1 /*****************************************************************************
2 * me.c: motion estimation
3 *****************************************************************************
4 * Copyright (C) 2003-2012 x264 project
6 * Authors: Loren Merritt <lorenm@u.washington.edu>
7 * Laurent Aimar <fenrir@via.ecp.fr>
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.
24 * This program is also available under a commercial proprietary license.
25 * For more information, contact us at licensing@x264.com.
26 *****************************************************************************/
28 #include "common/common.h"
29 #include "macroblock.h"
32 /* presets selected from good points on the speed-vs-quality curve of several test videos
33 * subpel_iters[i_subpel_refine] = { refine_hpel, refine_qpel, me_hpel, me_qpel }
34 * where me_* are the number of EPZS iterations run on all candidate block types,
35 * and refine_* are run only on the winner.
36 * the subme=8,9 values are much higher because any amount of satd search makes
37 * up its time by reducing the number of qpel-rd iterations. */
38 static const uint8_t subpel_iterations[][4] =
53 static const uint8_t mod6m1[8] = {5,0,1,2,3,4,5,0};
54 /* radius 2 hexagon. repeated entries are to avoid having to compute mod6 every time. */
55 static const int8_t hex2[8][2] = {{-1,-2}, {-2,0}, {-1,2}, {1,2}, {2,0}, {1,-2}, {-1,-2}, {-2,0}};
56 static const int8_t square1[9][2] = {{0,0}, {0,-1}, {0,1}, {-1,0}, {1,0}, {-1,-1}, {-1,1}, {1,-1}, {1,1}};
58 static void refine_subpel( x264_t *h, x264_me_t *m, int hpel_iters, int qpel_iters, int *p_halfpel_thresh, int b_refine_qpel );
60 #define BITS_MVD( mx, my )\
61 (p_cost_mvx[(mx)<<2] + p_cost_mvy[(my)<<2])
63 #define COST_MV( mx, my )\
65 int cost = h->pixf.fpelcmp[i_pixel]( p_fenc, FENC_STRIDE,\
66 &p_fref_w[(my)*stride+(mx)], stride )\
68 COPY3_IF_LT( bcost, cost, bmx, mx, bmy, my );\
71 #define COST_MV_HPEL( mx, my ) \
74 pixel *src = h->mc.get_ref( pix, &stride2, m->p_fref, stride, mx, my, bw, bh, &m->weight[0] ); \
75 int cost = h->pixf.fpelcmp[i_pixel]( p_fenc, FENC_STRIDE, src, stride2 ) \
76 + p_cost_mvx[ mx ] + p_cost_mvy[ my ]; \
77 COPY3_IF_LT( bpred_cost, cost, bpred_mx, mx, bpred_my, my ); \
80 #define COST_MV_X3_DIR( m0x, m0y, m1x, m1y, m2x, m2y, costs )\
82 pixel *pix_base = p_fref_w + bmx + bmy*stride;\
83 h->pixf.fpelcmp_x3[i_pixel]( p_fenc,\
84 pix_base + (m0x) + (m0y)*stride,\
85 pix_base + (m1x) + (m1y)*stride,\
86 pix_base + (m2x) + (m2y)*stride,\
88 (costs)[0] += BITS_MVD( bmx+(m0x), bmy+(m0y) );\
89 (costs)[1] += BITS_MVD( bmx+(m1x), bmy+(m1y) );\
90 (costs)[2] += BITS_MVD( bmx+(m2x), bmy+(m2y) );\
93 #define COST_MV_X4_DIR( m0x, m0y, m1x, m1y, m2x, m2y, m3x, m3y, costs )\
95 pixel *pix_base = p_fref_w + bmx + bmy*stride;\
96 h->pixf.fpelcmp_x4[i_pixel]( p_fenc,\
97 pix_base + (m0x) + (m0y)*stride,\
98 pix_base + (m1x) + (m1y)*stride,\
99 pix_base + (m2x) + (m2y)*stride,\
100 pix_base + (m3x) + (m3y)*stride,\
102 (costs)[0] += BITS_MVD( bmx+(m0x), bmy+(m0y) );\
103 (costs)[1] += BITS_MVD( bmx+(m1x), bmy+(m1y) );\
104 (costs)[2] += BITS_MVD( bmx+(m2x), bmy+(m2y) );\
105 (costs)[3] += BITS_MVD( bmx+(m3x), bmy+(m3y) );\
108 #define COST_MV_X4( m0x, m0y, m1x, m1y, m2x, m2y, m3x, m3y )\
110 pixel *pix_base = p_fref_w + omx + omy*stride;\
111 h->pixf.fpelcmp_x4[i_pixel]( p_fenc,\
112 pix_base + (m0x) + (m0y)*stride,\
113 pix_base + (m1x) + (m1y)*stride,\
114 pix_base + (m2x) + (m2y)*stride,\
115 pix_base + (m3x) + (m3y)*stride,\
117 costs[0] += BITS_MVD( omx+(m0x), omy+(m0y) );\
118 costs[1] += BITS_MVD( omx+(m1x), omy+(m1y) );\
119 costs[2] += BITS_MVD( omx+(m2x), omy+(m2y) );\
120 costs[3] += BITS_MVD( omx+(m3x), omy+(m3y) );\
121 COPY3_IF_LT( bcost, costs[0], bmx, omx+(m0x), bmy, omy+(m0y) );\
122 COPY3_IF_LT( bcost, costs[1], bmx, omx+(m1x), bmy, omy+(m1y) );\
123 COPY3_IF_LT( bcost, costs[2], bmx, omx+(m2x), bmy, omy+(m2y) );\
124 COPY3_IF_LT( bcost, costs[3], bmx, omx+(m3x), bmy, omy+(m3y) );\
127 #define COST_MV_X3_ABS( m0x, m0y, m1x, m1y, m2x, m2y )\
129 h->pixf.fpelcmp_x3[i_pixel]( p_fenc,\
130 p_fref_w + (m0x) + (m0y)*stride,\
131 p_fref_w + (m1x) + (m1y)*stride,\
132 p_fref_w + (m2x) + (m2y)*stride,\
134 costs[0] += p_cost_mvx[(m0x)<<2]; /* no cost_mvy */\
135 costs[1] += p_cost_mvx[(m1x)<<2];\
136 costs[2] += p_cost_mvx[(m2x)<<2];\
137 COPY3_IF_LT( bcost, costs[0], bmx, m0x, bmy, m0y );\
138 COPY3_IF_LT( bcost, costs[1], bmx, m1x, bmy, m1y );\
139 COPY3_IF_LT( bcost, costs[2], bmx, m2x, bmy, m2y );\
145 #define DIA1_ITER( mx, my )\
148 COST_MV_X4( 0,-1, 0,1, -1,0, 1,0 );\
151 #define CROSS( start, x_max, y_max )\
154 if( (x_max) <= X264_MIN(mv_x_max-omx, omx-mv_x_min) )\
155 for( ; i < (x_max)-2; i+=4 )\
156 COST_MV_X4( i,0, -i,0, i+2,0, -i-2,0 );\
157 for( ; i < (x_max); i+=2 )\
159 if( omx+i <= mv_x_max )\
160 COST_MV( omx+i, omy );\
161 if( omx-i >= mv_x_min )\
162 COST_MV( omx-i, omy );\
165 if( (y_max) <= X264_MIN(mv_y_max-omy, omy-mv_y_min) )\
166 for( ; i < (y_max)-2; i+=4 )\
167 COST_MV_X4( 0,i, 0,-i, 0,i+2, 0,-i-2 );\
168 for( ; i < (y_max); i+=2 )\
170 if( omy+i <= mv_y_max )\
171 COST_MV( omx, omy+i );\
172 if( omy-i >= mv_y_min )\
173 COST_MV( omx, omy-i );\
177 void x264_me_search_ref( x264_t *h, x264_me_t *m, int16_t (*mvc)[2], int i_mvc, int *p_halfpel_thresh )
179 const int bw = x264_pixel_size[m->i_pixel].w;
180 const int bh = x264_pixel_size[m->i_pixel].h;
181 const int i_pixel = m->i_pixel;
182 const int stride = m->i_stride[0];
183 int i_me_range = h->param.analyse.i_me_range;
185 int bpred_mx = 0, bpred_my = 0, bpred_cost = COST_MAX;
186 int omx, omy, pmx, pmy;
187 pixel *p_fenc = m->p_fenc[0];
188 pixel *p_fref_w = m->p_fref_w;
189 ALIGNED_ARRAY_16( pixel, pix,[16*16] );
193 int mv_x_min = h->mb.mv_min_fpel[0];
194 int mv_y_min = h->mb.mv_min_fpel[1];
195 int mv_x_max = h->mb.mv_max_fpel[0];
196 int mv_y_max = h->mb.mv_max_fpel[1];
197 int mv_x_min_qpel = mv_x_min << 2;
198 int mv_y_min_qpel = mv_y_min << 2;
199 int mv_x_max_qpel = mv_x_max << 2;
200 int mv_y_max_qpel = mv_y_max << 2;
201 /* Special version of pack to allow shortcuts in CHECK_MVRANGE */
202 #define pack16to32_mask2(mx,my) ((mx<<16)|(my&0x7FFF))
203 uint32_t mv_min = pack16to32_mask2( -mv_x_min, -mv_y_min );
204 uint32_t mv_max = pack16to32_mask2( mv_x_max, mv_y_max )|0x8000;
206 #define CHECK_MVRANGE(mx,my) (!(((pack16to32_mask2(mx,my) + mv_min) | (mv_max - pack16to32_mask2(mx,my))) & 0x80004000))
208 const uint16_t *p_cost_mvx = m->p_cost_mv - m->mvp[0];
209 const uint16_t *p_cost_mvy = m->p_cost_mv - m->mvp[1];
212 bmx = x264_clip3( m->mvp[0], mv_x_min_qpel, mv_x_max_qpel );
213 bmy = x264_clip3( m->mvp[1], mv_y_min_qpel, mv_y_max_qpel );
214 pmx = ( bmx + 2 ) >> 2;
215 pmy = ( bmy + 2 ) >> 2;
218 /* try extra predictors if provided */
219 if( h->mb.i_subpel_refine >= 3 )
221 pmv = pack16to32_mask(bmx,bmy);
223 COST_MV_HPEL( bmx, bmy );
224 for( int i = 0; i < i_mvc; i++ )
226 if( M32( mvc[i] ) && (pmv != M32( mvc[i] )) )
228 int mx = x264_clip3( mvc[i][0], mv_x_min_qpel, mv_x_max_qpel );
229 int my = x264_clip3( mvc[i][1], mv_y_min_qpel, mv_y_max_qpel );
230 COST_MV_HPEL( mx, my );
233 bmx = ( bpred_mx + 2 ) >> 2;
234 bmy = ( bpred_my + 2 ) >> 2;
242 /* Because we are rounding the predicted motion vector to fullpel, there will be
243 * an extra MV cost in 15 out of 16 cases. However, when the predicted MV is
244 * chosen as the best predictor, it is often the case that the subpel search will
245 * result in a vector at or next to the predicted motion vector. Therefore, it is
246 * sensible to omit the cost of the MV from the rounded MVP to avoid unfairly
247 * biasing against use of the predicted motion vector. */
248 bcost = h->pixf.fpelcmp[i_pixel]( p_fenc, FENC_STRIDE, &p_fref_w[bmy*stride+bmx], stride );
249 pmv = pack16to32_mask( bmx, bmy );
252 ALIGNED_ARRAY_8( int16_t, mvc_fpel,[16],[2] );
253 x264_predictor_roundclip( mvc_fpel, mvc, i_mvc, mv_x_min, mv_x_max, mv_y_min, mv_y_max );
255 for( int i = 1; i <= i_mvc; i++ )
257 if( M32( mvc_fpel[i-1] ) && (pmv != M32( mvc[i-1] )) )
259 int mx = mvc_fpel[i-1][0];
260 int my = mvc_fpel[i-1][1];
261 int cost = h->pixf.fpelcmp[i_pixel]( p_fenc, FENC_STRIDE, &p_fref_w[my*stride+mx], stride ) + BITS_MVD( mx, my );
262 cost = (cost << 4) + i;
263 COPY1_IF_LT( bcost, cost );
268 bmx = mvc_fpel[(bcost&15)-1][0];
269 bmy = mvc_fpel[(bcost&15)-1][1];
278 switch( h->mb.i_me_method )
282 /* diamond search, radius 1 */
287 COST_MV_X4_DIR( 0,-1, 0,1, -1,0, 1,0, costs );
288 COPY1_IF_LT( bcost, (costs[0]<<4)+1 );
289 COPY1_IF_LT( bcost, (costs[1]<<4)+3 );
290 COPY1_IF_LT( bcost, (costs[2]<<4)+4 );
291 COPY1_IF_LT( bcost, (costs[3]<<4)+12 );
294 bmx -= (bcost<<28)>>30;
295 bmy -= (bcost<<30)>>30;
297 } while( --i && CHECK_MVRANGE(bmx, bmy) );
305 /* hexagon search, radius 2 */
307 for( int i = 0; i < i_me_range/2; i++ )
309 omx = bmx; omy = bmy;
310 COST_MV( omx-2, omy );
311 COST_MV( omx-1, omy+2 );
312 COST_MV( omx+1, omy+2 );
313 COST_MV( omx+2, omy );
314 COST_MV( omx+1, omy-2 );
315 COST_MV( omx-1, omy-2 );
316 if( bmx == omx && bmy == omy )
318 if( !CHECK_MVRANGE(bmx, bmy) )
322 /* equivalent to the above, but eliminates duplicate candidates */
325 COST_MV_X3_DIR( -2,0, -1, 2, 1, 2, costs );
326 COST_MV_X3_DIR( 2,0, 1,-2, -1,-2, costs+3 );
328 COPY1_IF_LT( bcost, (costs[0]<<3)+2 );
329 COPY1_IF_LT( bcost, (costs[1]<<3)+3 );
330 COPY1_IF_LT( bcost, (costs[2]<<3)+4 );
331 COPY1_IF_LT( bcost, (costs[3]<<3)+5 );
332 COPY1_IF_LT( bcost, (costs[4]<<3)+6 );
333 COPY1_IF_LT( bcost, (costs[5]<<3)+7 );
337 int dir = (bcost&7)-2;
338 bmx += hex2[dir+1][0];
339 bmy += hex2[dir+1][1];
341 /* half hexagon, not overlapping the previous iteration */
342 for( int i = (i_me_range>>1) - 1; i > 0 && CHECK_MVRANGE(bmx, bmy); i-- )
344 COST_MV_X3_DIR( hex2[dir+0][0], hex2[dir+0][1],
345 hex2[dir+1][0], hex2[dir+1][1],
346 hex2[dir+2][0], hex2[dir+2][1],
349 COPY1_IF_LT( bcost, (costs[0]<<3)+1 );
350 COPY1_IF_LT( bcost, (costs[1]<<3)+2 );
351 COPY1_IF_LT( bcost, (costs[2]<<3)+3 );
356 bmx += hex2[dir+1][0];
357 bmy += hex2[dir+1][1];
364 COST_MV_X4_DIR( 0,-1, 0,1, -1,0, 1,0, costs );
365 COPY2_IF_LT( bcost, costs[0], dir, 1 );
366 COPY2_IF_LT( bcost, costs[1], dir, 2 );
367 COPY2_IF_LT( bcost, costs[2], dir, 3 );
368 COPY2_IF_LT( bcost, costs[3], dir, 4 );
369 COST_MV_X4_DIR( -1,-1, -1,1, 1,-1, 1,1, costs );
370 COPY2_IF_LT( bcost, costs[0], dir, 5 );
371 COPY2_IF_LT( bcost, costs[1], dir, 6 );
372 COPY2_IF_LT( bcost, costs[2], dir, 7 );
373 COPY2_IF_LT( bcost, costs[3], dir, 8 );
374 bmx += square1[dir][0];
375 bmy += square1[dir][1];
381 /* Uneven-cross Multi-Hexagon-grid Search
382 * as in JM, except with different early termination */
384 static const uint8_t x264_pixel_size_shift[7] = { 0, 1, 1, 2, 3, 3, 4 };
389 /* refine predictors */
391 DIA1_ITER( pmx, pmy );
395 if( i_pixel == PIXEL_4x4 )
399 if( (bmx | bmy) && ((bmx-pmx) | (bmy-pmy)) )
400 DIA1_ITER( bmx, bmy );
401 if( bcost == ucost2 )
403 omx = bmx; omy = bmy;
405 /* early termination */
406 #define SAD_THRESH(v) ( bcost < ( v >> x264_pixel_size_shift[i_pixel] ) )
407 if( bcost == ucost2 && SAD_THRESH(2000) )
409 COST_MV_X4( 0,-2, -1,-1, 1,-1, -2,0 );
410 COST_MV_X4( 2, 0, -1, 1, 1, 1, 0,2 );
411 if( bcost == ucost1 && SAD_THRESH(500) )
413 if( bcost == ucost2 )
415 int range = (i_me_range>>1) | 1;
416 CROSS( 3, range, range );
417 COST_MV_X4( -1,-2, 1,-2, -2,-1, 2,-1 );
418 COST_MV_X4( -2, 1, 2, 1, -1, 2, 1, 2 );
419 if( bcost == ucost2 )
421 cross_start = range + 2;
425 /* adaptive search range */
428 /* range multipliers based on casual inspection of some statistics of
429 * average distance between current predictor and final mv found by ESA.
430 * these have not been tuned much by actual encoding. */
431 static const uint8_t range_mul[4][4] =
439 int sad_ctx, mvd_ctx;
444 if( i_pixel == PIXEL_16x16 )
445 /* mvc is probably the same as mvp, so the difference isn't meaningful.
446 * but prediction usually isn't too bad, so just use medium range */
449 mvd = abs( m->mvp[0] - mvc[0][0] )
450 + abs( m->mvp[1] - mvc[0][1] );
454 /* calculate the degree of agreement between predictors. */
455 /* in 16x16, mvc includes all the neighbors used to make mvp,
456 * so don't count mvp separately. */
459 if( i_pixel != PIXEL_16x16 )
461 mvd = abs( m->mvp[0] - mvc[0][0] )
462 + abs( m->mvp[1] - mvc[0][1] );
465 mvd += x264_predictor_difference( mvc, i_mvc );
468 sad_ctx = SAD_THRESH(1000) ? 0
469 : SAD_THRESH(2000) ? 1
470 : SAD_THRESH(4000) ? 2 : 3;
471 mvd_ctx = mvd < 10*denom ? 0
473 : mvd < 40*denom ? 2 : 3;
475 i_me_range = i_me_range * range_mul[mvd_ctx][sad_ctx] >> 2;
478 /* FIXME if the above DIA2/OCT2/CROSS found a new mv, it has not updated omx/omy.
479 * we are still centered on the same place as the DIA2. is this desirable? */
480 CROSS( cross_start, i_me_range, i_me_range>>1 );
482 COST_MV_X4( -2,-2, -2,2, 2,-2, 2,2 );
485 omx = bmx; omy = bmy;
486 const uint16_t *p_cost_omvx = p_cost_mvx + omx*4;
487 const uint16_t *p_cost_omvy = p_cost_mvy + omy*4;
491 static const int8_t hex4[16][2] = {
492 { 0,-4}, { 0, 4}, {-2,-3}, { 2,-3},
493 {-4,-2}, { 4,-2}, {-4,-1}, { 4,-1},
494 {-4, 0}, { 4, 0}, {-4, 1}, { 4, 1},
495 {-4, 2}, { 4, 2}, {-2, 3}, { 2, 3},
498 if( 4*i > X264_MIN4( mv_x_max-omx, omx-mv_x_min,
499 mv_y_max-omy, omy-mv_y_min ) )
501 for( int j = 0; j < 16; j++ )
503 int mx = omx + hex4[j][0]*i;
504 int my = omy + hex4[j][1]*i;
505 if( CHECK_MVRANGE(mx, my) )
512 pixel *pix_base = p_fref_w + omx + (omy-4*i)*stride;
514 #define SADS(k,x0,y0,x1,y1,x2,y2,x3,y3)\
515 h->pixf.fpelcmp_x4[i_pixel]( p_fenc,\
516 pix_base x0*i+(y0-2*k+4)*dy,\
517 pix_base x1*i+(y1-2*k+4)*dy,\
518 pix_base x2*i+(y2-2*k+4)*dy,\
519 pix_base x3*i+(y3-2*k+4)*dy,\
520 stride, costs+4*k );\
522 #define ADD_MVCOST(k,x,y) costs[k] += p_cost_omvx[x*4*i] + p_cost_omvy[y*4*i]
523 #define MIN_MV(k,x,y) COPY2_IF_LT( bcost, costs[k], dir, x*16+(y&15) )
524 SADS( 0, +0,-4, +0,+4, -2,-3, +2,-3 );
525 SADS( 1, -4,-2, +4,-2, -4,-1, +4,-1 );
526 SADS( 2, -4,+0, +4,+0, -4,+1, +4,+1 );
527 SADS( 3, -4,+2, +4,+2, -2,+3, +2,+3 );
528 ADD_MVCOST( 0, 0,-4 );
529 ADD_MVCOST( 1, 0, 4 );
530 ADD_MVCOST( 2,-2,-3 );
531 ADD_MVCOST( 3, 2,-3 );
532 ADD_MVCOST( 4,-4,-2 );
533 ADD_MVCOST( 5, 4,-2 );
534 ADD_MVCOST( 6,-4,-1 );
535 ADD_MVCOST( 7, 4,-1 );
536 ADD_MVCOST( 8,-4, 0 );
537 ADD_MVCOST( 9, 4, 0 );
538 ADD_MVCOST( 10,-4, 1 );
539 ADD_MVCOST( 11, 4, 1 );
540 ADD_MVCOST( 12,-4, 2 );
541 ADD_MVCOST( 13, 4, 2 );
542 ADD_MVCOST( 14,-2, 3 );
543 ADD_MVCOST( 15, 2, 3 );
565 bmx = omx + i*(dir>>4);
566 bmy = omy + i*((dir<<28)>>28);
569 } while( ++i <= i_me_range>>2 );
570 if( bmy <= mv_y_max && bmy >= mv_y_min && bmx <= mv_x_max && bmx >= mv_x_min )
578 const int min_x = X264_MAX( bmx - i_me_range, mv_x_min );
579 const int min_y = X264_MAX( bmy - i_me_range, mv_y_min );
580 const int max_x = X264_MIN( bmx + i_me_range, mv_x_max );
581 const int max_y = X264_MIN( bmy + i_me_range, mv_y_max );
582 /* SEA is fastest in multiples of 4 */
583 const int width = (max_x - min_x + 3) & ~3;
585 /* plain old exhaustive search */
586 for( int my = min_y; my <= max_y; my++ )
587 for( int mx = min_x; mx < min_x + width; mx++ )
590 /* successive elimination by comparing DC before a full SAD,
591 * because sum(abs(diff)) >= abs(diff(sum)). */
592 uint16_t *sums_base = m->integral;
593 ALIGNED_16( static pixel zero[8*FENC_STRIDE] ) = {0};
594 ALIGNED_ARRAY_16( int, enc_dc,[4] );
595 int sad_size = i_pixel <= PIXEL_8x8 ? PIXEL_8x8 : PIXEL_4x4;
596 int delta = x264_pixel_size[sad_size].w;
597 int16_t *xs = h->scratch_buffer;
599 uint16_t *cost_fpel_mvx = h->cost_mv_fpel[h->mb.i_qp][-m->mvp[0]&3] + (-m->mvp[0]>>2);
601 h->pixf.sad_x4[sad_size]( zero, p_fenc, p_fenc+delta,
602 p_fenc+delta*FENC_STRIDE, p_fenc+delta+delta*FENC_STRIDE,
603 FENC_STRIDE, enc_dc );
605 sums_base += stride * (h->fenc->i_lines[0] + PADV*2);
606 if( i_pixel == PIXEL_16x16 || i_pixel == PIXEL_8x16 || i_pixel == PIXEL_4x8 )
608 if( i_pixel == PIXEL_8x16 || i_pixel == PIXEL_4x8 )
609 enc_dc[1] = enc_dc[2];
611 if( h->mb.i_me_method == X264_ME_TESA )
613 // ADS threshold, then SAD threshold, then keep the best few SADs, then SATD
614 mvsad_t *mvsads = (mvsad_t *)(xs + ((width+15)&~15) + 4);
615 int nmvsad = 0, limit;
616 int sad_thresh = i_me_range <= 16 ? 10 : i_me_range <= 24 ? 11 : 12;
617 int bsad = h->pixf.sad[i_pixel]( p_fenc, FENC_STRIDE, p_fref_w+bmy*stride+bmx, stride )
618 + BITS_MVD( bmx, bmy );
619 for( int my = min_y; my <= max_y; my++ )
622 int ycost = p_cost_mvy[my<<2];
626 xn = h->pixf.ads[i_pixel]( enc_dc, sums_base + min_x + my * stride, delta,
627 cost_fpel_mvx+min_x, xs, width, bsad * 17 >> 4 );
628 for( i = 0; i < xn-2; i += 3 )
630 pixel *ref = p_fref_w+min_x+my*stride;
632 h->pixf.sad_x3[i_pixel]( p_fenc, ref+xs[i], ref+xs[i+1], ref+xs[i+2], stride, sads );
633 for( int j = 0; j < 3; j++ )
635 int sad = sads[j] + cost_fpel_mvx[xs[i+j]];
636 if( sad < bsad*sad_thresh>>3 )
638 COPY1_IF_LT( bsad, sad );
639 mvsads[nmvsad].sad = sad + ycost;
640 mvsads[nmvsad].mv[0] = min_x+xs[i+j];
641 mvsads[nmvsad].mv[1] = my;
648 int mx = min_x+xs[i];
649 int sad = h->pixf.sad[i_pixel]( p_fenc, FENC_STRIDE, p_fref_w+mx+my*stride, stride )
650 + cost_fpel_mvx[xs[i]];
651 if( sad < bsad*sad_thresh>>3 )
653 COPY1_IF_LT( bsad, sad );
654 mvsads[nmvsad].sad = sad + ycost;
655 mvsads[nmvsad].mv[0] = mx;
656 mvsads[nmvsad].mv[1] = my;
663 limit = i_me_range >> 1;
664 sad_thresh = bsad*sad_thresh>>3;
665 while( nmvsad > limit*2 && sad_thresh > bsad )
668 // halve the range if the domain is too large... eh, close enough
669 sad_thresh = (sad_thresh + bsad) >> 1;
670 for( i = 0; i < nmvsad && mvsads[i].sad <= sad_thresh; i++ );
671 for( int j = i; j < nmvsad; j++ )
674 if( WORD_SIZE == 8 && sizeof(mvsad_t) == 8 )
676 uint64_t mvsad = M64( &mvsads[i] ) = M64( &mvsads[j] );
685 CP32( mvsads[i].mv, mvsads[j].mv );
688 i += (sad - (sad_thresh+1)) >> 31;
692 while( nmvsad > limit )
695 for( int i = 1; i < nmvsad; i++ )
696 if( mvsads[i].sad > mvsads[bi].sad )
699 if( sizeof( mvsad_t ) == sizeof( uint64_t ) )
700 CP64( &mvsads[bi], &mvsads[nmvsad] );
702 mvsads[bi] = mvsads[nmvsad];
704 for( int i = 0; i < nmvsad; i++ )
705 COST_MV( mvsads[i].mv[0], mvsads[i].mv[1] );
710 for( int my = min_y; my <= max_y; my++ )
713 int ycost = p_cost_mvy[my<<2];
717 xn = h->pixf.ads[i_pixel]( enc_dc, sums_base + min_x + my * stride, delta,
718 cost_fpel_mvx+min_x, xs, width, bcost );
719 for( i = 0; i < xn-2; i += 3 )
720 COST_MV_X3_ABS( min_x+xs[i],my, min_x+xs[i+1],my, min_x+xs[i+2],my );
723 COST_MV( min_x+xs[i], my );
732 if( bpred_cost < bcost )
736 m->cost = bpred_cost;
745 /* compute the real cost */
746 m->cost_mv = p_cost_mvx[ m->mv[0] ] + p_cost_mvy[ m->mv[1] ];
747 if( bmx == pmx && bmy == pmy && h->mb.i_subpel_refine < 3 )
748 m->cost += m->cost_mv;
751 if( h->mb.i_subpel_refine >= 2 )
753 int hpel = subpel_iterations[h->mb.i_subpel_refine][2];
754 int qpel = subpel_iterations[h->mb.i_subpel_refine][3];
755 refine_subpel( h, m, hpel, qpel, p_halfpel_thresh, 0 );
760 void x264_me_refine_qpel( x264_t *h, x264_me_t *m )
762 int hpel = subpel_iterations[h->mb.i_subpel_refine][0];
763 int qpel = subpel_iterations[h->mb.i_subpel_refine][1];
765 if( m->i_pixel <= PIXEL_8x8 )
766 m->cost -= m->i_ref_cost;
768 refine_subpel( h, m, hpel, qpel, NULL, 1 );
771 void x264_me_refine_qpel_refdupe( x264_t *h, x264_me_t *m, int *p_halfpel_thresh )
773 refine_subpel( h, m, 0, X264_MIN( 2, subpel_iterations[h->mb.i_subpel_refine][3] ), p_halfpel_thresh, 0 );
776 #define COST_MV_SAD( mx, my ) \
779 pixel *src = h->mc.get_ref( pix, &stride, m->p_fref, m->i_stride[0], mx, my, bw, bh, &m->weight[0] ); \
780 int cost = h->pixf.fpelcmp[i_pixel]( m->p_fenc[0], FENC_STRIDE, src, stride ) \
781 + p_cost_mvx[ mx ] + p_cost_mvy[ my ]; \
782 COPY3_IF_LT( bcost, cost, bmx, mx, bmy, my ); \
785 #define COST_MV_SATD( mx, my, dir ) \
786 if( b_refine_qpel || (dir^1) != odir ) \
789 pixel *src = h->mc.get_ref( pix, &stride, &m->p_fref[0], m->i_stride[0], mx, my, bw, bh, &m->weight[0] ); \
790 int cost = h->pixf.mbcmp_unaligned[i_pixel]( m->p_fenc[0], FENC_STRIDE, src, stride ) \
791 + p_cost_mvx[ mx ] + p_cost_mvy[ my ]; \
792 if( b_chroma_me && cost < bcost ) \
797 src = h->mc.get_ref( pix, &stride, &m->p_fref[4], m->i_stride[1], mx, my, bw, bh, &m->weight[1] ); \
798 cost += h->pixf.mbcmp_unaligned[i_pixel]( m->p_fenc[1], FENC_STRIDE, src, stride ); \
802 src = h->mc.get_ref( pix, &stride, &m->p_fref[8], m->i_stride[2], mx, my, bw, bh, &m->weight[2] ); \
803 cost += h->pixf.mbcmp_unaligned[i_pixel]( m->p_fenc[2], FENC_STRIDE, src, stride ); \
808 h->mc.mc_chroma( pix, pix+8, 16, m->p_fref[4], m->i_stride[1], \
809 mx, 2*(my+mvy_offset)>>chroma_v_shift, bw>>1, bh>>chroma_v_shift ); \
810 if( m->weight[1].weightfn ) \
811 m->weight[1].weightfn[bw>>3]( pix, 16, pix, 16, &m->weight[1], bh>>chroma_v_shift ); \
812 cost += h->pixf.mbcmp[chromapix]( m->p_fenc[1], FENC_STRIDE, pix, 16 ); \
815 if( m->weight[2].weightfn ) \
816 m->weight[2].weightfn[bw>>3]( pix+8, 16, pix+8, 16, &m->weight[2], bh>>chroma_v_shift ); \
817 cost += h->pixf.mbcmp[chromapix]( m->p_fenc[2], FENC_STRIDE, pix+8, 16 ); \
821 COPY4_IF_LT( bcost, cost, bmx, mx, bmy, my, bdir, dir ); \
824 static void refine_subpel( x264_t *h, x264_me_t *m, int hpel_iters, int qpel_iters, int *p_halfpel_thresh, int b_refine_qpel )
826 const int bw = x264_pixel_size[m->i_pixel].w;
827 const int bh = x264_pixel_size[m->i_pixel].h;
828 const uint16_t *p_cost_mvx = m->p_cost_mv - m->mvp[0];
829 const uint16_t *p_cost_mvy = m->p_cost_mv - m->mvp[1];
830 const int i_pixel = m->i_pixel;
831 const int b_chroma_me = h->mb.b_chroma_me && (i_pixel <= PIXEL_8x8 || CHROMA444);
832 int chromapix = h->luma2chroma_pixel[i_pixel];
833 int chroma_v_shift = CHROMA_V_SHIFT;
834 int mvy_offset = chroma_v_shift & MB_INTERLACED & m->i_ref ? (h->mb.i_mb_y & 1)*4 - 2 : 0;
836 ALIGNED_ARRAY_16( pixel, pix,[64*18] ); // really 17x17x2, but round up for alignment
843 /* try the subpel component of the predicted mv */
844 if( hpel_iters && h->mb.i_subpel_refine < 3 )
846 int mx = x264_clip3( m->mvp[0], h->mb.mv_min_spel[0]+2, h->mb.mv_max_spel[0]-2 );
847 int my = x264_clip3( m->mvp[1], h->mb.mv_min_spel[1]+2, h->mb.mv_max_spel[1]-2 );
848 if( (mx-bmx)|(my-bmy) )
849 COST_MV_SAD( mx, my );
852 /* halfpel diamond search */
853 for( int i = hpel_iters; i > 0; i-- )
855 int omx = bmx, omy = bmy;
857 int stride = 64; // candidates are either all hpel or all qpel, so one stride is enough
858 pixel *src0, *src1, *src2, *src3;
859 src0 = h->mc.get_ref( pix, &stride, m->p_fref, m->i_stride[0], omx, omy-2, bw, bh+1, &m->weight[0] );
860 src2 = h->mc.get_ref( pix+32, &stride, m->p_fref, m->i_stride[0], omx-2, omy, bw+4, bh, &m->weight[0] );
861 src1 = src0 + stride;
863 h->pixf.fpelcmp_x4[i_pixel]( m->p_fenc[0], src0, src1, src2, src3, stride, costs );
864 COPY2_IF_LT( bcost, costs[0] + p_cost_mvx[omx ] + p_cost_mvy[omy-2], bmy, omy-2 );
865 COPY2_IF_LT( bcost, costs[1] + p_cost_mvx[omx ] + p_cost_mvy[omy+2], bmy, omy+2 );
866 COPY3_IF_LT( bcost, costs[2] + p_cost_mvx[omx-2] + p_cost_mvy[omy ], bmx, omx-2, bmy, omy );
867 COPY3_IF_LT( bcost, costs[3] + p_cost_mvx[omx+2] + p_cost_mvy[omy ], bmx, omx+2, bmy, omy );
868 if( (bmx == omx) & (bmy == omy) )
872 if( !b_refine_qpel && (h->pixf.mbcmp_unaligned[0] != h->pixf.fpelcmp[0] || b_chroma_me) )
875 COST_MV_SATD( bmx, bmy, -1 );
878 /* early termination when examining multiple reference frames */
879 if( p_halfpel_thresh )
881 if( (bcost*7)>>3 > *p_halfpel_thresh )
886 // don't need cost_mv
889 else if( bcost < *p_halfpel_thresh )
890 *p_halfpel_thresh = bcost;
893 /* quarterpel diamond search */
894 if( h->mb.i_subpel_refine != 1 )
897 for( int i = qpel_iters; i > 0; i-- )
899 if( bmy <= h->mb.mv_min_spel[1] || bmy >= h->mb.mv_max_spel[1] || bmx <= h->mb.mv_min_spel[0] || bmx >= h->mb.mv_max_spel[0] )
902 int omx = bmx, omy = bmy;
903 COST_MV_SATD( omx, omy - 1, 0 );
904 COST_MV_SATD( omx, omy + 1, 1 );
905 COST_MV_SATD( omx - 1, omy, 2 );
906 COST_MV_SATD( omx + 1, omy, 3 );
907 if( (bmx == omx) & (bmy == omy) )
911 /* Special simplified case for subme=1 */
912 else if( bmy > h->mb.mv_min_spel[1] && bmy < h->mb.mv_max_spel[1] && bmx > h->mb.mv_min_spel[0] && bmx < h->mb.mv_max_spel[0] )
915 int omx = bmx, omy = bmy;
916 /* We have to use mc_luma because all strides must be the same to use fpelcmp_x4 */
917 h->mc.mc_luma( pix , 64, m->p_fref, m->i_stride[0], omx, omy-1, bw, bh, &m->weight[0] );
918 h->mc.mc_luma( pix+16, 64, m->p_fref, m->i_stride[0], omx, omy+1, bw, bh, &m->weight[0] );
919 h->mc.mc_luma( pix+32, 64, m->p_fref, m->i_stride[0], omx-1, omy, bw, bh, &m->weight[0] );
920 h->mc.mc_luma( pix+48, 64, m->p_fref, m->i_stride[0], omx+1, omy, bw, bh, &m->weight[0] );
921 h->pixf.fpelcmp_x4[i_pixel]( m->p_fenc[0], pix, pix+16, pix+32, pix+48, 64, costs );
922 COPY2_IF_LT( bcost, costs[0] + p_cost_mvx[omx ] + p_cost_mvy[omy-1], bmy, omy-1 );
923 COPY2_IF_LT( bcost, costs[1] + p_cost_mvx[omx ] + p_cost_mvy[omy+1], bmy, omy+1 );
924 COPY3_IF_LT( bcost, costs[2] + p_cost_mvx[omx-1] + p_cost_mvy[omy ], bmx, omx-1, bmy, omy );
925 COPY3_IF_LT( bcost, costs[3] + p_cost_mvx[omx+1] + p_cost_mvy[omy ], bmx, omx+1, bmy, omy );
931 m->cost_mv = p_cost_mvx[bmx] + p_cost_mvy[bmy];
934 #define BIME_CACHE( dx, dy, list )\
936 x264_me_t *m = m##list;\
937 int i = 4 + 3*dx + dy;\
938 int mvx = bm##list##x+dx;\
939 int mvy = bm##list##y+dy;\
940 stride[0][list][i] = bw;\
941 src[0][list][i] = h->mc.get_ref( pixy_buf[list][i], &stride[0][list][i], &m->p_fref[0],\
942 m->i_stride[0], mvx, mvy, bw, bh, x264_weight_none );\
947 stride[1][list][i] = bw;\
948 src[1][list][i] = h->mc.get_ref( pixu_buf[list][i], &stride[1][list][i], &m->p_fref[4],\
949 m->i_stride[1], mvx, mvy, bw, bh, x264_weight_none );\
950 stride[2][list][i] = bw;\
951 src[2][list][i] = h->mc.get_ref( pixv_buf[list][i], &stride[2][list][i], &m->p_fref[8],\
952 m->i_stride[2], mvx, mvy, bw, bh, x264_weight_none );\
955 h->mc.mc_chroma( pixu_buf[list][i], pixv_buf[list][i], 8, m->p_fref[4], m->i_stride[1],\
956 mvx, 2*(mvy+mv##list##y_offset)>>chroma_v_shift, bw>>1, bh>>chroma_v_shift );\
960 #define SATD_THRESH(cost) (cost+(cost>>4))
962 /* Don't unroll the BIME_CACHE loop. I couldn't find any way to force this
963 * other than making its iteration count not a compile-time constant. */
964 int x264_iter_kludge = 0;
966 static void ALWAYS_INLINE x264_me_refine_bidir( x264_t *h, x264_me_t *m0, x264_me_t *m1, int i_weight, int i8, int i_lambda2, int rd )
970 int s8 = X264_SCAN8_0 + 2*x + 16*y;
971 int16_t *cache0_mv = h->mb.cache.mv[0][s8];
972 int16_t *cache1_mv = h->mb.cache.mv[1][s8];
973 const int i_pixel = m0->i_pixel;
974 const int bw = x264_pixel_size[i_pixel].w;
975 const int bh = x264_pixel_size[i_pixel].h;
976 ALIGNED_ARRAY_16( pixel, pixy_buf,[2],[9][16*16] );
977 ALIGNED_ARRAY_16( pixel, pixu_buf,[2],[9][16*16] );
978 ALIGNED_ARRAY_16( pixel, pixv_buf,[2],[9][16*16] );
980 int chromapix = h->luma2chroma_pixel[i_pixel];
981 int chroma_v_shift = CHROMA_V_SHIFT;
982 int chroma_x = (8 >> CHROMA_H_SHIFT) * x;
983 int chroma_y = (8 >> chroma_v_shift) * y;
984 pixel *pix = &h->mb.pic.p_fdec[0][8*x + 8*y*FDEC_STRIDE];
985 pixel *pixu = &h->mb.pic.p_fdec[1][chroma_x + chroma_y*FDEC_STRIDE];
986 pixel *pixv = &h->mb.pic.p_fdec[2][chroma_x + chroma_y*FDEC_STRIDE];
987 int ref0 = h->mb.cache.ref[0][s8];
988 int ref1 = h->mb.cache.ref[1][s8];
989 const int mv0y_offset = chroma_v_shift & MB_INTERLACED & ref0 ? (h->mb.i_mb_y & 1)*4 - 2 : 0;
990 const int mv1y_offset = chroma_v_shift & MB_INTERLACED & ref1 ? (h->mb.i_mb_y & 1)*4 - 2 : 0;
992 int bm0x = m0->mv[0];
993 int bm0y = m0->mv[1];
994 int bm1x = m1->mv[0];
995 int bm1y = m1->mv[1];
996 int bcost = COST_MAX;
997 int mc_list0 = 1, mc_list1 = 1;
998 uint64_t bcostrd = COST_MAX64;
1000 /* each byte of visited represents 8 possible m1y positions, so a 4D array isn't needed */
1001 ALIGNED_ARRAY_16( uint8_t, visited,[8],[8][8] );
1002 /* all permutations of an offset in up to 2 of the dimensions */
1003 ALIGNED_4( static const int8_t dia4d[33][4] ) =
1006 {0,0,0,1}, {0,0,0,-1}, {0,0,1,0}, {0,0,-1,0},
1007 {0,1,0,0}, {0,-1,0,0}, {1,0,0,0}, {-1,0,0,0},
1008 {0,0,1,1}, {0,0,-1,-1},{0,1,1,0}, {0,-1,-1,0},
1009 {1,1,0,0}, {-1,-1,0,0},{1,0,0,1}, {-1,0,0,-1},
1010 {0,1,0,1}, {0,-1,0,-1},{1,0,1,0}, {-1,0,-1,0},
1011 {0,0,-1,1},{0,0,1,-1}, {0,-1,1,0},{0,1,-1,0},
1012 {-1,1,0,0},{1,-1,0,0}, {1,0,0,-1},{-1,0,0,1},
1013 {0,-1,0,1},{0,1,0,-1}, {-1,0,1,0},{1,0,-1,0},
1016 if( bm0y < h->mb.mv_min_spel[1] + 8 || bm1y < h->mb.mv_min_spel[1] + 8 ||
1017 bm0y > h->mb.mv_max_spel[1] - 8 || bm1y > h->mb.mv_max_spel[1] - 8 ||
1018 bm0x < h->mb.mv_min_spel[0] + 8 || bm1x < h->mb.mv_min_spel[0] + 8 ||
1019 bm0x > h->mb.mv_max_spel[0] - 8 || bm1x > h->mb.mv_max_spel[0] - 8 )
1022 if( rd && m0->i_pixel != PIXEL_16x16 && i8 != 0 )
1024 x264_mb_predict_mv( h, 0, i8<<2, bw>>2, m0->mvp );
1025 x264_mb_predict_mv( h, 1, i8<<2, bw>>2, m1->mvp );
1028 const uint16_t *p_cost_m0x = m0->p_cost_mv - m0->mvp[0];
1029 const uint16_t *p_cost_m0y = m0->p_cost_mv - m0->mvp[1];
1030 const uint16_t *p_cost_m1x = m1->p_cost_mv - m1->mvp[0];
1031 const uint16_t *p_cost_m1y = m1->p_cost_mv - m1->mvp[1];
1033 h->mc.memzero_aligned( visited, sizeof(uint8_t[8][8][8]) );
1035 for( int pass = 0; pass < 8; pass++ )
1038 /* check all mv pairs that differ in at most 2 components from the current mvs. */
1039 /* doesn't do chroma ME. this probably doesn't matter, as the gains
1040 * from bidir ME are the same with and without chroma ME. */
1043 for( int j = x264_iter_kludge; j < 9; j++ )
1044 BIME_CACHE( square1[j][0], square1[j][1], 0 );
1047 for( int j = x264_iter_kludge; j < 9; j++ )
1048 BIME_CACHE( square1[j][0], square1[j][1], 1 );
1050 for( int j = !!pass; j < 33; j++ )
1052 int m0x = dia4d[j][0] + bm0x;
1053 int m0y = dia4d[j][1] + bm0y;
1054 int m1x = dia4d[j][2] + bm1x;
1055 int m1y = dia4d[j][3] + bm1y;
1056 if( !pass || !((visited[(m0x)&7][(m0y)&7][(m1x)&7] & (1<<((m1y)&7)))) )
1058 int i0 = 4 + 3*dia4d[j][0] + dia4d[j][1];
1059 int i1 = 4 + 3*dia4d[j][2] + dia4d[j][3];
1060 visited[(m0x)&7][(m0y)&7][(m1x)&7] |= (1<<((m1y)&7));
1061 h->mc.avg[i_pixel]( pix, FDEC_STRIDE, src[0][0][i0], stride[0][0][i0], src[0][1][i1], stride[0][1][i1], i_weight );
1062 int cost = h->pixf.mbcmp[i_pixel]( m0->p_fenc[0], FENC_STRIDE, pix, FDEC_STRIDE )
1063 + p_cost_m0x[m0x] + p_cost_m0y[m0y] + p_cost_m1x[m1x] + p_cost_m1y[m1y];
1066 if( cost < SATD_THRESH(bcost) )
1068 bcost = X264_MIN( cost, bcost );
1069 M32( cache0_mv ) = pack16to32_mask(m0x,m0y);
1070 M32( cache1_mv ) = pack16to32_mask(m1x,m1y);
1073 h->mc.avg[i_pixel]( pixu, FDEC_STRIDE, src[1][0][i0], stride[1][0][i0], src[1][1][i1], stride[1][1][i1], i_weight );
1074 h->mc.avg[i_pixel]( pixv, FDEC_STRIDE, src[2][0][i0], stride[2][0][i0], src[2][1][i1], stride[2][1][i1], i_weight );
1078 h->mc.avg[chromapix]( pixu, FDEC_STRIDE, pixu_buf[0][i0], 8, pixu_buf[1][i1], 8, i_weight );
1079 h->mc.avg[chromapix]( pixv, FDEC_STRIDE, pixv_buf[0][i0], 8, pixv_buf[1][i1], 8, i_weight );
1081 uint64_t costrd = x264_rd_cost_part( h, i_lambda2, i8*4, m0->i_pixel );
1082 COPY2_IF_LT( bcostrd, costrd, bestj, j );
1086 COPY2_IF_LT( bcost, cost, bestj, j );
1093 bm0x += dia4d[bestj][0];
1094 bm0y += dia4d[bestj][1];
1095 bm1x += dia4d[bestj][2];
1096 bm1y += dia4d[bestj][3];
1098 mc_list0 = M16( &dia4d[bestj][0] );
1099 mc_list1 = M16( &dia4d[bestj][2] );
1104 x264_macroblock_cache_mv ( h, 2*x, 2*y, bw>>2, bh>>2, 0, pack16to32_mask(bm0x, bm0y) );
1105 amvd = pack8to16( X264_MIN(abs(bm0x - m0->mvp[0]),33), X264_MIN(abs(bm0y - m0->mvp[1]),33) );
1106 x264_macroblock_cache_mvd( h, 2*x, 2*y, bw>>2, bh>>2, 0, amvd );
1108 x264_macroblock_cache_mv ( h, 2*x, 2*y, bw>>2, bh>>2, 1, pack16to32_mask(bm1x, bm1y) );
1109 amvd = pack8to16( X264_MIN(abs(bm1x - m1->mvp[0]),33), X264_MIN(abs(bm1y - m1->mvp[1]),33) );
1110 x264_macroblock_cache_mvd( h, 2*x, 2*y, bw>>2, bh>>2, 1, amvd );
1119 void x264_me_refine_bidir_satd( x264_t *h, x264_me_t *m0, x264_me_t *m1, int i_weight )
1121 x264_me_refine_bidir( h, m0, m1, i_weight, 0, 0, 0 );
1124 void x264_me_refine_bidir_rd( x264_t *h, x264_me_t *m0, x264_me_t *m1, int i_weight, int i8, int i_lambda2 )
1126 /* Motion compensation is done as part of bidir_rd; don't repeat
1127 * it in encoding. */
1128 h->mb.b_skip_mc = 1;
1129 x264_me_refine_bidir( h, m0, m1, i_weight, i8, i_lambda2, 1 );
1130 h->mb.b_skip_mc = 0;
1134 #define COST_MV_SATD( mx, my, dst, avoid_mvp ) \
1136 if( !avoid_mvp || !(mx == pmx && my == pmy) ) \
1138 h->mc.mc_luma( pix, FDEC_STRIDE, m->p_fref, m->i_stride[0], mx, my, bw, bh, &m->weight[0] ); \
1139 dst = h->pixf.mbcmp[i_pixel]( m->p_fenc[0], FENC_STRIDE, pix, FDEC_STRIDE ) \
1140 + p_cost_mvx[mx] + p_cost_mvy[my]; \
1141 COPY1_IF_LT( bsatd, dst ); \
1147 #define COST_MV_RD( mx, my, satd, do_dir, mdir ) \
1149 if( satd <= SATD_THRESH(bsatd) ) \
1152 M32( cache_mv ) = pack16to32_mask(mx,my); \
1155 h->mc.mc_luma( pixu, FDEC_STRIDE, &m->p_fref[4], m->i_stride[1], mx, my, bw, bh, &m->weight[1] ); \
1156 h->mc.mc_luma( pixv, FDEC_STRIDE, &m->p_fref[8], m->i_stride[2], mx, my, bw, bh, &m->weight[2] ); \
1158 else if( m->i_pixel <= PIXEL_8x8 ) \
1160 h->mc.mc_chroma( pixu, pixv, FDEC_STRIDE, m->p_fref[4], m->i_stride[1], \
1161 mx, 2*(my+mvy_offset)>>chroma_v_shift, bw>>1, bh>>chroma_v_shift ); \
1162 if( m->weight[1].weightfn ) \
1163 m->weight[1].weightfn[bw>>3]( pixu, FDEC_STRIDE, pixu, FDEC_STRIDE, &m->weight[1], bh>>chroma_v_shift ); \
1164 if( m->weight[2].weightfn ) \
1165 m->weight[2].weightfn[bw>>3]( pixv, FDEC_STRIDE, pixv, FDEC_STRIDE, &m->weight[2], bh>>chroma_v_shift ); \
1167 cost = x264_rd_cost_part( h, i_lambda2, i4, m->i_pixel ); \
1168 COPY4_IF_LT( bcost, cost, bmx, mx, bmy, my, dir, do_dir?mdir:dir ); \
1172 void x264_me_refine_qpel_rd( x264_t *h, x264_me_t *m, int i_lambda2, int i4, int i_list )
1174 int16_t *cache_mv = h->mb.cache.mv[i_list][x264_scan8[i4]];
1175 const uint16_t *p_cost_mvx, *p_cost_mvy;
1176 const int bw = x264_pixel_size[m->i_pixel].w;
1177 const int bh = x264_pixel_size[m->i_pixel].h;
1178 const int i_pixel = m->i_pixel;
1179 int chroma_v_shift = CHROMA_V_SHIFT;
1180 int mvy_offset = chroma_v_shift & MB_INTERLACED & m->i_ref ? (h->mb.i_mb_y & 1)*4 - 2 : 0;
1182 uint64_t bcost = COST_MAX64;
1185 int omx, omy, pmx, pmy;
1191 pixel *pix = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[i4]];
1195 pixu = &h->mb.pic.p_fdec[1][block_idx_xy_fdec[i4]];
1196 pixv = &h->mb.pic.p_fdec[2][block_idx_xy_fdec[i4]];
1200 pixu = &h->mb.pic.p_fdec[1][(i8>>1)*(8*FDEC_STRIDE>>chroma_v_shift)+(i8&1)*4];
1201 pixv = &h->mb.pic.p_fdec[2][(i8>>1)*(8*FDEC_STRIDE>>chroma_v_shift)+(i8&1)*4];
1204 h->mb.b_skip_mc = 1;
1206 if( m->i_pixel != PIXEL_16x16 && i4 != 0 )
1207 x264_mb_predict_mv( h, i_list, i4, bw>>2, m->mvp );
1210 p_cost_mvx = m->p_cost_mv - pmx;
1211 p_cost_mvy = m->p_cost_mv - pmy;
1212 COST_MV_SATD( bmx, bmy, bsatd, 0 );
1213 if( m->i_pixel != PIXEL_16x16 )
1214 COST_MV_RD( bmx, bmy, 0, 0, 0 )
1218 /* check the predicted mv */
1219 if( (bmx != pmx || bmy != pmy)
1220 && pmx >= h->mb.mv_min_spel[0] && pmx <= h->mb.mv_max_spel[0]
1221 && pmy >= h->mb.mv_min_spel[1] && pmy <= h->mb.mv_max_spel[1] )
1223 COST_MV_SATD( pmx, pmy, satd, 0 );
1224 COST_MV_RD ( pmx, pmy, satd, 0, 0 );
1225 /* The hex motion search is guaranteed to not repeat the center candidate,
1226 * so if pmv is chosen, set the "MV to avoid checking" to bmv instead. */
1227 if( bmx == pmx && bmy == pmy )
1234 if( bmy < h->mb.mv_min_spel[1] + 3 || bmy > h->mb.mv_max_spel[1] - 3 ||
1235 bmx < h->mb.mv_min_spel[0] + 3 || bmx > h->mb.mv_max_spel[0] - 3 )
1237 h->mb.b_skip_mc = 0;
1241 /* subpel hex search, same pattern as ME HEX. */
1245 for( int j = 0; j < 6; j++ )
1247 COST_MV_SATD( omx + hex2[j+1][0], omy + hex2[j+1][1], satd, 1 );
1248 COST_MV_RD ( omx + hex2[j+1][0], omy + hex2[j+1][1], satd, 1, j );
1253 /* half hexagon, not overlapping the previous iteration */
1254 for( int i = 1; i < 10; i++ )
1256 const int odir = mod6m1[dir+1];
1257 if( bmy < h->mb.mv_min_spel[1] + 3 ||
1258 bmy > h->mb.mv_max_spel[1] - 3 )
1263 for( int j = 0; j < 3; j++ )
1265 COST_MV_SATD( omx + hex2[odir+j][0], omy + hex2[odir+j][1], satd, 1 );
1266 COST_MV_RD ( omx + hex2[odir+j][0], omy + hex2[odir+j][1], satd, 1, odir-1+j );
1273 /* square refine, same pattern as ME HEX. */
1276 for( int i = 0; i < 8; i++ )
1278 COST_MV_SATD( omx + square1[i+1][0], omy + square1[i+1][1], satd, 1 );
1279 COST_MV_RD ( omx + square1[i+1][0], omy + square1[i+1][1], satd, 0, 0 );
1285 x264_macroblock_cache_mv ( h, block_idx_x[i4], block_idx_y[i4], bw>>2, bh>>2, i_list, pack16to32_mask(bmx, bmy) );
1286 amvd = pack8to16( X264_MIN(abs(bmx - m->mvp[0]),66), X264_MIN(abs(bmy - m->mvp[1]),66) );
1287 x264_macroblock_cache_mvd( h, block_idx_x[i4], block_idx_y[i4], bw>>2, bh>>2, i_list, amvd );
1288 h->mb.b_skip_mc = 0;