_stride[0],
550 mvx0, mvy0, 4*width, 4*height, weight_none );
551 src1 = h->mc.get_ref( tmp1, &i_stride1, h->mb.pic.p_fref[1][i_ref1], h->mb.pic.i_stride[0],
552 mvx1, mvy1, 4*width, 4*height, weight_none );
553 h->mc.avg[i_mode]( &h->mb.pic.p_fdec[0][4*y*FDEC_STRIDE+4*x], FDEC_STRIDE,
554 src0, i_stride0, src1, i_stride1, weight );
556 if( h->mb.b_interlaced & i_ref0 )
557 mvy0 += (h->mb.i_mb_y & 1)*4 - 2;
558 if( h->mb.b_interlaced & i_ref1 )
559 mvy1 += (h->mb.i_mb_y & 1)*4 - 2;
561 h->mc.mc_chroma( tmp0, 16, h->mb.pic.p_fref[0][i_ref0][4], h->mb.pic.i_stride[1],
562 mvx0, mvy0, 2*width, 2*height );
563 h->mc.mc_chroma( tmp1, 16, h->mb.pic.p_fref[1][i_ref1][4], h->mb.pic.i_stride[1],
564 mvx1, mvy1, 2*width, 2*height );
565 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 );
566 h->mc.mc_chroma( tmp0, 16, h->mb.pic.p_fref[0][i_ref0][5], h->mb.pic.i_stride[2],
567 mvx0, mvy0, 2*width, 2*height );
568 h->mc.mc_chroma( tmp1, 16, h->mb.pic.p_fref[1][i_ref1][5], h->mb.pic.i_stride[2],
569 mvx1, mvy1, 2*width, 2*height );
570 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 );
573 static void x264_mb_mc_direct8x8( x264_t *h, int x, int y )
575 const int i8 = x264_scan8[0] + x + 8*y;
577 if( h->mb.cache.ref[0][i8] >= 0 )
578 if( h->mb.cache.ref[1][i8] >= 0 )
579 x264_mb_mc_01xywh( h, x, y, 2, 2 );
581 x264_mb_mc_0xywh( h, x, y, 2, 2 );
583 x264_mb_mc_1xywh( h, x, y, 2, 2 );
586 void x264_mb_mc_8x8( x264_t *h, int i8 )
588 const int x = 2*(i8&1);
589 const int y = 2*(i8>>1);
590 switch( h->mb.i_sub_partition[i8] )
593 x264_mb_mc_0xywh( h, x, y, 2, 2 );
596 x264_mb_mc_0xywh( h, x, y+0, 2, 1 );
597 x264_mb_mc_0xywh( h, x, y+1, 2, 1 );
600 x264_mb_mc_0xywh( h, x+0, y, 1, 2 );
601 x264_mb_mc_0xywh( h, x+1, y, 1, 2 );
604 x264_mb_mc_0xywh( h, x+0, y+0, 1, 1 );
605 x264_mb_mc_0xywh( h, x+1, y+0, 1, 1 );
606 x264_mb_mc_0xywh( h, x+0, y+1, 1, 1 );
607 x264_mb_mc_0xywh( h, x+1, y+1, 1, 1 );
610 x264_mb_mc_1xywh( h, x, y, 2, 2 );
613 x264_mb_mc_01xywh( h, x, y, 2, 2 );
616 x264_mb_mc_direct8x8( h, x, y );
621 void x264_mb_mc( x264_t *h )
623 if( h->mb.i_type == P_L0 )
625 if( h->mb.i_partition == D_16x16 )
627 x264_mb_mc_0xywh( h, 0, 0, 4, 4 );
629 else if( h->mb.i_partition == D_16x8 )
631 x264_mb_mc_0xywh( h, 0, 0, 4, 2 );
632 x264_mb_mc_0xywh( h, 0, 2, 4, 2 );
634 else if( h->mb.i_partition == D_8x16 )
636 x264_mb_mc_0xywh( h, 0, 0, 2, 4 );
637 x264_mb_mc_0xywh( h, 2, 0, 2, 4 );
640 else if( h->mb.i_type == P_8x8 || h->mb.i_type == B_8x8 )
643 for( i = 0; i < 4; i++ )
644 x264_mb_mc_8x8( h, i );
646 else if( h->mb.i_type == B_SKIP || h->mb.i_type == B_DIRECT )
648 x264_mb_mc_direct8x8( h, 0, 0 );
649 x264_mb_mc_direct8x8( h, 2, 0 );
650 x264_mb_mc_direct8x8( h, 0, 2 );
651 x264_mb_mc_direct8x8( h, 2, 2 );
655 const uint8_t *b_list0 = x264_mb_type_list_table[h->mb.i_type][0];
656 const uint8_t *b_list1 = x264_mb_type_list_table[h->mb.i_type][1];
658 if( h->mb.i_partition == D_16x16 )
660 if( b_list0[0] && b_list1[0] ) x264_mb_mc_01xywh( h, 0, 0, 4, 4 );
661 else if( b_list0[0] ) x264_mb_mc_0xywh ( h, 0, 0, 4, 4 );
662 else if( b_list1[0] ) x264_mb_mc_1xywh ( h, 0, 0, 4, 4 );
664 else if( h->mb.i_partition == D_16x8 )
666 if( b_list0[0] && b_list1[0] ) x264_mb_mc_01xywh( h, 0, 0, 4, 2 );
667 else if( b_list0[0] ) x264_mb_mc_0xywh ( h, 0, 0, 4, 2 );
668 else if( b_list1[0] ) x264_mb_mc_1xywh ( h, 0, 0, 4, 2 );
670 if( b_list0[1] && b_list1[1] ) x264_mb_mc_01xywh( h, 0, 2, 4, 2 );
671 else if( b_list0[1] ) x264_mb_mc_0xywh ( h, 0, 2, 4, 2 );
672 else if( b_list1[1] ) x264_mb_mc_1xywh ( h, 0, 2, 4, 2 );
674 else if( h->mb.i_partition == D_8x16 )
676 if( b_list0[0] && b_list1[0] ) x264_mb_mc_01xywh( h, 0, 0, 2, 4 );
677 else if( b_list0[0] ) x264_mb_mc_0xywh ( h, 0, 0, 2, 4 );
678 else if( b_list1[0] ) x264_mb_mc_1xywh ( h, 0, 0, 2, 4 );
680 if( b_list0[1] && b_list1[1] ) x264_mb_mc_01xywh( h, 2, 0, 2, 4 );
681 else if( b_list0[1] ) x264_mb_mc_0xywh ( h, 2, 0, 2, 4 );
682 else if( b_list1[1] ) x264_mb_mc_1xywh ( h, 2, 0, 2, 4 );
687 int x264_macroblock_cache_init( x264_t *h )
690 int i_mb_count = h->mb.i_mb_count;
692 h->mb.i_mb_stride = h->sps->i_mb_width;
693 h->mb.i_b8_stride = h->sps->i_mb_width * 2;
694 h->mb.i_b4_stride = h->sps->i_mb_width * 4;
696 h->mb.b_interlaced = h->param.b_interlaced;
698 CHECKED_MALLOC( h->mb.qp, i_mb_count * sizeof(int8_t) );
699 CHECKED_MALLOC( h->mb.cbp, i_mb_count * sizeof(int16_t) );
700 CHECKED_MALLOC( h->mb.skipbp, i_mb_count * sizeof(int8_t) );
701 CHECKED_MALLOC( h->mb.mb_transform_size, i_mb_count * sizeof(int8_t) );
703 /* 0 -> 3 top(4), 4 -> 6 : left(3) */
704 CHECKED_MALLOC( h->mb.intra4x4_pred_mode, i_mb_count * 8 * sizeof(int8_t) );
707 CHECKED_MALLOC( h->mb.non_zero_count, i_mb_count * 24 * sizeof(uint8_t) );
709 if( h->param.b_cabac )
711 CHECKED_MALLOC( h->mb.chroma_pred_mode, i_mb_count * sizeof(int8_t) );
712 CHECKED_MALLOC( h->mb.mvd[0], 2*16 * i_mb_count * sizeof(int16_t) );
713 CHECKED_MALLOC( h->mb.mvd[1], 2*16 * i_mb_count * sizeof(int16_t) );
718 int i_refs = X264_MIN(16, (i ? 1 + !!h->param.i_bframe_pyramid : h->param.i_frame_reference) ) << h->param.b_interlaced;
719 if( h->param.analyse.i_weighted_pred == X264_WEIGHTP_SMART )
720 i_refs = X264_MIN(16, i_refs + 2); //smart weights add two duplicate frames
721 else if( h->param.analyse.i_weighted_pred == X264_WEIGHTP_BLIND )
722 i_refs = X264_MIN(16, i_refs + 1); //blind weights add one duplicate frame
724 for( j=0; j < i_refs; j++ )
725 CHECKED_MALLOC( h->mb.mvr[i][j], 2 * i_mb_count * sizeof(int16_t) );
728 if( h->param.analyse.i_weighted_pred )
730 int i_padv = PADV << h->param.b_interlaced;
731 #define ALIGN(x,a) (((x)+((a)-1))&~((a)-1))
732 int align = h->param.cpu&X264_CPU_CACHELINE_64 ? 64 : h->param.cpu&X264_CPU_CACHELINE_32 ? 32 : 16;
733 int i_stride, luma_plane_size;
736 if( h->param.analyse.i_weighted_pred == X264_WEIGHTP_FAKE )
738 // only need buffer for lookahead
739 if( !h->param.i_sync_lookahead || h == h->thread[h->param.i_threads] )
741 // Fake analysis only works on lowres
742 i_stride = ALIGN( h->sps->i_mb_width*8 + 2*PADH, align );
743 luma_plane_size = i_stride * (h->sps->i_mb_height*8+2*i_padv);
744 // Only need 1 buffer for analysis
752 i_stride = ALIGN( h->sps->i_mb_width*16 + 2*PADH, align );
753 luma_plane_size = i_stride * (h->sps->i_mb_height*16+2*i_padv);
755 if( h->param.analyse.i_weighted_pred == X264_WEIGHTP_SMART )
756 //SMART can weight one ref and one offset -1
759 //blind only has one weighted copy (offset -1)
763 for( i = 0; i < numweightbuf; i++ )
764 CHECKED_MALLOC( h->mb.p_weight_buf[i], luma_plane_size );
768 for( i=0; i<=h->param.b_interlaced; i++ )
771 /* shouldn't really be initialized, just silences a valgrind false-positive in predict_8x8_filter_mmx */
772 CHECKED_MALLOCZERO( h->mb.intra_border_backup[i][j], (h->sps->i_mb_width*16+32)>>!!j );
773 h->mb.intra_border_backup[i][j] += 8;
776 /* init with not available (for top right idx=7,15) */
777 memset( h->mb.cache.ref[0], -2, X264_SCAN8_SIZE * sizeof( int8_t ) );
778 memset( h->mb.cache.ref[1], -2, X264_SCAN8_SIZE * sizeof( int8_t ) );
783 void x264_macroblock_cache_end( x264_t *h )
786 for( i=0; i<=h->param.b_interlaced; i++ )
788 x264_free( h->mb.intra_border_backup[i][j] - 8 );
790 for( j=0; j<32; j++ )
791 x264_free( h->mb.mvr[i][j] );
792 for( i=0; i<16; i++ )
793 x264_free( h->mb.p_weight_buf[i] );
795 if( h->param.b_cabac )
797 x264_free( h->mb.chroma_pred_mode );
798 x264_free( h->mb.mvd[0] );
799 x264_free( h->mb.mvd[1] );
801 x264_free( h->mb.intra4x4_pred_mode );
802 x264_free( h->mb.non_zero_count );
803 x264_free( h->mb.mb_transform_size );
804 x264_free( h->mb.skipbp );
805 x264_free( h->mb.cbp );
806 x264_free( h->mb.qp );
808 void x264_macroblock_slice_init( x264_t *h )
812 h->mb.mv[0] = h->fdec->mv[0];
813 h->mb.mv[1] = h->fdec->mv[1];
814 h->mb.ref[0] = h->fdec->ref[0];
815 h->mb.ref[1] = h->fdec->ref[1];
816 h->mb.type = h->fdec->mb_type;
818 h->fdec->i_ref[0] = h->i_ref0;
819 h->fdec->i_ref[1] = h->i_ref1;
820 for( i = 0; i < h->i_ref0; i++ )
821 h->fdec->ref_poc[0][i] = h->fref0[i]->i_poc;
822 if( h->sh.i_type == SLICE_TYPE_B )
824 for( i = 0; i < h->i_ref1; i++ )
825 h->fdec->ref_poc[1][i] = h->fref1[i]->i_poc;
827 map_col_to_list0(-1) = -1;
828 map_col_to_list0(-2) = -2;
829 for( i = 0; i < h->fref1[0]->i_ref[0]; i++ )
831 int poc = h->fref1[0]->ref_poc[0][i];
832 map_col_to_list0(i) = -2;
833 for( j = 0; j < h->i_ref0; j++ )
834 if( h->fref0[j]->i_poc == poc )
836 map_col_to_list0(i) = j;
841 if( h->sh.i_type == SLICE_TYPE_P )
842 memset( h->mb.cache.skip, 0, X264_SCAN8_SIZE * sizeof( int8_t ) );
844 setup_inverse_delta_pocs( h );
846 h->mb.i_neighbour4[6] =
847 h->mb.i_neighbour4[9] =
848 h->mb.i_neighbour4[12] =
849 h->mb.i_neighbour4[14] = MB_LEFT|MB_TOP|MB_TOPLEFT|MB_TOPRIGHT;
850 h->mb.i_neighbour4[3] =
851 h->mb.i_neighbour4[7] =
852 h->mb.i_neighbour4[11] =
853 h->mb.i_neighbour4[13] =
854 h->mb.i_neighbour4[15] =
855 h->mb.i_neighbour8[3] = MB_LEFT|MB_TOP|MB_TOPLEFT;
858 void x264_macroblock_thread_init( x264_t *h )
870 h->mb.pic.p_fenc[0] = h->mb.pic.fenc_buf;
871 h->mb.pic.p_fenc[1] = h->mb.pic.fenc_buf + 16*FENC_STRIDE;
872 h->mb.pic.p_fenc[2] = h->mb.pic.fenc_buf + 16*FENC_STRIDE + 8;
873 h->mb.pic.p_fdec[0] = h->mb.pic.fdec_buf + 2*FDEC_STRIDE;
874 h->mb.pic.p_fdec[1] = h->mb.pic.fdec_buf + 19*FDEC_STRIDE;
875 h->mb.pic.p_fdec[2] = h->mb.pic.fdec_buf + 19*FDEC_STRIDE + 16;
878 void x264_prefetch_fenc( x264_t *h, x264_frame_t *fenc, int i_mb_x, int i_mb_y )
880 int stride_y = fenc->i_stride[0];
881 int stride_uv = fenc->i_stride[1];
882 int off_y = 16 * (i_mb_x + i_mb_y * stride_y);
883 int off_uv = 8 * (i_mb_x + i_mb_y * stride_uv);
884 h->mc.prefetch_fenc( fenc->plane[0]+off_y, stride_y,
885 fenc->plane[1+(i_mb_x&1)]+off_uv, stride_uv, i_mb_x );
888 static NOINLINE void copy_column8( uint8_t *dst, uint8_t *src )
890 // input pointers are offset by 4 rows because that's faster (smaller instruction size on x86)
892 for( i = -4; i < 4; i++ )
893 dst[i*FDEC_STRIDE] = src[i*FDEC_STRIDE];
896 static void ALWAYS_INLINE x264_macroblock_load_pic_pointers( x264_t *h, int i_mb_x, int i_mb_y, int i)
898 const int w = (i == 0 ? 16 : 8);
899 const int i_stride = h->fdec->i_stride[!!i];
900 const int i_stride2 = i_stride << h->mb.b_interlaced;
901 const int i_pix_offset = h->mb.b_interlaced
902 ? w * (i_mb_x + (i_mb_y&~1) * i_stride) + (i_mb_y&1) * i_stride
903 : w * (i_mb_x + i_mb_y * i_stride);
904 const uint8_t *plane_fdec = &h->fdec->plane[i][i_pix_offset];
905 const uint8_t *intra_fdec = h->param.b_sliced_threads ? plane_fdec-i_stride2 :
906 &h->mb.intra_border_backup[i_mb_y & h->sh.b_mbaff][i][i_mb_x*16>>!!i];
907 int ref_pix_offset[2] = { i_pix_offset, i_pix_offset };
908 x264_frame_t **fref[2] = { h->fref0, h->fref1 };
910 if( h->mb.b_interlaced )
911 ref_pix_offset[1] += (1-2*(i_mb_y&1)) * i_stride;
912 h->mb.pic.i_stride[i] = i_stride2;
913 h->mb.pic.p_fenc_plane[i] = &h->fenc->plane[i][i_pix_offset];
914 h->mc.copy[i?PIXEL_8x8:PIXEL_16x16]( h->mb.pic.p_fenc[i], FENC_STRIDE,
915 h->mb.pic.p_fenc_plane[i], i_stride2, w );
917 memcpy( &h->mb.pic.p_fdec[i][-1-FDEC_STRIDE], intra_fdec-1, w*3/2+1 );
919 memset( &h->mb.pic.p_fdec[i][-1-FDEC_STRIDE], 0, w*3/2+1 );
920 if( h->mb.b_interlaced || h->mb.b_reencode_mb )
921 for( j = 0; j < w; j++ )
922 h->mb.pic.p_fdec[i][-1+j*FDEC_STRIDE] = plane_fdec[-1+j*i_stride2];
923 for( j = 0; j < h->mb.pic.i_fref[0]; j++ )
925 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]];
928 for( k = 1; k < 4; k++ )
929 h->mb.pic.p_fref[0][j][k] = &fref[0][j >> h->mb.b_interlaced]->filtered[k][ref_pix_offset[j&1]];
930 if( h->sh.weight[j][0].weightfn )
931 h->mb.pic.p_fref_w[j] = &h->fenc->weighted[j >> h->mb.b_interlaced][ref_pix_offset[j&1]];
933 h->mb.pic.p_fref_w[j] = h->mb.pic.p_fref[0][j][0];
936 if( h->sh.i_type == SLICE_TYPE_B )
937 for( j = 0; j < h->mb.pic.i_fref[1]; j++ )
939 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]];
941 for( k = 1; k < 4; k++ )
942 h->mb.pic.p_fref[1][j][k] = &fref[1][j >> h->mb.b_interlaced]->filtered[k][ref_pix_offset[j&1]];
946 void x264_macroblock_cache_load( x264_t *h, int i_mb_x, int i_mb_y )
948 int i_mb_xy = i_mb_y * h->mb.i_mb_stride + i_mb_x;
949 int i_mb_4x4 = 4*(i_mb_y * h->mb.i_b4_stride + i_mb_x);
950 int i_mb_8x8 = 2*(i_mb_y * h->mb.i_b8_stride + i_mb_x);
951 int i_top_y = i_mb_y - (1 << h->mb.b_interlaced);
952 int i_top_xy = i_top_y * h->mb.i_mb_stride + i_mb_x;
953 int i_top_4x4 = (4*i_top_y+3) * h->mb.i_b4_stride + 4*i_mb_x;
954 int i_top_8x8 = (2*i_top_y+1) * h->mb.i_b8_stride + 2*i_mb_x;
956 int i_top_type = -1; /* gcc warn */
962 h->mb.i_mb_x = i_mb_x;
963 h->mb.i_mb_y = i_mb_y;
964 h->mb.i_mb_xy = i_mb_xy;
965 h->mb.i_b8_xy = i_mb_8x8;
966 h->mb.i_b4_xy = i_mb_4x4;
967 h->mb.i_mb_top_xy = i_top_xy;
968 h->mb.i_neighbour = 0;
969 h->mb.i_neighbour_intra = 0;
972 if( i_top_xy >= h->sh.i_first_mb )
974 h->mb.i_mb_type_top =
975 i_top_type = h->mb.type[i_top_xy];
976 h->mb.cache.i_cbp_top = h->mb.cbp[i_top_xy];
978 h->mb.i_neighbour |= MB_TOP;
980 if( !h->param.b_constrained_intra || IS_INTRA( i_top_type ) )
981 h->mb.i_neighbour_intra |= MB_TOP;
984 CP32( &h->mb.cache.intra4x4_pred_mode[x264_scan8[0] - 8], &h->mb.intra4x4_pred_mode[i_top_xy][0] );
986 /* load non_zero_count */
987 CP32( &h->mb.cache.non_zero_count[x264_scan8[0] - 8], &h->mb.non_zero_count[i_top_xy][12] );
988 /* shift because x264_scan8[16] is misaligned */
989 M32( &h->mb.cache.non_zero_count[x264_scan8[16+0] - 9] ) = M16( &h->mb.non_zero_count[i_top_xy][18] ) << 8;
990 M32( &h->mb.cache.non_zero_count[x264_scan8[16+4] - 9] ) = M16( &h->mb.non_zero_count[i_top_xy][22] ) << 8;
994 h->mb.i_mb_type_top = -1;
995 h->mb.cache.i_cbp_top = -1;
998 M32( &h->mb.cache.intra4x4_pred_mode[x264_scan8[0] - 8] ) = 0xFFFFFFFFU;
1000 /* load non_zero_count */
1001 M32( &h->mb.cache.non_zero_count[x264_scan8[ 0] - 8] ) = 0x80808080U;
1002 M32( &h->mb.cache.non_zero_count[x264_scan8[16+0] - 9] ) = 0x80808080U;
1003 M32( &h->mb.cache.non_zero_count[x264_scan8[16+4] - 9] ) = 0x80808080U;
1006 if( i_mb_x > 0 && i_mb_xy > h->sh.i_first_mb )
1008 i_left_xy = i_mb_xy - 1;
1009 h->mb.i_mb_type_left =
1010 i_left_type = h->mb.type[i_left_xy];
1011 h->mb.cache.i_cbp_left = h->mb.cbp[h->mb.i_mb_xy - 1];
1013 h->mb.i_neighbour |= MB_LEFT;
1015 if( !h->param.b_constrained_intra || IS_INTRA( i_left_type ) )
1016 h->mb.i_neighbour_intra |= MB_LEFT;
1019 h->mb.cache.intra4x4_pred_mode[x264_scan8[0 ] - 1] = h->mb.intra4x4_pred_mode[i_left_xy][4];
1020 h->mb.cache.intra4x4_pred_mode[x264_scan8[2 ] - 1] = h->mb.intra4x4_pred_mode[i_left_xy][5];
1021 h->mb.cache.intra4x4_pred_mode[x264_scan8[8 ] - 1] = h->mb.intra4x4_pred_mode[i_left_xy][6];
1022 h->mb.cache.intra4x4_pred_mode[x264_scan8[10] - 1] = h->mb.intra4x4_pred_mode[i_left_xy][3];
1024 /* load non_zero_count */
1025 h->mb.cache.non_zero_count[x264_scan8[0 ] - 1] = h->mb.non_zero_count[i_left_xy][3];
1026 h->mb.cache.non_zero_count[x264_scan8[2 ] - 1] = h->mb.non_zero_count[i_left_xy][7];
1027 h->mb.cache.non_zero_count[x264_scan8[8 ] - 1] = h->mb.non_zero_count[i_left_xy][11];
1028 h->mb.cache.non_zero_count[x264_scan8[10] - 1] = h->mb.non_zero_count[i_left_xy][15];
1030 h->mb.cache.non_zero_count[x264_scan8[16+0] - 1] = h->mb.non_zero_count[i_left_xy][16+1];
1031 h->mb.cache.non_zero_count[x264_scan8[16+2] - 1] = h->mb.non_zero_count[i_left_xy][16+3];
1033 h->mb.cache.non_zero_count[x264_scan8[16+4+0] - 1] = h->mb.non_zero_count[i_left_xy][16+4+1];
1034 h->mb.cache.non_zero_count[x264_scan8[16+4+2] - 1] = h->mb.non_zero_count[i_left_xy][16+4+3];
1038 h->mb.i_mb_type_left = -1;
1039 h->mb.cache.i_cbp_left = -1;
1041 h->mb.cache.intra4x4_pred_mode[x264_scan8[0 ] - 1] =
1042 h->mb.cache.intra4x4_pred_mode[x264_scan8[2 ] - 1] =
1043 h->mb.cache.intra4x4_pred_mode[x264_scan8[8 ] - 1] =
1044 h->mb.cache.intra4x4_pred_mode[x264_scan8[10] - 1] = -1;
1046 /* load non_zero_count */
1047 h->mb.cache.non_zero_count[x264_scan8[0 ] - 1] =
1048 h->mb.cache.non_zero_count[x264_scan8[2 ] - 1] =
1049 h->mb.cache.non_zero_count[x264_scan8[8 ] - 1] =
1050 h->mb.cache.non_zero_count[x264_scan8[10] - 1] =
1051 h->mb.cache.non_zero_count[x264_scan8[16+0] - 1] =
1052 h->mb.cache.non_zero_count[x264_scan8[16+2] - 1] =
1053 h->mb.cache.non_zero_count[x264_scan8[16+4+0] - 1] =
1054 h->mb.cache.non_zero_count[x264_scan8[16+4+2] - 1] = 0x80;
1057 if( i_mb_x < h->sps->i_mb_width - 1 && i_top_xy + 1 >= h->sh.i_first_mb )
1059 h->mb.i_neighbour |= MB_TOPRIGHT;
1060 h->mb.i_mb_type_topright = h->mb.type[ i_top_xy + 1 ];
1061 if( !h->param.b_constrained_intra || IS_INTRA( h->mb.i_mb_type_topright ) )
1062 h->mb.i_neighbour_intra |= MB_TOPRIGHT;
1065 h->mb.i_mb_type_topright = -1;
1066 if( i_mb_x > 0 && i_top_xy - 1 >= h->sh.i_first_mb )
1068 h->mb.i_neighbour |= MB_TOPLEFT;
1069 h->mb.i_mb_type_topleft = h->mb.type[ i_top_xy - 1 ];
1070 if( !h->param.b_constrained_intra || IS_INTRA( h->mb.i_mb_type_topleft ) )
1071 h->mb.i_neighbour_intra |= MB_TOPLEFT;
1074 h->mb.i_mb_type_topleft = -1;
1076 if( h->pps->b_transform_8x8_mode )
1078 h->mb.cache.i_neighbour_transform_size =
1079 ( i_left_type >= 0 && h->mb.mb_transform_size[i_left_xy] )
1080 + ( i_top_type >= 0 && h->mb.mb_transform_size[i_top_xy] );
1085 h->mb.pic.i_fref[0] = h->i_ref0 << h->mb.b_interlaced;
1086 h->mb.pic.i_fref[1] = h->i_ref1 << h->mb.b_interlaced;
1087 h->mb.cache.i_neighbour_interlaced =
1088 !!(h->mb.i_neighbour & MB_LEFT)
1089 + !!(h->mb.i_neighbour & MB_TOP);
1092 if( !h->mb.b_interlaced && !h->mb.b_reencode_mb )
1094 copy_column8( h->mb.pic.p_fdec[0]-1+ 4*FDEC_STRIDE, h->mb.pic.p_fdec[0]+15+ 4*FDEC_STRIDE );
1095 copy_column8( h->mb.pic.p_fdec[0]-1+12*FDEC_STRIDE, h->mb.pic.p_fdec[0]+15+12*FDEC_STRIDE );
1096 copy_column8( h->mb.pic.p_fdec[1]-1+ 4*FDEC_STRIDE, h->mb.pic.p_fdec[1]+ 7+ 4*FDEC_STRIDE );
1097 copy_column8( h->mb.pic.p_fdec[2]-1+ 4*FDEC_STRIDE, h->mb.pic.p_fdec[2]+ 7+ 4*FDEC_STRIDE );
1100 /* load picture pointers */
1101 x264_macroblock_load_pic_pointers( h, i_mb_x, i_mb_y, 0 );
1102 x264_macroblock_load_pic_pointers( h, i_mb_x, i_mb_y, 1 );
1103 x264_macroblock_load_pic_pointers( h, i_mb_x, i_mb_y, 2 );
1105 if( h->fdec->integral )
1107 assert( !h->mb.b_interlaced );
1108 for( i = 0; i < h->mb.pic.i_fref[0]; i++ )
1109 h->mb.pic.p_integral[0][i] = &h->fref0[i]->integral[ 16 * ( i_mb_x + i_mb_y * h->fdec->i_stride[0] )];
1110 for( i = 0; i < h->mb.pic.i_fref[1]; i++ )
1111 h->mb.pic.p_integral[1][i] = &h->fref1[i]->integral[ 16 * ( i_mb_x + i_mb_y * h->fdec->i_stride[0] )];
1114 x264_prefetch_fenc( h, h->fenc, i_mb_x, i_mb_y );
1116 /* load ref/mv/mvd */
1117 if( h->sh.i_type != SLICE_TYPE_I )
1119 const int s8x8 = h->mb.i_b8_stride;
1120 const int s4x4 = h->mb.i_b4_stride;
1124 for( i_list = 0; i_list < (h->sh.i_type == SLICE_TYPE_B ? 2 : 1 ); i_list++ )
1127 h->mb.cache.ref[i_list][x264_scan8[5 ]+1] =
1128 h->mb.cache.ref[i_list][x264_scan8[7 ]+1] =
1129 h->mb.cache.ref[i_list][x264_scan8[13]+1] = -2;
1132 if( h->mb.i_neighbour & MB_TOPLEFT )
1134 const int i8 = x264_scan8[0] - 1 - 1*8;
1135 const int ir = i_top_8x8 - 1;
1136 const int iv = i_top_4x4 - 1;
1137 h->mb.cache.ref[i_list][i8] = h->mb.ref[i_list][ir];
1138 CP32( h->mb.cache.mv[i_list][i8], h->mb.mv[i_list][iv] );
1142 const int i8 = x264_scan8[0] - 1 - 1*8;
1143 h->mb.cache.ref[i_list][i8] = -2;
1144 M32( h->mb.cache.mv[i_list][i8] ) = 0;
1147 if( h->mb.i_neighbour & MB_TOP )
1149 const int i8 = x264_scan8[0] - 8;
1150 const int ir = i_top_8x8;
1151 const int iv = i_top_4x4;
1152 h->mb.cache.ref[i_list][i8+0] =
1153 h->mb.cache.ref[i_list][i8+1] = h->mb.ref[i_list][ir + 0];
1154 h->mb.cache.ref[i_list][i8+2] =
1155 h->mb.cache.ref[i_list][i8+3] = h->mb.ref[i_list][ir + 1];
1156 CP64( h->mb.cache.mv[i_list][i8+0], h->mb.mv[i_list][iv+0] );
1157 CP64( h->mb.cache.mv[i_list][i8+2], h->mb.mv[i_list][iv+2] );
1161 const int i8 = x264_scan8[0] - 8;
1162 M64( h->mb.cache.mv[i_list][i8+0] ) = 0;
1163 M64( h->mb.cache.mv[i_list][i8+2] ) = 0;
1164 M32( &h->mb.cache.ref[i_list][i8] ) = (uint8_t)(-2) * 0x01010101U;
1167 if( h->mb.i_neighbour & MB_TOPRIGHT )
1169 const int i8 = x264_scan8[0] + 4 - 1*8;
1170 const int ir = i_top_8x8 + 2;
1171 const int iv = i_top_4x4 + 4;
1172 h->mb.cache.ref[i_list][i8] = h->mb.ref[i_list][ir];
1173 CP32( h->mb.cache.mv[i_list][i8], h->mb.mv[i_list][iv] );
1177 const int i8 = x264_scan8[0] + 4 - 1*8;
1178 h->mb.cache.ref[i_list][i8] = -2;
1179 M32( h->mb.cache.mv[i_list][i8] ) = 0;
1182 if( h->mb.i_neighbour & MB_LEFT )
1184 const int i8 = x264_scan8[0] - 1;
1185 const int ir = i_mb_8x8 - 1;
1186 const int iv = i_mb_4x4 - 1;
1187 h->mb.cache.ref[i_list][i8+0*8] =
1188 h->mb.cache.ref[i_list][i8+1*8] = h->mb.ref[i_list][ir + 0*s8x8];
1189 h->mb.cache.ref[i_list][i8+2*8] =
1190 h->mb.cache.ref[i_list][i8+3*8] = h->mb.ref[i_list][ir + 1*s8x8];
1192 CP32( h->mb.cache.mv[i_list][i8+0*8], h->mb.mv[i_list][iv + 0*s4x4] );
1193 CP32( h->mb.cache.mv[i_list][i8+1*8], h->mb.mv[i_list][iv + 1*s4x4] );
1194 CP32( h->mb.cache.mv[i_list][i8+2*8], h->mb.mv[i_list][iv + 2*s4x4] );
1195 CP32( h->mb.cache.mv[i_list][i8+3*8], h->mb.mv[i_list][iv + 3*s4x4] );
1199 const int i8 = x264_scan8[0] - 1;
1200 for( i = 0; i < 4; i++ )
1202 h->mb.cache.ref[i_list][i8+i*8] = -2;
1203 M32( h->mb.cache.mv[i_list][i8+i*8] ) = 0;
1207 if( h->param.b_cabac )
1209 if( i_top_type >= 0 )
1211 const int i8 = x264_scan8[0] - 8;
1212 const int iv = i_top_4x4;
1213 CP64( h->mb.cache.mvd[i_list][i8+0], h->mb.mvd[i_list][iv+0] );
1214 CP64( h->mb.cache.mvd[i_list][i8+2], h->mb.mvd[i_list][iv+2] );
1218 const int i8 = x264_scan8[0] - 8;
1219 M64( h->mb.cache.mvd[i_list][i8+0] ) = 0;
1220 M64( h->mb.cache.mvd[i_list][i8+2] ) = 0;
1223 if( i_left_type >= 0 )
1225 const int i8 = x264_scan8[0] - 1;
1226 const int iv = i_mb_4x4 - 1;
1227 CP32( h->mb.cache.mvd[i_list][i8+0*8], h->mb.mvd[i_list][iv + 0*s4x4] );
1228 CP32( h->mb.cache.mvd[i_list][i8+1*8], h->mb.mvd[i_list][iv + 1*s4x4] );
1229 CP32( h->mb.cache.mvd[i_list][i8+2*8], h->mb.mvd[i_list][iv + 2*s4x4] );
1230 CP32( h->mb.cache.mvd[i_list][i8+3*8], h->mb.mvd[i_list][iv + 3*s4x4] );
1234 const int i8 = x264_scan8[0] - 1;
1235 for( i = 0; i < 4; i++ )
1236 M32( h->mb.cache.mvd[i_list][i8+i*8] ) = 0;
1242 if( h->sh.i_type == SLICE_TYPE_B )
1244 h->mb.bipred_weight = h->mb.bipred_weight_buf[h->mb.b_interlaced&(i_mb_y&1)];
1245 h->mb.dist_scale_factor = h->mb.dist_scale_factor_buf[h->mb.b_interlaced&(i_mb_y&1)];
1246 if( h->param.b_cabac )
1249 x264_macroblock_cache_skip( h, 0, 0, 4, 4, 0 );
1250 skipbp = i_left_type >= 0 ? h->mb.skipbp[i_left_xy] : 0;
1251 h->mb.cache.skip[x264_scan8[0] - 1] = skipbp & 0x2;
1252 h->mb.cache.skip[x264_scan8[8] - 1] = skipbp & 0x8;
1253 skipbp = i_top_type >= 0 ? h->mb.skipbp[i_top_xy] : 0;
1254 h->mb.cache.skip[x264_scan8[0] - 8] = skipbp & 0x4;
1255 h->mb.cache.skip[x264_scan8[4] - 8] = skipbp & 0x8;
1259 if( h->sh.i_type == SLICE_TYPE_P )
1260 x264_mb_predict_mv_pskip( h, h->mb.cache.pskip_mv );
1263 h->mb.i_neighbour4[0] =
1264 h->mb.i_neighbour8[0] = (h->mb.i_neighbour_intra & (MB_TOP|MB_LEFT|MB_TOPLEFT))
1265 | ((h->mb.i_neighbour_intra & MB_TOP) ? MB_TOPRIGHT : 0);
1266 h->mb.i_neighbour4[4] =
1267 h->mb.i_neighbour4[1] = MB_LEFT | ((h->mb.i_neighbour_intra & MB_TOP) ? (MB_TOP|MB_TOPLEFT|MB_TOPRIGHT) : 0);
1268 h->mb.i_neighbour4[2] =
1269 h->mb.i_neighbour4[8] =
1270 h->mb.i_neighbour4[10] =
1271 h->mb.i_neighbour8[2] = MB_TOP|MB_TOPRIGHT | ((h->mb.i_neighbour_intra & MB_LEFT) ? (MB_LEFT|MB_TOPLEFT) : 0);
1272 h->mb.i_neighbour4[5] =
1273 h->mb.i_neighbour8[1] = MB_LEFT | (h->mb.i_neighbour_intra & MB_TOPRIGHT)
1274 | ((h->mb.i_neighbour_intra & MB_TOP) ? MB_TOP|MB_TOPLEFT : 0);
1277 static void ALWAYS_INLINE x264_macroblock_store_pic( x264_t *h, int i )
1280 int i_stride = h->fdec->i_stride[!!i];
1281 int i_stride2 = i_stride << h->mb.b_interlaced;
1282 int i_pix_offset = h->mb.b_interlaced
1283 ? w * (h->mb.i_mb_x + (h->mb.i_mb_y&~1) * i_stride) + (h->mb.i_mb_y&1) * i_stride
1284 : w * (h->mb.i_mb_x + h->mb.i_mb_y * i_stride);
1285 h->mc.copy[i?PIXEL_8x8:PIXEL_16x16](
1286 &h->fdec->plane[i][i_pix_offset], i_stride2,
1287 h->mb.pic.p_fdec[i], FDEC_STRIDE, w );
1290 void x264_macroblock_cache_save( x264_t *h )
1292 const int i_mb_xy = h->mb.i_mb_xy;
1293 const int i_mb_type = x264_mb_type_fix[h->mb.i_type];
1294 const int s8x8 = h->mb.i_b8_stride;
1295 const int s4x4 = h->mb.i_b4_stride;
1296 const int i_mb_4x4 = h->mb.i_b4_xy;
1297 const int i_mb_8x8 = h->mb.i_b8_xy;
1299 /* GCC pessimizes direct stores to heap-allocated 8-bit arrays due to aliasing.*/
1300 /* By only dereferencing them once, we avoid this issue. */
1301 int8_t *intra4x4_pred_mode = h->mb.intra4x4_pred_mode[i_mb_xy];
1302 uint8_t *non_zero_count = h->mb.non_zero_count[i_mb_xy];
1306 x264_macroblock_store_pic( h, 0 );
1307 x264_macroblock_store_pic( h, 1 );
1308 x264_macroblock_store_pic( h, 2 );
1310 x264_prefetch_fenc( h, h->fdec, h->mb.i_mb_x, h->mb.i_mb_y );
1312 h->mb.type[i_mb_xy] = i_mb_type;
1313 h->mb.i_mb_prev_xy = i_mb_xy;
1316 if( i_mb_type == I_4x4 )
1318 CP32( &intra4x4_pred_mode[0], &h->mb.cache.intra4x4_pred_mode[x264_scan8[10]] );
1319 M32( &intra4x4_pred_mode[4] ) = pack8to32(h->mb.cache.intra4x4_pred_mode[x264_scan8[5] ],
1320 h->mb.cache.intra4x4_pred_mode[x264_scan8[7] ],
1321 h->mb.cache.intra4x4_pred_mode[x264_scan8[13] ], 0);
1323 else if( !h->param.b_constrained_intra || IS_INTRA(i_mb_type) )
1324 M64( intra4x4_pred_mode ) = I_PRED_4x4_DC * 0x0101010101010101ULL;
1326 M64( intra4x4_pred_mode ) = (uint8_t)(-1) * 0x0101010101010101ULL;
1329 if( i_mb_type == I_PCM )
1331 h->mb.qp[i_mb_xy] = 0;
1332 h->mb.i_last_dqp = 0;
1333 h->mb.i_cbp_chroma = 2;
1334 h->mb.i_cbp_luma = 0xf;
1335 h->mb.cbp[i_mb_xy] = 0x72f; /* all set */
1336 h->mb.b_transform_8x8 = 0;
1337 memset( non_zero_count, 16, 24 );
1341 /* save non zero count */
1342 CP32( &non_zero_count[0*4], &h->mb.cache.non_zero_count[x264_scan8[0]+0*8] );
1343 CP32( &non_zero_count[1*4], &h->mb.cache.non_zero_count[x264_scan8[0]+1*8] );
1344 CP32( &non_zero_count[2*4], &h->mb.cache.non_zero_count[x264_scan8[0]+2*8] );
1345 CP32( &non_zero_count[3*4], &h->mb.cache.non_zero_count[x264_scan8[0]+3*8] );
1346 M16( &non_zero_count[16+0*2] ) = M32( &h->mb.cache.non_zero_count[x264_scan8[16+0*2]-1] ) >> 8;
1347 M16( &non_zero_count[16+1*2] ) = M32( &h->mb.cache.non_zero_count[x264_scan8[16+1*2]-1] ) >> 8;
1348 M16( &non_zero_count[16+2*2] ) = M32( &h->mb.cache.non_zero_count[x264_scan8[16+2*2]-1] ) >> 8;
1349 M16( &non_zero_count[16+3*2] ) = M32( &h->mb.cache.non_zero_count[x264_scan8[16+3*2]-1] ) >> 8;
1351 if( h->mb.i_type != I_16x16 && h->mb.i_cbp_luma == 0 && h->mb.i_cbp_chroma == 0 )
1352 h->mb.i_qp = h->mb.i_last_qp;
1353 h->mb.qp[i_mb_xy] = h->mb.i_qp;
1354 h->mb.i_last_dqp = h->mb.i_qp - h->mb.i_last_qp;
1355 h->mb.i_last_qp = h->mb.i_qp;
1358 if( h->mb.i_cbp_luma == 0 && h->mb.i_type != I_8x8 )
1359 h->mb.b_transform_8x8 = 0;
1360 h->mb.mb_transform_size[i_mb_xy] = h->mb.b_transform_8x8;
1362 if( h->sh.i_type != SLICE_TYPE_I )
1364 if( !IS_INTRA( i_mb_type ) )
1366 h->mb.ref[0][i_mb_8x8+0+0*s8x8] = h->mb.cache.ref[0][x264_scan8[0]];
1367 h->mb.ref[0][i_mb_8x8+1+0*s8x8] = h->mb.cache.ref[0][x264_scan8[4]];
1368 h->mb.ref[0][i_mb_8x8+0+1*s8x8] = h->mb.cache.ref[0][x264_scan8[8]];
1369 h->mb.ref[0][i_mb_8x8+1+1*s8x8] = h->mb.cache.ref[0][x264_scan8[12]];
1370 for( y = 0; y < 4; y++ )
1372 CP64( h->mb.mv[0][i_mb_4x4+y*s4x4+0], h->mb.cache.mv[0][x264_scan8[0]+8*y+0] );
1373 CP64( h->mb.mv[0][i_mb_4x4+y*s4x4+2], h->mb.cache.mv[0][x264_scan8[0]+8*y+2] );
1375 if( h->sh.i_type == SLICE_TYPE_B )
1377 h->mb.ref[1][i_mb_8x8+0+0*s8x8] = h->mb.cache.ref[1][x264_scan8[0]];
1378 h->mb.ref[1][i_mb_8x8+1+0*s8x8] = h->mb.cache.ref[1][x264_scan8[4]];
1379 h->mb.ref[1][i_mb_8x8+0+1*s8x8] = h->mb.cache.ref[1][x264_scan8[8]];
1380 h->mb.ref[1][i_mb_8x8+1+1*s8x8] = h->mb.cache.ref[1][x264_scan8[12]];
1381 for( y = 0; y < 4; y++ )
1383 CP64( h->mb.mv[1][i_mb_4x4+y*s4x4+0], h->mb.cache.mv[1][x264_scan8[0]+8*y+0] );
1384 CP64( h->mb.mv[1][i_mb_4x4+y*s4x4+2], h->mb.cache.mv[1][x264_scan8[0]+8*y+2] );
1391 for( i_list = 0; i_list < (h->sh.i_type == SLICE_TYPE_B ? 2 : 1 ); i_list++ )
1393 M16( &h->mb.ref[i_list][i_mb_8x8+0*s8x8] ) = (uint8_t)(-1) * 0x0101;
1394 M16( &h->mb.ref[i_list][i_mb_8x8+1*s8x8] ) = (uint8_t)(-1) * 0x0101;
1395 for( y = 0; y < 4; y++ )
1397 M64( h->mb.mv[i_list][i_mb_4x4+y*s4x4+0] ) = 0;
1398 M64( h->mb.mv[i_list][i_mb_4x4+y*s4x4+2] ) = 0;
1404 if( h->param.b_cabac )
1406 if( IS_INTRA(i_mb_type) && i_mb_type != I_PCM )
1407 h->mb.chroma_pred_mode[i_mb_xy] = x264_mb_pred_mode8x8c_fix[ h->mb.i_chroma_pred_mode ];
1409 h->mb.chroma_pred_mode[i_mb_xy] = I_PRED_CHROMA_DC;
1411 if( !IS_INTRA( i_mb_type ) && !IS_SKIP( i_mb_type ) && !IS_DIRECT( i_mb_type ) )
1413 for( y = 0; y < 4; y++ )
1415 CP64( h->mb.mvd[0][i_mb_4x4+y*s4x4+0], h->mb.cache.mvd[0][x264_scan8[0]+8*y+0] );
1416 CP64( h->mb.mvd[0][i_mb_4x4+y*s4x4+2], h->mb.cache.mvd[0][x264_scan8[0]+8*y+2] );
1418 if( h->sh.i_type == SLICE_TYPE_B )
1419 for( y = 0; y < 4; y++ )
1421 CP64( h->mb.mvd[1][i_mb_4x4+y*s4x4+0], h->mb.cache.mvd[1][x264_scan8[0]+8*y+0] );
1422 CP64( h->mb.mvd[1][i_mb_4x4+y*s4x4+2], h->mb.cache.mvd[1][x264_scan8[0]+8*y+2] );
1427 for( y = 0; y < 4; y++ )
1429 M64( h->mb.mvd[0][i_mb_4x4+y*s4x4+0] ) = 0;
1430 M64( h->mb.mvd[0][i_mb_4x4+y*s4x4+2] ) = 0;
1432 if( h->sh.i_type == SLICE_TYPE_B )
1433 for( y = 0; y < 4; y++ )
1435 M64( h->mb.mvd[1][i_mb_4x4+y*s4x4+0] ) = 0;
1436 M64( h->mb.mvd[1][i_mb_4x4+y*s4x4+2] ) = 0;
1440 if( h->sh.i_type == SLICE_TYPE_B )
1442 if( i_mb_type == B_SKIP || i_mb_type == B_DIRECT )
1443 h->mb.skipbp[i_mb_xy] = 0xf;
1444 else if( i_mb_type == B_8x8 )
1446 int skipbp = ( h->mb.i_sub_partition[0] == D_DIRECT_8x8 ) << 0;
1447 skipbp |= ( h->mb.i_sub_partition[1] == D_DIRECT_8x8 ) << 1;
1448 skipbp |= ( h->mb.i_sub_partition[2] == D_DIRECT_8x8 ) << 2;
1449 skipbp |= ( h->mb.i_sub_partition[3] == D_DIRECT_8x8 ) << 3;
1450 h->mb.skipbp[i_mb_xy] = skipbp;
1453 h->mb.skipbp[i_mb_xy] = 0;
1459 void x264_macroblock_bipred_init( x264_t *h )
1461 int i_ref0, i_ref1, field;
1462 for( field = 0; field <= h->sh.b_mbaff; field++ )
1463 for( i_ref0 = 0; i_ref0 < (h->i_ref0<<h->sh.b_mbaff); i_ref0++ )
1465 int poc0 = h->fref0[i_ref0>>h->sh.b_mbaff]->i_poc;
1466 if( h->sh.b_mbaff && field^(i_ref0&1) )
1467 poc0 += h->sh.i_delta_poc_bottom;
1468 for( i_ref1 = 0; i_ref1 < (h->i_ref1<<h->sh.b_mbaff); i_ref1++ )
1470 int dist_scale_factor;
1471 int poc1 = h->fref1[i_ref1>>h->sh.b_mbaff]->i_poc;
1472 if( h->sh.b_mbaff && field^(i_ref1&1) )
1473 poc1 += h->sh.i_delta_poc_bottom;
1474 int cur_poc = h->fdec->i_poc + field*h->sh.i_delta_poc_bottom;
1475 int td = x264_clip3( poc1 - poc0, -128, 127 );
1476 if( td == 0 /* || pic0 is a long-term ref */ )
1477 dist_scale_factor = 256;
1480 int tb = x264_clip3( cur_poc - poc0, -128, 127 );
1481 int tx = (16384 + (abs(td) >> 1)) / td;
1482 dist_scale_factor = x264_clip3( (tb * tx + 32) >> 6, -1024, 1023 );
1485 h->mb.dist_scale_factor_buf[field][i_ref0][i_ref1] = dist_scale_factor;
1487 dist_scale_factor >>= 2;
1488 if( h->param.analyse.b_weighted_bipred
1489 && dist_scale_factor >= -64
1490 && dist_scale_factor <= 128 )
1492 h->mb.bipred_weight_buf[field][i_ref0][i_ref1] = 64 - dist_scale_factor;
1493 // ssse3 implementation of biweight doesn't support the extrema.
1494 // if we ever generate them, we'll have to drop that optimization.
1495 assert( dist_scale_factor >= -63 && dist_scale_factor <= 127 );
1498 h->mb.bipred_weight_buf[field][i_ref0][i_ref1] = 32;