1 /*****************************************************************************
2 * me.c: h264 encoder library (Motion Estimation)
3 *****************************************************************************
4 * Copyright (C) 2003-2008 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.
23 *****************************************************************************/
25 #include "common/common.h"
26 #include "macroblock.h"
29 /* presets selected from good points on the speed-vs-quality curve of several test videos
30 * subpel_iters[i_subpel_refine] = { refine_hpel, refine_qpel, me_hpel, me_qpel }
31 * where me_* are the number of EPZS iterations run on all candidate block types,
32 * and refine_* are run only on the winner.
33 * the subme=8,9 values are much higher because any amount of satd search makes
34 * up its time by reducing the number of qpel-rd iterations. */
35 static const int subpel_iterations[][4] =
49 static const int mod6m1[8] = {5,0,1,2,3,4,5,0};
50 /* radius 2 hexagon. repeated entries are to avoid having to compute mod6 every time. */
51 static const int hex2[8][2] = {{-1,-2}, {-2,0}, {-1,2}, {1,2}, {2,0}, {1,-2}, {-1,-2}, {-2,0}};
52 static const int square1[9][2] = {{0,0}, {0,-1}, {0,1}, {-1,0}, {1,0}, {-1,-1}, {-1,1}, {1,-1}, {1,1}};
54 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 );
56 #define BITS_MVD( mx, my )\
57 (p_cost_mvx[(mx)<<2] + p_cost_mvy[(my)<<2])
59 #define COST_MV( mx, my )\
61 int cost = h->pixf.fpelcmp[i_pixel]( p_fenc, FENC_STRIDE,\
62 &p_fref_w[(my)*stride+(mx)], stride )\
64 COPY3_IF_LT( bcost, cost, bmx, mx, bmy, my );\
67 #define COST_MV_HPEL( mx, my ) \
70 uint8_t *src = h->mc.get_ref( pix, &stride2, m->p_fref, stride, mx, my, bw, bh, &m->weight[0] ); \
71 int cost = h->pixf.fpelcmp[i_pixel]( p_fenc, FENC_STRIDE, src, stride2 ) \
72 + p_cost_mvx[ mx ] + p_cost_mvy[ my ]; \
73 COPY3_IF_LT( bpred_cost, cost, bpred_mx, mx, bpred_my, my ); \
76 #define COST_MV_X3_DIR( m0x, m0y, m1x, m1y, m2x, m2y, costs )\
78 uint8_t *pix_base = p_fref_w + bmx + bmy*stride;\
79 h->pixf.fpelcmp_x3[i_pixel]( p_fenc,\
80 pix_base + (m0x) + (m0y)*stride,\
81 pix_base + (m1x) + (m1y)*stride,\
82 pix_base + (m2x) + (m2y)*stride,\
84 (costs)[0] += BITS_MVD( bmx+(m0x), bmy+(m0y) );\
85 (costs)[1] += BITS_MVD( bmx+(m1x), bmy+(m1y) );\
86 (costs)[2] += BITS_MVD( bmx+(m2x), bmy+(m2y) );\
89 #define COST_MV_X4_DIR( m0x, m0y, m1x, m1y, m2x, m2y, m3x, m3y, costs )\
91 uint8_t *pix_base = p_fref_w + bmx + bmy*stride;\
92 h->pixf.fpelcmp_x4[i_pixel]( p_fenc,\
93 pix_base + (m0x) + (m0y)*stride,\
94 pix_base + (m1x) + (m1y)*stride,\
95 pix_base + (m2x) + (m2y)*stride,\
96 pix_base + (m3x) + (m3y)*stride,\
98 (costs)[0] += BITS_MVD( bmx+(m0x), bmy+(m0y) );\
99 (costs)[1] += BITS_MVD( bmx+(m1x), bmy+(m1y) );\
100 (costs)[2] += BITS_MVD( bmx+(m2x), bmy+(m2y) );\
101 (costs)[3] += BITS_MVD( bmx+(m3x), bmy+(m3y) );\
104 #define COST_MV_X4( m0x, m0y, m1x, m1y, m2x, m2y, m3x, m3y )\
106 uint8_t *pix_base = p_fref_w + omx + omy*stride;\
107 h->pixf.fpelcmp_x4[i_pixel]( p_fenc,\
108 pix_base + (m0x) + (m0y)*stride,\
109 pix_base + (m1x) + (m1y)*stride,\
110 pix_base + (m2x) + (m2y)*stride,\
111 pix_base + (m3x) + (m3y)*stride,\
113 costs[0] += BITS_MVD( omx+(m0x), omy+(m0y) );\
114 costs[1] += BITS_MVD( omx+(m1x), omy+(m1y) );\
115 costs[2] += BITS_MVD( omx+(m2x), omy+(m2y) );\
116 costs[3] += BITS_MVD( omx+(m3x), omy+(m3y) );\
117 COPY3_IF_LT( bcost, costs[0], bmx, omx+(m0x), bmy, omy+(m0y) );\
118 COPY3_IF_LT( bcost, costs[1], bmx, omx+(m1x), bmy, omy+(m1y) );\
119 COPY3_IF_LT( bcost, costs[2], bmx, omx+(m2x), bmy, omy+(m2y) );\
120 COPY3_IF_LT( bcost, costs[3], bmx, omx+(m3x), bmy, omy+(m3y) );\
123 #define COST_MV_X3_ABS( m0x, m0y, m1x, m1y, m2x, m2y )\
125 h->pixf.fpelcmp_x3[i_pixel]( p_fenc,\
126 p_fref_w + (m0x) + (m0y)*stride,\
127 p_fref_w + (m1x) + (m1y)*stride,\
128 p_fref_w + (m2x) + (m2y)*stride,\
130 costs[0] += p_cost_mvx[(m0x)<<2]; /* no cost_mvy */\
131 costs[1] += p_cost_mvx[(m1x)<<2];\
132 costs[2] += p_cost_mvx[(m2x)<<2];\
133 COPY3_IF_LT( bcost, costs[0], bmx, m0x, bmy, m0y );\
134 COPY3_IF_LT( bcost, costs[1], bmx, m1x, bmy, m1y );\
135 COPY3_IF_LT( bcost, costs[2], bmx, m2x, bmy, m2y );\
141 #define DIA1_ITER( mx, my )\
144 COST_MV_X4( 0,-1, 0,1, -1,0, 1,0 );\
147 #define CROSS( start, x_max, y_max )\
150 if( x_max <= X264_MIN(mv_x_max-omx, omx-mv_x_min) )\
151 for( ; i < x_max-2; i+=4 )\
152 COST_MV_X4( i,0, -i,0, i+2,0, -i-2,0 );\
153 for( ; i < x_max; i+=2 )\
155 if( omx+i <= mv_x_max )\
156 COST_MV( omx+i, omy );\
157 if( omx-i >= mv_x_min )\
158 COST_MV( omx-i, omy );\
161 if( y_max <= X264_MIN(mv_y_max-omy, omy-mv_y_min) )\
162 for( ; i < y_max-2; i+=4 )\
163 COST_MV_X4( 0,i, 0,-i, 0,i+2, 0,-i-2 );\
164 for( ; i < y_max; i+=2 )\
166 if( omy+i <= mv_y_max )\
167 COST_MV( omx, omy+i );\
168 if( omy-i >= mv_y_min )\
169 COST_MV( omx, omy-i );\
173 void x264_me_search_ref( x264_t *h, x264_me_t *m, int16_t (*mvc)[2], int i_mvc, int *p_halfpel_thresh )
175 const int bw = x264_pixel_size[m->i_pixel].w;
176 const int bh = x264_pixel_size[m->i_pixel].h;
177 const int i_pixel = m->i_pixel;
178 const int stride = m->i_stride[0];
179 int i_me_range = h->param.analyse.i_me_range;
181 int bpred_mx = 0, bpred_my = 0, bpred_cost = COST_MAX;
182 int omx, omy, pmx, pmy;
183 uint8_t *p_fenc = m->p_fenc[0];
184 uint8_t *p_fref_w = m->p_fref_w;
185 ALIGNED_ARRAY_16( uint8_t, pix,[16*16] );
191 int mv_x_min = h->mb.mv_min_fpel[0];
192 int mv_y_min = h->mb.mv_min_fpel[1];
193 int mv_x_max = h->mb.mv_max_fpel[0];
194 int mv_y_max = h->mb.mv_max_fpel[1];
196 #define CHECK_MVRANGE(mx,my) ( mx >= mv_x_min && mx <= mv_x_max && my >= mv_y_min && my <= mv_y_max )
198 const uint16_t *p_cost_mvx = m->p_cost_mv - m->mvp[0];
199 const uint16_t *p_cost_mvy = m->p_cost_mv - m->mvp[1];
201 bmx = x264_clip3( m->mvp[0], mv_x_min*4, mv_x_max*4 );
202 bmy = x264_clip3( m->mvp[1], mv_y_min*4, mv_y_max*4 );
203 pmx = ( bmx + 2 ) >> 2;
204 pmy = ( bmy + 2 ) >> 2;
207 /* try extra predictors if provided */
208 if( h->mb.i_subpel_refine >= 3 )
210 uint32_t bmv = pack16to32_mask(bmx,bmy);
211 COST_MV_HPEL( bmx, bmy );
212 for( i = 0; i < i_mvc; i++ )
214 if( M32( mvc[i] ) && (bmv - M32( mvc[i] )) )
216 int mx = x264_clip3( mvc[i][0], mv_x_min*4, mv_x_max*4 );
217 int my = x264_clip3( mvc[i][1], mv_y_min*4, mv_y_max*4 );
218 COST_MV_HPEL( mx, my );
221 bmx = ( bpred_mx + 2 ) >> 2;
222 bmy = ( bpred_my + 2 ) >> 2;
229 /* Because we are rounding the predicted motion vector to fullpel, there will be
230 * an extra MV cost in 15 out of 16 cases. However, when the predicted MV is
231 * chosen as the best predictor, it is often the case that the subpel search will
232 * result in a vector at or next to the predicted motion vector. Therefore, it is
233 * sensible to remove the cost of the MV from the rounded MVP to avoid unfairly
234 * biasing against use of the predicted motion vector. */
235 bcost -= BITS_MVD( pmx, pmy );
236 for( i = 0; i < i_mvc; i++ )
238 int mx = (mvc[i][0] + 2) >> 2;
239 int my = (mvc[i][1] + 2) >> 2;
240 if( (mx | my) && ((mx-bmx) | (my-bmy)) )
242 mx = x264_clip3( mx, mv_x_min, mv_x_max );
243 my = x264_clip3( my, mv_y_min, mv_y_max );
250 switch( h->mb.i_me_method )
253 /* diamond search, radius 1 */
258 COST_MV_X4_DIR( 0,-1, 0,1, -1,0, 1,0, costs );
259 COPY1_IF_LT( bcost, (costs[0]<<4)+1 );
260 COPY1_IF_LT( bcost, (costs[1]<<4)+3 );
261 COPY1_IF_LT( bcost, (costs[2]<<4)+4 );
262 COPY1_IF_LT( bcost, (costs[3]<<4)+12 );
265 bmx -= (bcost<<28)>>30;
266 bmy -= (bcost<<30)>>30;
268 if( !CHECK_MVRANGE(bmx, bmy) )
270 } while( ++i < i_me_range );
276 /* hexagon search, radius 2 */
278 for( i = 0; i < i_me_range/2; i++ )
280 omx = bmx; omy = bmy;
281 COST_MV( omx-2, omy );
282 COST_MV( omx-1, omy+2 );
283 COST_MV( omx+1, omy+2 );
284 COST_MV( omx+2, omy );
285 COST_MV( omx+1, omy-2 );
286 COST_MV( omx-1, omy-2 );
287 if( bmx == omx && bmy == omy )
289 if( !CHECK_MVRANGE(bmx, bmy) )
293 /* equivalent to the above, but eliminates duplicate candidates */
296 COST_MV_X3_DIR( -2,0, -1, 2, 1, 2, costs );
297 COST_MV_X3_DIR( 2,0, 1,-2, -1,-2, costs+3 );
299 COPY1_IF_LT( bcost, (costs[0]<<3)+2 );
300 COPY1_IF_LT( bcost, (costs[1]<<3)+3 );
301 COPY1_IF_LT( bcost, (costs[2]<<3)+4 );
302 COPY1_IF_LT( bcost, (costs[3]<<3)+5 );
303 COPY1_IF_LT( bcost, (costs[4]<<3)+6 );
304 COPY1_IF_LT( bcost, (costs[5]<<3)+7 );
309 bmx += hex2[dir+1][0];
310 bmy += hex2[dir+1][1];
311 /* half hexagon, not overlapping the previous iteration */
312 for( i = 1; i < i_me_range/2 && CHECK_MVRANGE(bmx, bmy); i++ )
314 COST_MV_X3_DIR( hex2[dir+0][0], hex2[dir+0][1],
315 hex2[dir+1][0], hex2[dir+1][1],
316 hex2[dir+2][0], hex2[dir+2][1],
319 COPY1_IF_LT( bcost, (costs[0]<<3)+1 );
320 COPY1_IF_LT( bcost, (costs[1]<<3)+2 );
321 COPY1_IF_LT( bcost, (costs[2]<<3)+3 );
326 bmx += hex2[dir+1][0];
327 bmy += hex2[dir+1][1];
334 COST_MV_X4_DIR( 0,-1, 0,1, -1,0, 1,0, costs );
335 COPY2_IF_LT( bcost, costs[0], dir, 1 );
336 COPY2_IF_LT( bcost, costs[1], dir, 2 );
337 COPY2_IF_LT( bcost, costs[2], dir, 3 );
338 COPY2_IF_LT( bcost, costs[3], dir, 4 );
339 COST_MV_X4_DIR( -1,-1, -1,1, 1,-1, 1,1, costs );
340 COPY2_IF_LT( bcost, costs[0], dir, 5 );
341 COPY2_IF_LT( bcost, costs[1], dir, 6 );
342 COPY2_IF_LT( bcost, costs[2], dir, 7 );
343 COPY2_IF_LT( bcost, costs[3], dir, 8 );
344 bmx += square1[dir][0];
345 bmy += square1[dir][1];
350 /* Uneven-cross Multi-Hexagon-grid Search
351 * as in JM, except with different early termination */
353 static const int x264_pixel_size_shift[7] = { 0, 1, 1, 2, 3, 3, 4 };
358 /* refine predictors */
360 DIA1_ITER( pmx, pmy );
364 if(i_pixel == PIXEL_4x4)
368 if( (bmx | bmy) && ((bmx-pmx) | (bmy-pmy)) )
369 DIA1_ITER( bmx, bmy );
370 if( bcost == ucost2 )
372 omx = bmx; omy = bmy;
374 /* early termination */
375 #define SAD_THRESH(v) ( bcost < ( v >> x264_pixel_size_shift[i_pixel] ) )
376 if( bcost == ucost2 && SAD_THRESH(2000) )
378 COST_MV_X4( 0,-2, -1,-1, 1,-1, -2,0 );
379 COST_MV_X4( 2, 0, -1, 1, 1, 1, 0,2 );
380 if( bcost == ucost1 && SAD_THRESH(500) )
382 if( bcost == ucost2 )
384 int range = (i_me_range>>1) | 1;
385 CROSS( 3, range, range );
386 COST_MV_X4( -1,-2, 1,-2, -2,-1, 2,-1 );
387 COST_MV_X4( -2, 1, 2, 1, -1, 2, 1, 2 );
388 if( bcost == ucost2 )
390 cross_start = range + 2;
394 /* adaptive search range */
397 /* range multipliers based on casual inspection of some statistics of
398 * average distance between current predictor and final mv found by ESA.
399 * these have not been tuned much by actual encoding. */
400 static const int range_mul[4][4] =
408 int sad_ctx, mvd_ctx;
413 if( i_pixel == PIXEL_16x16 )
414 /* mvc is probably the same as mvp, so the difference isn't meaningful.
415 * but prediction usually isn't too bad, so just use medium range */
418 mvd = abs( m->mvp[0] - mvc[0][0] )
419 + abs( m->mvp[1] - mvc[0][1] );
423 /* calculate the degree of agreement between predictors. */
424 /* in 16x16, mvc includes all the neighbors used to make mvp,
425 * so don't count mvp separately. */
428 if( i_pixel != PIXEL_16x16 )
430 mvd = abs( m->mvp[0] - mvc[0][0] )
431 + abs( m->mvp[1] - mvc[0][1] );
434 mvd += x264_predictor_difference( mvc, i_mvc );
437 sad_ctx = SAD_THRESH(1000) ? 0
438 : SAD_THRESH(2000) ? 1
439 : SAD_THRESH(4000) ? 2 : 3;
440 mvd_ctx = mvd < 10*denom ? 0
442 : mvd < 40*denom ? 2 : 3;
444 i_me_range = i_me_range * range_mul[mvd_ctx][sad_ctx] / 4;
447 /* FIXME if the above DIA2/OCT2/CROSS found a new mv, it has not updated omx/omy.
448 * we are still centered on the same place as the DIA2. is this desirable? */
449 CROSS( cross_start, i_me_range, i_me_range/2 );
451 COST_MV_X4( -2,-2, -2,2, 2,-2, 2,2 );
454 omx = bmx; omy = bmy;
455 const uint16_t *p_cost_omvx = p_cost_mvx + omx*4;
456 const uint16_t *p_cost_omvy = p_cost_mvy + omy*4;
460 static const int hex4[16][2] = {
461 { 0,-4}, { 0, 4}, {-2,-3}, { 2,-3},
462 {-4,-2}, { 4,-2}, {-4,-1}, { 4,-1},
463 {-4, 0}, { 4, 0}, {-4, 1}, { 4, 1},
464 {-4, 2}, { 4, 2}, {-2, 3}, { 2, 3},
467 if( 4*i > X264_MIN4( mv_x_max-omx, omx-mv_x_min,
468 mv_y_max-omy, omy-mv_y_min ) )
470 for( j = 0; j < 16; j++ )
472 int mx = omx + hex4[j][0]*i;
473 int my = omy + hex4[j][1]*i;
474 if( CHECK_MVRANGE(mx, my) )
481 uint8_t *pix_base = p_fref_w + omx + (omy-4*i)*stride;
483 #define SADS(k,x0,y0,x1,y1,x2,y2,x3,y3)\
484 h->pixf.fpelcmp_x4[i_pixel]( p_fenc,\
485 pix_base x0*i+(y0-2*k+4)*dy,\
486 pix_base x1*i+(y1-2*k+4)*dy,\
487 pix_base x2*i+(y2-2*k+4)*dy,\
488 pix_base x3*i+(y3-2*k+4)*dy,\
489 stride, costs+4*k );\
491 #define ADD_MVCOST(k,x,y) costs[k] += p_cost_omvx[x*4*i] + p_cost_omvy[y*4*i]
492 #define MIN_MV(k,x,y) COPY2_IF_LT( bcost, costs[k], dir, x*16+(y&15) )
493 SADS( 0, +0,-4, +0,+4, -2,-3, +2,-3 );
494 SADS( 1, -4,-2, +4,-2, -4,-1, +4,-1 );
495 SADS( 2, -4,+0, +4,+0, -4,+1, +4,+1 );
496 SADS( 3, -4,+2, +4,+2, -2,+3, +2,+3 );
497 ADD_MVCOST( 0, 0,-4 );
498 ADD_MVCOST( 1, 0, 4 );
499 ADD_MVCOST( 2,-2,-3 );
500 ADD_MVCOST( 3, 2,-3 );
501 ADD_MVCOST( 4,-4,-2 );
502 ADD_MVCOST( 5, 4,-2 );
503 ADD_MVCOST( 6,-4,-1 );
504 ADD_MVCOST( 7, 4,-1 );
505 ADD_MVCOST( 8,-4, 0 );
506 ADD_MVCOST( 9, 4, 0 );
507 ADD_MVCOST( 10,-4, 1 );
508 ADD_MVCOST( 11, 4, 1 );
509 ADD_MVCOST( 12,-4, 2 );
510 ADD_MVCOST( 13, 4, 2 );
511 ADD_MVCOST( 14,-2, 3 );
512 ADD_MVCOST( 15, 2, 3 );
534 bmx = omx + i*(dir>>4);
535 bmy = omy + i*((dir<<28)>>28);
538 } while( ++i <= i_me_range/4 );
539 if( bmy <= mv_y_max && bmy >= mv_y_min && bmx <= mv_x_max && bmx >= mv_x_min )
547 const int min_x = X264_MAX( bmx - i_me_range, mv_x_min );
548 const int min_y = X264_MAX( bmy - i_me_range, mv_y_min );
549 const int max_x = X264_MIN( bmx + i_me_range, mv_x_max );
550 const int max_y = X264_MIN( bmy + i_me_range, mv_y_max );
551 /* SEA is fastest in multiples of 4 */
552 const int width = (max_x - min_x + 3) & ~3;
555 /* plain old exhaustive search */
557 for( my = min_y; my <= max_y; my++ )
558 for( mx = min_x; mx <= max_x; mx++ )
561 /* successive elimination by comparing DC before a full SAD,
562 * because sum(abs(diff)) >= abs(diff(sum)). */
563 uint16_t *sums_base = m->integral;
564 /* due to a GCC bug on some platforms (win32?), zero[] may not actually be aligned.
565 * this is not a problem because it is not used for any SSE instructions. */
566 ALIGNED_16( static uint8_t zero[8*FENC_STRIDE] );
567 ALIGNED_ARRAY_16( int, enc_dc,[4] );
568 int sad_size = i_pixel <= PIXEL_8x8 ? PIXEL_8x8 : PIXEL_4x4;
569 int delta = x264_pixel_size[sad_size].w;
570 int16_t *xs = h->scratch_buffer;
572 uint16_t *cost_fpel_mvx = h->cost_mv_fpel[x264_lambda_tab[h->mb.i_qp]][-m->mvp[0]&3] + (-m->mvp[0]>>2);
574 h->pixf.sad_x4[sad_size]( zero, p_fenc, p_fenc+delta,
575 p_fenc+delta*FENC_STRIDE, p_fenc+delta+delta*FENC_STRIDE,
576 FENC_STRIDE, enc_dc );
578 sums_base += stride * (h->fenc->i_lines[0] + PADV*2);
579 if( i_pixel == PIXEL_16x16 || i_pixel == PIXEL_8x16 || i_pixel == PIXEL_4x8 )
581 if( i_pixel == PIXEL_8x16 || i_pixel == PIXEL_4x8 )
582 enc_dc[1] = enc_dc[2];
584 if( h->mb.i_me_method == X264_ME_TESA )
586 // ADS threshold, then SAD threshold, then keep the best few SADs, then SATD
587 mvsad_t *mvsads = (mvsad_t *)(xs + ((width+15)&~15));
588 int nmvsad = 0, limit;
589 int sad_thresh = i_me_range <= 16 ? 10 : i_me_range <= 24 ? 11 : 12;
590 int bsad = h->pixf.sad[i_pixel]( p_fenc, FENC_STRIDE, p_fref_w+bmy*stride+bmx, stride )
591 + BITS_MVD( bmx, bmy );
592 for( my = min_y; my <= max_y; my++ )
594 int ycost = p_cost_mvy[my<<2];
598 xn = h->pixf.ads[i_pixel]( enc_dc, sums_base + min_x + my * stride, delta,
599 cost_fpel_mvx+min_x, xs, width, bsad*17/16 );
600 for( i=0; i<xn-2; i+=3 )
602 uint8_t *ref = p_fref_w+min_x+my*stride;
604 h->pixf.sad_x3[i_pixel]( p_fenc, ref+xs[i], ref+xs[i+1], ref+xs[i+2], stride, sads );
607 int sad = sads[j] + cost_fpel_mvx[xs[i+j]];
608 if( sad < bsad*sad_thresh>>3 )
610 COPY1_IF_LT( bsad, sad );
611 mvsads[nmvsad].sad = sad + ycost;
612 mvsads[nmvsad].mx = min_x+xs[i+j];
613 mvsads[nmvsad].my = my;
620 int mx = min_x+xs[i];
621 int sad = h->pixf.sad[i_pixel]( p_fenc, FENC_STRIDE, p_fref_w+mx+my*stride, stride )
622 + cost_fpel_mvx[xs[i]];
623 if( sad < bsad*sad_thresh>>3 )
625 COPY1_IF_LT( bsad, sad );
626 mvsads[nmvsad].sad = sad + ycost;
627 mvsads[nmvsad].mx = mx;
628 mvsads[nmvsad].my = my;
635 limit = i_me_range / 2;
636 sad_thresh = bsad*sad_thresh>>3;
637 while( nmvsad > limit*2 && sad_thresh > bsad )
639 // halve the range if the domain is too large... eh, close enough
640 sad_thresh = (sad_thresh + bsad) >> 1;
641 for( i=0; i<nmvsad && mvsads[i].sad <= sad_thresh; i++ );
642 for( j=i; j<nmvsad; j++ )
644 /* mvsad_t is not guaranteed to be 8 bytes on all archs, so check before using explicit write-combining */
645 if( sizeof( mvsad_t ) == sizeof( uint64_t ) )
646 CP64( &mvsads[i], &mvsads[j] );
648 mvsads[i] = mvsads[j];
649 i += mvsads[j].sad <= sad_thresh;
653 while( nmvsad > limit )
655 int bsad = mvsads[0].sad;
657 for( i=1; i<nmvsad; i++ )
658 COPY2_IF_GT( bsad, mvsads[i].sad, bi, i );
660 mvsads[bi] = mvsads[nmvsad];
661 if( sizeof( mvsad_t ) == sizeof( uint64_t ) )
662 CP64( &mvsads[bi], &mvsads[nmvsad] );
664 mvsads[bi] = mvsads[nmvsad];
666 for( i=0; i<nmvsad; i++ )
667 COST_MV( mvsads[i].mx, mvsads[i].my );
672 for( my = min_y; my <= max_y; my++ )
674 int ycost = p_cost_mvy[my<<2];
678 xn = h->pixf.ads[i_pixel]( enc_dc, sums_base + min_x + my * stride, delta,
679 cost_fpel_mvx+min_x, xs, width, bcost );
680 for( i=0; i<xn-2; i+=3 )
681 COST_MV_X3_ABS( min_x+xs[i],my, min_x+xs[i+1],my, min_x+xs[i+2],my );
684 COST_MV( min_x+xs[i], my );
693 if( bpred_cost < bcost )
697 m->cost = bpred_cost;
706 /* compute the real cost */
707 m->cost_mv = p_cost_mvx[ m->mv[0] ] + p_cost_mvy[ m->mv[1] ];
708 if( bmx == pmx && bmy == pmy && h->mb.i_subpel_refine < 3 )
709 m->cost += m->cost_mv;
712 if( h->mb.i_subpel_refine >= 2 )
714 int hpel = subpel_iterations[h->mb.i_subpel_refine][2];
715 int qpel = subpel_iterations[h->mb.i_subpel_refine][3];
716 refine_subpel( h, m, hpel, qpel, p_halfpel_thresh, 0 );
721 void x264_me_refine_qpel( x264_t *h, x264_me_t *m )
723 int hpel = subpel_iterations[h->mb.i_subpel_refine][0];
724 int qpel = subpel_iterations[h->mb.i_subpel_refine][1];
726 if( m->i_pixel <= PIXEL_8x8 && h->sh.i_type == SLICE_TYPE_P )
727 m->cost -= m->i_ref_cost;
729 refine_subpel( h, m, hpel, qpel, NULL, 1 );
732 #define COST_MV_SAD( mx, my ) \
735 uint8_t *src = h->mc.get_ref( pix[0], &stride, m->p_fref, m->i_stride[0], mx, my, bw, bh, &m->weight[0] ); \
736 int cost = h->pixf.fpelcmp[i_pixel]( m->p_fenc[0], FENC_STRIDE, src, stride ) \
737 + p_cost_mvx[ mx ] + p_cost_mvy[ my ]; \
738 COPY3_IF_LT( bcost, cost, bmx, mx, bmy, my ); \
741 #define COST_MV_SATD( mx, my, dir ) \
742 if( b_refine_qpel || (dir^1) != odir ) \
745 uint8_t *src = h->mc.get_ref( pix[0], &stride, m->p_fref, m->i_stride[0], mx, my, bw, bh, &m->weight[0] ); \
746 int cost = h->pixf.mbcmp_unaligned[i_pixel]( m->p_fenc[0], FENC_STRIDE, src, stride ) \
747 + p_cost_mvx[ mx ] + p_cost_mvy[ my ]; \
748 if( b_chroma_me && cost < bcost ) \
750 h->mc.mc_chroma( pix[0], 8, m->p_fref[4], m->i_stride[1], mx, my + mvy_offset, bw/2, bh/2 ); \
751 if( m->weight[1].weightfn ) \
752 m->weight[1].weightfn[x264_pixel_size[i_pixel].w>>3]( pix[0], 8, pix[0], 8, \
753 &m->weight[1], x264_pixel_size[i_pixel].h>>1 ); \
754 cost += h->pixf.mbcmp[i_pixel+3]( m->p_fenc[1], FENC_STRIDE, pix[0], 8 ); \
757 h->mc.mc_chroma( pix[0], 8, m->p_fref[5], m->i_stride[1], mx, my + mvy_offset, bw/2, bh/2 ); \
758 cost += h->pixf.mbcmp[i_pixel+3]( m->p_fenc[2], FENC_STRIDE, pix[0], 8 ); \
759 if( m->weight[2].weightfn ) \
760 m->weight[2].weightfn[x264_pixel_size[i_pixel].w>>3]( pix[0], 8, pix[0], 8, \
761 &m->weight[2], x264_pixel_size[i_pixel].h>>1 ); \
773 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 )
775 const int bw = x264_pixel_size[m->i_pixel].w;
776 const int bh = x264_pixel_size[m->i_pixel].h;
777 const uint16_t *p_cost_mvx = m->p_cost_mv - m->mvp[0];
778 const uint16_t *p_cost_mvy = m->p_cost_mv - m->mvp[1];
779 const int i_pixel = m->i_pixel;
780 const int b_chroma_me = h->mb.b_chroma_me && i_pixel <= PIXEL_8x8;
781 const int mvy_offset = h->mb.b_interlaced & m->i_ref ? (h->mb.i_mb_y & 1)*4 - 2 : 0;
783 ALIGNED_ARRAY_16( uint8_t, pix,[2],[32*18] ); // really 17x17, but round up for alignment
792 /* try the subpel component of the predicted mv */
793 if( hpel_iters && h->mb.i_subpel_refine < 3 )
795 int mx = x264_clip3( m->mvp[0], h->mb.mv_min_spel[0]+2, h->mb.mv_max_spel[0]-2 );
796 int my = x264_clip3( m->mvp[1], h->mb.mv_min_spel[1]+2, h->mb.mv_max_spel[1]-2 );
797 if( (mx-bmx)|(my-bmy) )
798 COST_MV_SAD( mx, my );
801 /* halfpel diamond search */
802 for( i = hpel_iters; i > 0; i-- )
804 int omx = bmx, omy = bmy;
806 int stride = 32; // candidates are either all hpel or all qpel, so one stride is enough
807 uint8_t *src0, *src1, *src2, *src3;
808 src0 = h->mc.get_ref( pix[0], &stride, m->p_fref, m->i_stride[0], omx, omy-2, bw, bh+1, &m->weight[0] );
809 src2 = h->mc.get_ref( pix[1], &stride, m->p_fref, m->i_stride[0], omx-2, omy, bw+4, bh, &m->weight[0] );
810 src1 = src0 + stride;
812 h->pixf.fpelcmp_x4[i_pixel]( m->p_fenc[0], src0, src1, src2, src3, stride, costs );
813 COPY2_IF_LT( bcost, costs[0] + p_cost_mvx[omx ] + p_cost_mvy[omy-2], bmy, omy-2 );
814 COPY2_IF_LT( bcost, costs[1] + p_cost_mvx[omx ] + p_cost_mvy[omy+2], bmy, omy+2 );
815 COPY3_IF_LT( bcost, costs[2] + p_cost_mvx[omx-2] + p_cost_mvy[omy ], bmx, omx-2, bmy, omy );
816 COPY3_IF_LT( bcost, costs[3] + p_cost_mvx[omx+2] + p_cost_mvy[omy ], bmx, omx+2, bmy, omy );
817 if( (bmx == omx) & (bmy == omy) )
824 COST_MV_SATD( bmx, bmy, -1 );
827 /* early termination when examining multiple reference frames */
828 if( p_halfpel_thresh )
830 if( (bcost*7)>>3 > *p_halfpel_thresh )
835 // don't need cost_mv
838 else if( bcost < *p_halfpel_thresh )
839 *p_halfpel_thresh = bcost;
842 /* quarterpel diamond search */
844 for( i = qpel_iters; i > 0; i-- )
846 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] )
851 COST_MV_SATD( omx, omy - 1, 0 );
852 COST_MV_SATD( omx, omy + 1, 1 );
853 COST_MV_SATD( omx - 1, omy, 2 );
854 COST_MV_SATD( omx + 1, omy, 3 );
855 if( bmx == omx && bmy == omy )
862 m->cost_mv = p_cost_mvx[ bmx ] + p_cost_mvy[ bmy ];
865 #define BIME_CACHE( dx, dy, list ) \
867 x264_me_t *m = m##list;\
868 int i = 4 + 3*dx + dy; \
869 int mvx = om##list##x+dx;\
870 int mvy = om##list##y+dy;\
871 stride##list[i] = bw;\
872 src##list[i] = h->mc.get_ref( pixy_buf[list][i], &stride##list[i], m->p_fref, m->i_stride[0], mvx, mvy, bw, bh, weight_none ); \
875 h->mc.mc_chroma( pixu_buf[list][i], 8, m->p_fref[4], m->i_stride[1], mvx, mvy + mv##list##y_offset, bw>>1, bh>>1 );\
876 h->mc.mc_chroma( pixv_buf[list][i], 8, m->p_fref[5], m->i_stride[1], mvx, mvy + mv##list##y_offset, bw>>1, bh>>1 );\
880 #define SATD_THRESH 17/16
882 /* Don't unroll the BIME_CACHE loop. I couldn't find any way to force this
883 * other than making its iteration count not a compile-time constant. */
884 int x264_iter_kludge = 0;
886 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 )
888 static const int pixel_mv_offs[] = { 0, 4, 4*8, 0 };
889 int16_t *cache0_mv = h->mb.cache.mv[0][x264_scan8[i8*4]];
890 int16_t *cache0_mv2 = cache0_mv + pixel_mv_offs[m0->i_pixel];
891 int16_t *cache1_mv = h->mb.cache.mv[1][x264_scan8[i8*4]];
892 int16_t *cache1_mv2 = cache1_mv + pixel_mv_offs[m0->i_pixel];
893 const int i_pixel = m0->i_pixel;
894 const int bw = x264_pixel_size[i_pixel].w;
895 const int bh = x264_pixel_size[i_pixel].h;
896 const uint16_t *p_cost_m0x = m0->p_cost_mv - m0->mvp[0];
897 const uint16_t *p_cost_m0y = m0->p_cost_mv - m0->mvp[1];
898 const uint16_t *p_cost_m1x = m1->p_cost_mv - m1->mvp[0];
899 const uint16_t *p_cost_m1y = m1->p_cost_mv - m1->mvp[1];
900 ALIGNED_ARRAY_16( uint8_t, pixy_buf,[2],[9][16*16] );
901 ALIGNED_8( uint8_t pixu_buf[2][9][8*8] );
902 ALIGNED_8( uint8_t pixv_buf[2][9][8*8] );
905 uint8_t *pix = &h->mb.pic.p_fdec[0][(i8>>1)*8*FDEC_STRIDE+(i8&1)*8];
906 uint8_t *pixu = &h->mb.pic.p_fdec[1][(i8>>1)*4*FDEC_STRIDE+(i8&1)*4];
907 uint8_t *pixv = &h->mb.pic.p_fdec[2][(i8>>1)*4*FDEC_STRIDE+(i8&1)*4];
908 const int ref0 = h->mb.cache.ref[0][x264_scan8[i8*4]];
909 const int ref1 = h->mb.cache.ref[1][x264_scan8[i8*4]];
910 const int mv0y_offset = h->mb.b_interlaced & ref0 ? (h->mb.i_mb_y & 1)*4 - 2 : 0;
911 const int mv1y_offset = h->mb.b_interlaced & ref1 ? (h->mb.i_mb_y & 1)*4 - 2 : 0;
914 int bm0x = m0->mv[0], om0x = bm0x;
915 int bm0y = m0->mv[1], om0y = bm0y;
916 int bm1x = m1->mv[0], om1x = bm1x;
917 int bm1y = m1->mv[1], om1y = bm1y;
918 int bcost = COST_MAX;
921 int mc_list0 = 1, mc_list1 = 1;
922 uint64_t bcostrd = COST_MAX64;
923 /* each byte of visited represents 8 possible m1y positions, so a 4D array isn't needed */
924 ALIGNED_ARRAY_16( uint8_t, visited,[8],[8][8] );
925 /* all permutations of an offset in up to 2 of the dimensions */
926 static const int8_t dia4d[33][4] = {
928 {0,0,0,1}, {0,0,0,-1}, {0,0,1,0}, {0,0,-1,0},
929 {0,1,0,0}, {0,-1,0,0}, {1,0,0,0}, {-1,0,0,0},
930 {0,0,1,1}, {0,0,-1,-1},{0,1,1,0}, {0,-1,-1,0},
931 {1,1,0,0}, {-1,-1,0,0},{1,0,0,1}, {-1,0,0,-1},
932 {0,1,0,1}, {0,-1,0,-1},{1,0,1,0}, {-1,0,-1,0},
933 {0,0,-1,1},{0,0,1,-1}, {0,-1,1,0},{0,1,-1,0},
934 {-1,1,0,0},{1,-1,0,0}, {1,0,0,-1},{-1,0,0,1},
935 {0,-1,0,1},{0,1,0,-1}, {-1,0,1,0},{1,0,-1,0},
938 if( bm0y < h->mb.mv_min_spel[1] + 8 || bm1y < h->mb.mv_min_spel[1] + 8 ||
939 bm0y > h->mb.mv_max_spel[1] - 8 || bm1y > h->mb.mv_max_spel[1] - 8 ||
940 bm0x < h->mb.mv_min_spel[0] + 8 || bm1x < h->mb.mv_min_spel[0] + 8 ||
941 bm0x > h->mb.mv_max_spel[0] - 8 || bm1x > h->mb.mv_max_spel[0] - 8 )
944 h->mc.memzero_aligned( visited, sizeof(uint8_t[8][8][8]) );
946 for( pass = 0; pass < 8; pass++ )
948 /* check all mv pairs that differ in at most 2 components from the current mvs. */
949 /* doesn't do chroma ME. this probably doesn't matter, as the gains
950 * from bidir ME are the same with and without chroma ME. */
953 for( j = x264_iter_kludge; j < 9; j++ )
954 BIME_CACHE( square1[j][0], square1[j][1], 0 );
957 for( j = x264_iter_kludge; j < 9; j++ )
958 BIME_CACHE( square1[j][0], square1[j][1], 1 );
960 for( j = !!pass; j < 33; j++ )
962 int m0x = dia4d[j][0] + om0x;
963 int m0y = dia4d[j][1] + om0y;
964 int m1x = dia4d[j][2] + om1x;
965 int m1y = dia4d[j][3] + om1y;
966 if( !pass || !((visited[(m0x)&7][(m0y)&7][(m1x)&7] & (1<<((m1y)&7)))) )
968 int i0 = 4 + 3*(m0x-om0x) + (m0y-om0y);
969 int i1 = 4 + 3*(m1x-om1x) + (m1y-om1y);
970 visited[(m0x)&7][(m0y)&7][(m1x)&7] |= (1<<((m1y)&7));
971 h->mc.avg[i_pixel]( pix, FDEC_STRIDE, src0[i0], stride0[i0], src1[i1], stride1[i1], i_weight );
972 int cost = h->pixf.mbcmp[i_pixel]( m0->p_fenc[0], FENC_STRIDE, pix, FDEC_STRIDE )
973 + p_cost_m0x[m0x] + p_cost_m0y[m0y] + p_cost_m1x[m1x] + p_cost_m1y[m1y];
976 if( cost < bcost * SATD_THRESH )
978 bcost = X264_MIN( cost, bcost );
979 M32( cache0_mv ) = pack16to32_mask(m0x,m0y);
980 M32( cache0_mv2 ) = pack16to32_mask(m0x,m0y);
981 M32( cache1_mv ) = pack16to32_mask(m1x,m1y);
982 M32( cache1_mv2 ) = pack16to32_mask(m1x,m1y);
983 h->mc.avg[i_pixel+3]( pixu, FDEC_STRIDE, pixu_buf[0][i0], 8, pixu_buf[1][i1], 8, i_weight );
984 h->mc.avg[i_pixel+3]( pixv, FDEC_STRIDE, pixv_buf[0][i0], 8, pixv_buf[1][i1], 8, i_weight );
985 uint64_t costrd = x264_rd_cost_part( h, i_lambda2, i8*4, m0->i_pixel );
986 COPY5_IF_LT( bcostrd, costrd, bm0x, m0x, bm0y, m0y, bm1x, m1x, bm1y, m1y );
990 COPY5_IF_LT( bcost, cost, bm0x, m0x, bm0y, m0y, bm1x, m1x, bm1y, m1y );
994 mc_list0 = (om0x-bm0x)|(om0y-bm0y);
995 mc_list1 = (om1x-bm1x)|(om1y-bm1y);
996 if( !mc_list0 && !mc_list1 )
1011 void x264_me_refine_bidir_satd( x264_t *h, x264_me_t *m0, x264_me_t *m1, int i_weight )
1013 x264_me_refine_bidir( h, m0, m1, i_weight, 0, 0, 0 );
1016 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 )
1018 /* Motion compensation is done as part of bidir_rd; don't repeat
1019 * it in encoding. */
1020 h->mb.b_skip_mc = 1;
1021 x264_me_refine_bidir( h, m0, m1, i_weight, i8, i_lambda2, 1 );
1022 h->mb.b_skip_mc = 0;
1026 #define COST_MV_SATD( mx, my, dst, avoid_mvp ) \
1028 if( !avoid_mvp || !(mx == pmx && my == pmy) ) \
1030 h->mc.mc_luma( pix, FDEC_STRIDE, m->p_fref, m->i_stride[0], mx, my, bw, bh, &m->weight[0] ); \
1031 dst = h->pixf.mbcmp[i_pixel]( m->p_fenc[0], FENC_STRIDE, pix, FDEC_STRIDE ) \
1032 + p_cost_mvx[mx] + p_cost_mvy[my]; \
1033 COPY1_IF_LT( bsatd, dst ); \
1039 #define COST_MV_RD( mx, my, satd, do_dir, mdir ) \
1041 if( satd <= bsatd * SATD_THRESH ) \
1044 M32( cache_mv ) = pack16to32_mask(mx,my); \
1045 M32( cache_mv2 ) = pack16to32_mask(mx,my); \
1046 if( m->i_pixel <= PIXEL_8x8 )\
1048 h->mc.mc_chroma( pixu, FDEC_STRIDE, m->p_fref[4], m->i_stride[1], mx, my + mvy_offset, bw>>1, bh>>1 );\
1049 h->mc.mc_chroma( pixv, FDEC_STRIDE, m->p_fref[5], m->i_stride[1], mx, my + mvy_offset, bw>>1, bh>>1 );\
1051 cost = x264_rd_cost_part( h, i_lambda2, i4, m->i_pixel ); \
1052 COPY4_IF_LT( bcost, cost, bmx, mx, bmy, my, dir, do_dir?mdir:dir ); \
1056 void x264_me_refine_qpel_rd( x264_t *h, x264_me_t *m, int i_lambda2, int i4, int i_list )
1058 // don't have to fill the whole mv cache rectangle
1059 static const int pixel_mv_offs[] = { 0, 4, 4*8, 0, 2, 2*8, 0 };
1060 int16_t *cache_mv = h->mb.cache.mv[i_list][x264_scan8[i4]];
1061 int16_t *cache_mv2 = cache_mv + pixel_mv_offs[m->i_pixel];
1062 const uint16_t *p_cost_mvx, *p_cost_mvy;
1063 const int bw = x264_pixel_size[m->i_pixel].w;
1064 const int bh = x264_pixel_size[m->i_pixel].h;
1065 const int i_pixel = m->i_pixel;
1066 const int mvy_offset = h->mb.b_interlaced & m->i_ref ? (h->mb.i_mb_y & 1)*4 - 2 : 0;
1068 uint64_t bcost = COST_MAX64;
1071 int omx, omy, pmx, pmy, i, j;
1077 uint8_t *pix = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[i4]];
1078 uint8_t *pixu = &h->mb.pic.p_fdec[1][(i8>>1)*4*FDEC_STRIDE+(i8&1)*4];
1079 uint8_t *pixv = &h->mb.pic.p_fdec[2][(i8>>1)*4*FDEC_STRIDE+(i8&1)*4];
1081 h->mb.b_skip_mc = 1;
1083 if( m->i_pixel != PIXEL_16x16 && i4 != 0 )
1084 x264_mb_predict_mv( h, i_list, i4, bw>>2, m->mvp );
1087 p_cost_mvx = m->p_cost_mv - pmx;
1088 p_cost_mvy = m->p_cost_mv - pmy;
1089 COST_MV_SATD( bmx, bmy, bsatd, 0 );
1090 if( m->i_pixel != PIXEL_16x16 )
1091 COST_MV_RD( bmx, bmy, 0, 0, 0 )
1095 /* check the predicted mv */
1096 if( (bmx != pmx || bmy != pmy)
1097 && pmx >= h->mb.mv_min_spel[0] && pmx <= h->mb.mv_max_spel[0]
1098 && pmy >= h->mb.mv_min_spel[1] && pmy <= h->mb.mv_max_spel[1] )
1100 COST_MV_SATD( pmx, pmy, satd, 0 );
1101 COST_MV_RD ( pmx, pmy, satd, 0, 0 );
1102 /* The hex motion search is guaranteed to not repeat the center candidate,
1103 * so if pmv is chosen, set the "MV to avoid checking" to bmv instead. */
1104 if( bmx == pmx && bmy == pmy )
1111 if( bmy < h->mb.mv_min_spel[1] + 3 || bmy > h->mb.mv_max_spel[1] - 3 ||
1112 bmx < h->mb.mv_min_spel[0] + 3 || bmx > h->mb.mv_max_spel[0] - 3 )
1114 h->mb.b_skip_mc = 0;
1118 /* subpel hex search, same pattern as ME HEX. */
1122 for( j=0; j<6; j++ )
1124 COST_MV_SATD( omx + hex2[j+1][0], omy + hex2[j+1][1], satd, 1 );
1125 COST_MV_RD ( omx + hex2[j+1][0], omy + hex2[j+1][1], satd, 1, j );
1130 /* half hexagon, not overlapping the previous iteration */
1131 for( i = 1; i < 10; i++ )
1133 const int odir = mod6m1[dir+1];
1134 if( bmy < h->mb.mv_min_spel[1] + 3 ||
1135 bmy > h->mb.mv_max_spel[1] - 3 )
1140 for( j=0; j<3; j++ )
1142 COST_MV_SATD( omx + hex2[odir+j][0], omy + hex2[odir+j][1], satd, 1 );
1143 COST_MV_RD ( omx + hex2[odir+j][0], omy + hex2[odir+j][1], satd, 1, odir-1+j );
1150 /* square refine, same pattern as ME HEX. */
1153 for( i=0; i<8; i++ )
1155 COST_MV_SATD( omx + square1[i+1][0], omy + square1[i+1][1], satd, 1 );
1156 COST_MV_RD ( omx + square1[i+1][0], omy + square1[i+1][1], satd, 0, 0 );
1162 x264_macroblock_cache_mv ( h, block_idx_x[i4], block_idx_y[i4], bw>>2, bh>>2, i_list, pack16to32_mask(bmx, bmy) );
1163 x264_macroblock_cache_mvd( h, block_idx_x[i4], block_idx_y[i4], bw>>2, bh>>2, i_list, pack16to32_mask(bmx - m->mvp[0], bmy - m->mvp[1]) );
1164 h->mb.b_skip_mc = 0;