1 /*****************************************************************************
2 * macroblock.c: macroblock common functions
3 *****************************************************************************
4 * Copyright (C) 2003-2011 x264 project
6 * Authors: Fiona Glaser <fiona@x264.com>
7 * Laurent Aimar <fenrir@via.ecp.fr>
8 * Loren Merritt <lorenm@u.washington.edu>
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 *****************************************************************************/
29 #include "encoder/me.h"
31 static NOINLINE void x264_mb_mc_0xywh( x264_t *h, int x, int y, int width, int height )
33 int i8 = x264_scan8[0]+x+8*y;
34 int i_ref = h->mb.cache.ref[0][i8];
35 int mvx = x264_clip3( h->mb.cache.mv[0][i8][0], h->mb.mv_min[0], h->mb.mv_max[0] ) + 4*4*x;
36 int mvy = x264_clip3( h->mb.cache.mv[0][i8][1], h->mb.mv_min[1], h->mb.mv_max[1] ) + 4*4*y;
38 h->mc.mc_luma( &h->mb.pic.p_fdec[0][4*y*FDEC_STRIDE+4*x], FDEC_STRIDE,
39 h->mb.pic.p_fref[0][i_ref], h->mb.pic.i_stride[0],
40 mvx, mvy, 4*width, 4*height, &h->sh.weight[i_ref][0] );
42 // chroma is offset if MCing from a field of opposite parity
43 if( h->mb.b_interlaced & i_ref )
44 mvy += (h->mb.i_mb_y & 1)*4 - 2;
46 h->mc.mc_chroma( &h->mb.pic.p_fdec[1][2*y*FDEC_STRIDE+2*x],
47 &h->mb.pic.p_fdec[2][2*y*FDEC_STRIDE+2*x], FDEC_STRIDE,
48 h->mb.pic.p_fref[0][i_ref][4], h->mb.pic.i_stride[1],
49 mvx, mvy, 2*width, 2*height );
51 if( h->sh.weight[i_ref][1].weightfn )
52 h->sh.weight[i_ref][1].weightfn[width>>1]( &h->mb.pic.p_fdec[1][2*y*FDEC_STRIDE+2*x], FDEC_STRIDE,
53 &h->mb.pic.p_fdec[1][2*y*FDEC_STRIDE+2*x], FDEC_STRIDE,
54 &h->sh.weight[i_ref][1], height*2 );
55 if( h->sh.weight[i_ref][2].weightfn )
56 h->sh.weight[i_ref][2].weightfn[width>>1]( &h->mb.pic.p_fdec[2][2*y*FDEC_STRIDE+2*x], FDEC_STRIDE,
57 &h->mb.pic.p_fdec[2][2*y*FDEC_STRIDE+2*x], FDEC_STRIDE,
58 &h->sh.weight[i_ref][2],height*2 );
61 static NOINLINE void x264_mb_mc_1xywh( x264_t *h, int x, int y, int width, int height )
63 int i8 = x264_scan8[0]+x+8*y;
64 int i_ref = h->mb.cache.ref[1][i8];
65 int mvx = x264_clip3( h->mb.cache.mv[1][i8][0], h->mb.mv_min[0], h->mb.mv_max[0] ) + 4*4*x;
66 int mvy = x264_clip3( h->mb.cache.mv[1][i8][1], h->mb.mv_min[1], h->mb.mv_max[1] ) + 4*4*y;
68 h->mc.mc_luma( &h->mb.pic.p_fdec[0][4*y*FDEC_STRIDE+4*x], FDEC_STRIDE,
69 h->mb.pic.p_fref[1][i_ref], h->mb.pic.i_stride[0],
70 mvx, mvy, 4*width, 4*height, weight_none );
72 if( h->mb.b_interlaced & i_ref )
73 mvy += (h->mb.i_mb_y & 1)*4 - 2;
75 h->mc.mc_chroma( &h->mb.pic.p_fdec[1][2*y*FDEC_STRIDE+2*x],
76 &h->mb.pic.p_fdec[2][2*y*FDEC_STRIDE+2*x], FDEC_STRIDE,
77 h->mb.pic.p_fref[1][i_ref][4], h->mb.pic.i_stride[1],
78 mvx, mvy, 2*width, 2*height );
81 static NOINLINE void x264_mb_mc_01xywh( x264_t *h, int x, int y, int width, int height )
83 int i8 = x264_scan8[0]+x+8*y;
84 int i_ref0 = h->mb.cache.ref[0][i8];
85 int i_ref1 = h->mb.cache.ref[1][i8];
86 int weight = h->mb.bipred_weight[i_ref0][i_ref1];
87 int mvx0 = x264_clip3( h->mb.cache.mv[0][i8][0], h->mb.mv_min[0], h->mb.mv_max[0] ) + 4*4*x;
88 int mvx1 = x264_clip3( h->mb.cache.mv[1][i8][0], h->mb.mv_min[0], h->mb.mv_max[0] ) + 4*4*x;
89 int mvy0 = x264_clip3( h->mb.cache.mv[0][i8][1], h->mb.mv_min[1], h->mb.mv_max[1] ) + 4*4*y;
90 int mvy1 = x264_clip3( h->mb.cache.mv[1][i8][1], h->mb.mv_min[1], h->mb.mv_max[1] ) + 4*4*y;
91 int i_mode = x264_size2pixel[height][width];
92 int i_stride0 = 16, i_stride1 = 16;
93 ALIGNED_ARRAY_16( pixel, tmp0,[16*16] );
94 ALIGNED_ARRAY_16( pixel, tmp1,[16*16] );
97 src0 = h->mc.get_ref( tmp0, &i_stride0, h->mb.pic.p_fref[0][i_ref0], h->mb.pic.i_stride[0],
98 mvx0, mvy0, 4*width, 4*height, weight_none );
99 src1 = h->mc.get_ref( tmp1, &i_stride1, h->mb.pic.p_fref[1][i_ref1], h->mb.pic.i_stride[0],
100 mvx1, mvy1, 4*width, 4*height, weight_none );
101 h->mc.avg[i_mode]( &h->mb.pic.p_fdec[0][4*y*FDEC_STRIDE+4*x], FDEC_STRIDE,
102 src0, i_stride0, src1, i_stride1, weight );
104 if( h->mb.b_interlaced & i_ref0 )
105 mvy0 += (h->mb.i_mb_y & 1)*4 - 2;
106 if( h->mb.b_interlaced & i_ref1 )
107 mvy1 += (h->mb.i_mb_y & 1)*4 - 2;
109 h->mc.mc_chroma( tmp0, tmp0+8, 16, h->mb.pic.p_fref[0][i_ref0][4], h->mb.pic.i_stride[1],
110 mvx0, mvy0, 2*width, 2*height );
111 h->mc.mc_chroma( tmp1, tmp1+8, 16, h->mb.pic.p_fref[1][i_ref1][4], h->mb.pic.i_stride[1],
112 mvx1, mvy1, 2*width, 2*height );
113 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 );
114 h->mc.avg[i_mode+3]( &h->mb.pic.p_fdec[2][2*y*FDEC_STRIDE+2*x], FDEC_STRIDE, tmp0+8, 16, tmp1+8, 16, weight );
117 void x264_mb_mc_8x8( x264_t *h, int i8 )
122 if( h->sh.i_type == SLICE_TYPE_P )
124 switch( h->mb.i_sub_partition[i8] )
127 x264_mb_mc_0xywh( h, x, y, 2, 2 );
130 x264_mb_mc_0xywh( h, x, y+0, 2, 1 );
131 x264_mb_mc_0xywh( h, x, y+1, 2, 1 );
134 x264_mb_mc_0xywh( h, x+0, y, 1, 2 );
135 x264_mb_mc_0xywh( h, x+1, y, 1, 2 );
138 x264_mb_mc_0xywh( h, x+0, y+0, 1, 1 );
139 x264_mb_mc_0xywh( h, x+1, y+0, 1, 1 );
140 x264_mb_mc_0xywh( h, x+0, y+1, 1, 1 );
141 x264_mb_mc_0xywh( h, x+1, y+1, 1, 1 );
147 int scan8 = x264_scan8[0] + x + 8*y;
149 if( h->mb.cache.ref[0][scan8] >= 0 )
150 if( h->mb.cache.ref[1][scan8] >= 0 )
151 x264_mb_mc_01xywh( h, x, y, 2, 2 );
153 x264_mb_mc_0xywh( h, x, y, 2, 2 );
155 x264_mb_mc_1xywh( h, x, y, 2, 2 );
159 void x264_mb_mc( x264_t *h )
161 if( h->mb.i_partition == D_8x8 )
163 for( int i = 0; i < 4; i++ )
164 x264_mb_mc_8x8( h, i );
168 int ref0a = h->mb.cache.ref[0][x264_scan8[ 0]];
169 int ref0b = h->mb.cache.ref[0][x264_scan8[12]];
170 int ref1a = h->mb.cache.ref[1][x264_scan8[ 0]];
171 int ref1b = h->mb.cache.ref[1][x264_scan8[12]];
173 if( h->mb.i_partition == D_16x16 )
176 if( ref1a >= 0 ) x264_mb_mc_01xywh( h, 0, 0, 4, 4 );
177 else x264_mb_mc_0xywh ( h, 0, 0, 4, 4 );
178 else x264_mb_mc_1xywh ( h, 0, 0, 4, 4 );
180 else if( h->mb.i_partition == D_16x8 )
183 if( ref1a >= 0 ) x264_mb_mc_01xywh( h, 0, 0, 4, 2 );
184 else x264_mb_mc_0xywh ( h, 0, 0, 4, 2 );
185 else x264_mb_mc_1xywh ( h, 0, 0, 4, 2 );
188 if( ref1b >= 0 ) x264_mb_mc_01xywh( h, 0, 2, 4, 2 );
189 else x264_mb_mc_0xywh ( h, 0, 2, 4, 2 );
190 else x264_mb_mc_1xywh ( h, 0, 2, 4, 2 );
192 else if( h->mb.i_partition == D_8x16 )
195 if( ref1a >= 0 ) x264_mb_mc_01xywh( h, 0, 0, 2, 4 );
196 else x264_mb_mc_0xywh ( h, 0, 0, 2, 4 );
197 else x264_mb_mc_1xywh ( h, 0, 0, 2, 4 );
200 if( ref1b >= 0 ) x264_mb_mc_01xywh( h, 2, 0, 2, 4 );
201 else x264_mb_mc_0xywh ( h, 2, 0, 2, 4 );
202 else x264_mb_mc_1xywh ( h, 2, 0, 2, 4 );
207 int x264_macroblock_cache_allocate( x264_t *h )
209 int i_mb_count = h->mb.i_mb_count;
211 h->mb.i_mb_stride = h->mb.i_mb_width;
212 h->mb.i_b8_stride = h->mb.i_mb_width * 2;
213 h->mb.i_b4_stride = h->mb.i_mb_width * 4;
215 h->mb.b_interlaced = h->param.b_interlaced;
217 CHECKED_MALLOC( h->mb.qp, i_mb_count * sizeof(int8_t) );
218 CHECKED_MALLOC( h->mb.cbp, i_mb_count * sizeof(int16_t) );
219 CHECKED_MALLOC( h->mb.skipbp, i_mb_count * sizeof(int8_t) );
220 CHECKED_MALLOC( h->mb.mb_transform_size, i_mb_count * sizeof(int8_t) );
221 CHECKED_MALLOC( h->mb.slice_table, i_mb_count * sizeof(uint16_t) );
222 memset( h->mb.slice_table, -1, i_mb_count * sizeof(uint16_t) );
224 /* 0 -> 3 top(4), 4 -> 6 : left(3) */
225 CHECKED_MALLOC( h->mb.intra4x4_pred_mode, i_mb_count * 8 * sizeof(int8_t) );
228 CHECKED_MALLOC( h->mb.non_zero_count, i_mb_count * 24 * sizeof(uint8_t) );
230 if( h->param.b_cabac )
232 CHECKED_MALLOC( h->mb.chroma_pred_mode, i_mb_count * sizeof(int8_t) );
233 CHECKED_MALLOC( h->mb.mvd[0], i_mb_count * sizeof( **h->mb.mvd ) );
234 CHECKED_MALLOC( h->mb.mvd[1], i_mb_count * sizeof( **h->mb.mvd ) );
237 for( int i = 0; i < 2; i++ )
239 int i_refs = X264_MIN(X264_REF_MAX, (i ? 1 + !!h->param.i_bframe_pyramid : h->param.i_frame_reference) ) << h->param.b_interlaced;
240 if( h->param.analyse.i_weighted_pred == X264_WEIGHTP_SMART )
241 i_refs = X264_MIN(X264_REF_MAX, i_refs + 1 + (BIT_DEPTH == 8)); //smart weights add two duplicate frames, one in >8-bit
243 for( int j = !i; j < i_refs; j++ )
245 CHECKED_MALLOC( h->mb.mvr[i][j], 2 * (i_mb_count + 1) * sizeof(int16_t) );
246 M32( h->mb.mvr[i][j][0] ) = 0;
251 if( h->param.analyse.i_weighted_pred )
253 int i_padv = PADV << h->param.b_interlaced;
254 int luma_plane_size = 0;
257 if( h->param.analyse.i_weighted_pred == X264_WEIGHTP_FAKE )
259 // only need buffer for lookahead
260 if( !h->param.i_sync_lookahead || h == h->thread[h->param.i_threads] )
262 // Fake analysis only works on lowres
263 luma_plane_size = h->fdec->i_stride_lowres * (h->mb.i_mb_height*8+2*i_padv);
264 // Only need 1 buffer for analysis
272 luma_plane_size = h->fdec->i_stride[0] * (h->mb.i_mb_height*16+2*i_padv);
274 if( h->param.analyse.i_weighted_pred == X264_WEIGHTP_SMART )
275 //smart can weight one ref and one offset -1 in 8-bit
276 numweightbuf = 1 + (BIT_DEPTH == 8);
278 //simple only has one weighted ref
282 for( int i = 0; i < numweightbuf; i++ )
283 CHECKED_MALLOC( h->mb.p_weight_buf[i], luma_plane_size * sizeof(pixel) );
290 void x264_macroblock_cache_free( x264_t *h )
292 for( int i = 0; i < 2; i++ )
293 for( int j = !i; j < X264_REF_MAX*2; j++ )
294 if( h->mb.mvr[i][j] )
295 x264_free( h->mb.mvr[i][j]-1 );
296 for( int i = 0; i < X264_REF_MAX; i++ )
297 x264_free( h->mb.p_weight_buf[i] );
299 if( h->param.b_cabac )
301 x264_free( h->mb.chroma_pred_mode );
302 x264_free( h->mb.mvd[0] );
303 x264_free( h->mb.mvd[1] );
305 x264_free( h->mb.slice_table );
306 x264_free( h->mb.intra4x4_pred_mode );
307 x264_free( h->mb.non_zero_count );
308 x264_free( h->mb.mb_transform_size );
309 x264_free( h->mb.skipbp );
310 x264_free( h->mb.cbp );
311 x264_free( h->mb.qp );
314 int x264_macroblock_thread_allocate( x264_t *h, int b_lookahead )
318 for( int i = 0; i <= 4*h->param.b_interlaced; i++ )
319 for( int j = 0; j < 2; j++ )
321 /* shouldn't really be initialized, just silences a valgrind false-positive in predict_8x8_filter_mmx */
322 CHECKED_MALLOCZERO( h->intra_border_backup[i][j], (h->sps->i_mb_width*16+32) * sizeof(pixel) );
323 h->intra_border_backup[i][j] += 16;
324 if( !h->param.b_interlaced )
325 h->intra_border_backup[1][j] = h->intra_border_backup[i][j];
327 for( int i = 0; i <= h->param.b_interlaced; i++ )
329 CHECKED_MALLOC( h->deblock_strength[i], sizeof(**h->deblock_strength) * h->mb.i_mb_width );
330 h->deblock_strength[1] = h->deblock_strength[i];
334 /* Allocate scratch buffer */
335 int scratch_size = 0;
338 int buf_hpel = (h->thread[0]->fdec->i_width[0]+48) * sizeof(int16_t);
339 int buf_ssim = h->param.analyse.b_ssim * 8 * (h->param.i_width/4+3) * sizeof(int);
340 int me_range = X264_MIN(h->param.analyse.i_me_range, h->param.analyse.i_mv_range);
341 int buf_tesa = (h->param.analyse.i_me_method >= X264_ME_ESA) *
342 ((me_range*2+24) * sizeof(int16_t) + (me_range+4) * (me_range+1) * 4 * sizeof(mvsad_t));
343 scratch_size = X264_MAX3( buf_hpel, buf_ssim, buf_tesa );
345 int buf_mbtree = h->param.rc.b_mb_tree * ((h->mb.i_mb_width+3)&~3) * sizeof(int);
346 scratch_size = X264_MAX( scratch_size, buf_mbtree );
348 CHECKED_MALLOC( h->scratch_buffer, scratch_size );
350 h->scratch_buffer = NULL;
357 void x264_macroblock_thread_free( x264_t *h, int b_lookahead )
361 for( int i = 0; i <= h->param.b_interlaced; i++ )
362 x264_free( h->deblock_strength[i] );
363 for( int i = 0; i <= 4*h->param.b_interlaced; i++ )
364 for( int j = 0; j < 2; j++ )
365 x264_free( h->intra_border_backup[i][j] - 16 );
367 x264_free( h->scratch_buffer );
370 void x264_macroblock_slice_init( x264_t *h )
372 h->mb.mv[0] = h->fdec->mv[0];
373 h->mb.mv[1] = h->fdec->mv[1];
374 h->mb.mvr[0][0] = h->fdec->mv16x16;
375 h->mb.ref[0] = h->fdec->ref[0];
376 h->mb.ref[1] = h->fdec->ref[1];
377 h->mb.type = h->fdec->mb_type;
378 h->mb.partition = h->fdec->mb_partition;
379 h->mb.field = h->fdec->field;
381 h->fdec->i_ref[0] = h->i_ref[0];
382 h->fdec->i_ref[1] = h->i_ref[1];
383 for( int i = 0; i < h->i_ref[0]; i++ )
384 h->fdec->ref_poc[0][i] = h->fref[0][i]->i_poc;
385 if( h->sh.i_type == SLICE_TYPE_B )
387 for( int i = 0; i < h->i_ref[1]; i++ )
388 h->fdec->ref_poc[1][i] = h->fref[1][i]->i_poc;
390 map_col_to_list0(-1) = -1;
391 map_col_to_list0(-2) = -2;
392 for( int i = 0; i < h->fref[1][0]->i_ref[0]; i++ )
394 int poc = h->fref[1][0]->ref_poc[0][i];
395 map_col_to_list0(i) = -2;
396 for( int j = 0; j < h->i_ref[0]; j++ )
397 if( h->fref[0][j]->i_poc == poc )
399 map_col_to_list0(i) = j;
404 else if( h->sh.i_type == SLICE_TYPE_P )
406 memset( h->mb.cache.skip, 0, sizeof( h->mb.cache.skip ) );
408 if( h->sh.i_disable_deblocking_filter_idc != 1 && h->param.analyse.i_weighted_pred == X264_WEIGHTP_SMART )
410 deblock_ref_table(-2) = -2;
411 deblock_ref_table(-1) = -1;
412 for( int i = 0; i < h->i_ref[0] << h->sh.b_mbaff; i++ )
414 /* Mask off high bits to avoid frame num collisions with -1/-2.
415 * In current x264 frame num values don't cover a range of more
416 * than 32, so 6 bits is enough for uniqueness. */
417 if( !h->mb.b_interlaced )
418 deblock_ref_table(i) = h->fref[0][i]->i_frame_num&63;
420 deblock_ref_table(i) = ((h->fref[0][i>>1]->i_frame_num&63)<<1) + (i&1);
425 /* init with not available (for top right idx=7,15) */
426 memset( h->mb.cache.ref, -2, sizeof( h->mb.cache.ref ) );
428 if( h->i_ref[0] > 0 )
429 for( int field = 0; field <= h->sh.b_mbaff; field++ )
431 int curpoc = h->fdec->i_poc + h->fdec->i_delta_poc[field];
432 int refpoc = h->fref[0][0]->i_poc + h->fref[0][0]->i_delta_poc[field];
433 int delta = curpoc - refpoc;
435 h->fdec->inv_ref_poc[field] = (256 + delta/2) / delta;
438 h->mb.i_neighbour4[6] =
439 h->mb.i_neighbour4[9] =
440 h->mb.i_neighbour4[12] =
441 h->mb.i_neighbour4[14] = MB_LEFT|MB_TOP|MB_TOPLEFT|MB_TOPRIGHT;
442 h->mb.i_neighbour4[3] =
443 h->mb.i_neighbour4[7] =
444 h->mb.i_neighbour4[11] =
445 h->mb.i_neighbour4[13] =
446 h->mb.i_neighbour4[15] =
447 h->mb.i_neighbour8[3] = MB_LEFT|MB_TOP|MB_TOPLEFT;
450 void x264_macroblock_thread_init( x264_t *h )
452 h->mb.i_me_method = h->param.analyse.i_me_method;
453 h->mb.i_subpel_refine = h->param.analyse.i_subpel_refine;
454 if( h->sh.i_type == SLICE_TYPE_B && (h->mb.i_subpel_refine == 6 || h->mb.i_subpel_refine == 8) )
455 h->mb.i_subpel_refine--;
456 h->mb.b_chroma_me = h->param.analyse.b_chroma_me &&
457 ((h->sh.i_type == SLICE_TYPE_P && h->mb.i_subpel_refine >= 5) ||
458 (h->sh.i_type == SLICE_TYPE_B && h->mb.i_subpel_refine >= 9));
459 h->mb.b_dct_decimate = h->sh.i_type == SLICE_TYPE_B ||
460 (h->param.analyse.b_dct_decimate && h->sh.i_type != SLICE_TYPE_I);
461 h->mb.i_mb_prev_xy = -1;
473 h->mb.pic.p_fenc[0] = h->mb.pic.fenc_buf;
474 h->mb.pic.p_fenc[1] = h->mb.pic.fenc_buf + 16*FENC_STRIDE;
475 h->mb.pic.p_fenc[2] = h->mb.pic.fenc_buf + 16*FENC_STRIDE + 8;
476 h->mb.pic.p_fdec[0] = h->mb.pic.fdec_buf + 2*FDEC_STRIDE;
477 h->mb.pic.p_fdec[1] = h->mb.pic.fdec_buf + 19*FDEC_STRIDE;
478 h->mb.pic.p_fdec[2] = h->mb.pic.fdec_buf + 19*FDEC_STRIDE + 16;
481 void x264_prefetch_fenc( x264_t *h, x264_frame_t *fenc, int i_mb_x, int i_mb_y )
483 int stride_y = fenc->i_stride[0];
484 int stride_uv = fenc->i_stride[1];
485 int off_y = 16 * i_mb_x + 16 * i_mb_y * stride_y;
486 int off_uv = 16 * i_mb_x + 8 * i_mb_y * stride_uv;
487 h->mc.prefetch_fenc( fenc->plane[0]+off_y, stride_y,
488 fenc->plane[1]+off_uv, stride_uv, i_mb_x );
491 NOINLINE void x264_copy_column8( pixel *dst, pixel *src )
493 // input pointers are offset by 4 rows because that's faster (smaller instruction size on x86)
494 for( int i = -4; i < 4; i++ )
495 dst[i*FDEC_STRIDE] = src[i*FDEC_STRIDE];
498 static void ALWAYS_INLINE x264_macroblock_load_pic_pointers( x264_t *h, int mb_x, int mb_y, int i, int b_mbaff )
500 int w = (i ? 8 : 16);
501 int i_stride = h->fdec->i_stride[i];
502 int i_stride2 = i_stride << h->mb.b_interlaced;
503 int i_pix_offset = h->mb.b_interlaced
504 ? 16 * mb_x + w * (mb_y&~1) * i_stride + (mb_y&1) * i_stride
505 : 16 * mb_x + w * mb_y * i_stride;
506 pixel *plane_fdec = &h->fdec->plane[i][i_pix_offset];
507 int fdec_idx = b_mbaff ? (h->mb.b_interlaced ? (3 + (mb_y&1)) : (mb_y&1) ? 2 : 4) : 0;
508 pixel *intra_fdec = &h->intra_border_backup[fdec_idx][i][mb_x*16];
509 int ref_pix_offset[2] = { i_pix_offset, i_pix_offset };
510 /* ref_pix_offset[0] references the current field and [1] the opposite field. */
511 if( h->mb.b_interlaced )
512 ref_pix_offset[1] += (1-2*(mb_y&1)) * i_stride;
513 h->mb.pic.i_stride[i] = i_stride2;
514 h->mb.pic.p_fenc_plane[i] = &h->fenc->plane[i][i_pix_offset];
517 h->mc.load_deinterleave_8x8x2_fenc( h->mb.pic.p_fenc[1], h->mb.pic.p_fenc_plane[1], i_stride2 );
518 memcpy( h->mb.pic.p_fdec[1]-FDEC_STRIDE, intra_fdec, 8*sizeof(pixel) );
519 memcpy( h->mb.pic.p_fdec[2]-FDEC_STRIDE, intra_fdec+8, 8*sizeof(pixel) );
522 h->mb.pic.p_fdec[1][-FDEC_STRIDE-1] = intra_fdec[-1-8];
523 h->mb.pic.p_fdec[2][-FDEC_STRIDE-1] = intra_fdec[-1];
528 h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fenc[0], FENC_STRIDE, h->mb.pic.p_fenc_plane[0], i_stride2, 16 );
529 memcpy( h->mb.pic.p_fdec[0]-FDEC_STRIDE, intra_fdec, 24*sizeof(pixel) );
531 h->mb.pic.p_fdec[0][-FDEC_STRIDE-1] = intra_fdec[-1];
535 for( int j = 0; j < w; j++ )
538 h->mb.pic.p_fdec[1][-1+j*FDEC_STRIDE] = plane_fdec[-2+j*i_stride2];
539 h->mb.pic.p_fdec[2][-1+j*FDEC_STRIDE] = plane_fdec[-1+j*i_stride2];
542 h->mb.pic.p_fdec[0][-1+j*FDEC_STRIDE] = plane_fdec[-1+j*i_stride2];
544 pixel *plane_src, **filtered_src;
545 for( int j = 0; j < h->mb.pic.i_fref[0]; j++ )
547 // Interpolate between pixels in same field.
548 if( h->mb.b_interlaced )
550 plane_src = h->fref[0][j>>1]->plane_fld[i];
551 filtered_src = h->fref[0][j>>1]->filtered_fld;
555 plane_src = h->fref[0][j]->plane[i];
556 filtered_src = h->fref[0][j]->filtered;
558 h->mb.pic.p_fref[0][j][i?4:0] = plane_src + ref_pix_offset[j&1];
562 for( int k = 1; k < 4; k++ )
563 h->mb.pic.p_fref[0][j][k] = filtered_src[k] + ref_pix_offset[j&1];
564 if( h->sh.weight[j][0].weightfn )
565 h->mb.pic.p_fref_w[j] = &h->fenc->weighted[j >> h->mb.b_interlaced][ref_pix_offset[j&1]];
567 h->mb.pic.p_fref_w[j] = h->mb.pic.p_fref[0][j][0];
570 if( h->sh.i_type == SLICE_TYPE_B )
571 for( int j = 0; j < h->mb.pic.i_fref[1]; j++ )
573 if( h->mb.b_interlaced )
575 plane_src = h->fref[1][j>>1]->plane_fld[i];
576 filtered_src = h->fref[1][j>>1]->filtered_fld;
580 plane_src = h->fref[1][j]->plane[i];
581 filtered_src = h->fref[1][j]->filtered;
583 h->mb.pic.p_fref[1][j][i?4:0] = plane_src + ref_pix_offset[j&1];
586 for( int k = 1; k < 4; k++ )
587 h->mb.pic.p_fref[1][j][k] = filtered_src[k] + ref_pix_offset[j&1];
591 x264_left_table_t left_indices[4] =
593 /* Current is progressive */
594 {{ 4, 4, 5, 5}, { 3, 3, 7, 7}, {16+1, 16+1, 16+4+1, 16+4+1}, {0, 0, 1, 1}, {0, 0, 0, 0}},
595 {{ 6, 6, 3, 3}, {11, 11, 15, 15}, {16+3, 16+3, 16+4+3, 16+4+3}, {2, 2, 3, 3}, {1, 1, 1, 1}},
596 /* Current is interlaced */
597 {{ 4, 6, 4, 6}, { 3, 11, 3, 11}, {16+1, 16+1, 16+4+1, 16+4+1}, {0, 2, 0, 2}, {0, 1, 0, 1}},
599 {{ 4, 5, 6, 3}, { 3, 7, 11, 15}, {16+1, 16+3, 16+4+1, 16+4+3}, {0, 1, 2, 3}, {0, 0, 1, 1}}
602 static void inline x264_macroblock_cache_load_neighbours( x264_t *h, int mb_x, int mb_y )
604 int top_y = mb_y - (1 << h->mb.b_interlaced);
605 int top = top_y * h->mb.i_mb_stride + mb_x;
609 h->mb.i_mb_xy = mb_y * h->mb.i_mb_stride + mb_x;
610 h->mb.i_b8_xy = 2*(mb_y * h->mb.i_b8_stride + mb_x);
611 h->mb.i_b4_xy = 4*(mb_y * h->mb.i_b4_stride + mb_x);
613 h->mb.left_b8[1] = -1;
615 h->mb.left_b4[1] = -1;
616 h->mb.i_neighbour = 0;
617 h->mb.i_neighbour_intra = 0;
618 h->mb.i_neighbour_frame = 0;
619 h->mb.i_mb_top_xy = -1;
620 h->mb.i_mb_top_y = -1;
621 h->mb.i_mb_left_xy[0] = h->mb.i_mb_left_xy[1] = -1;
622 h->mb.i_mb_topleft_xy = -1;
623 h->mb.i_mb_topright_xy = -1;
624 h->mb.i_mb_topleft_y = -1;
625 h->mb.i_mb_topright_y = -1;
626 h->mb.i_mb_top_mbpair_xy = h->mb.i_mb_xy - 2*h->mb.i_mb_stride;
627 h->mb.i_mb_type_top = -1;
628 h->mb.i_mb_type_left[0] = h->mb.i_mb_type_left[1] = -1;
629 h->mb.i_mb_type_topleft = -1;
630 h->mb.i_mb_type_topright = -1;
631 h->mb.left_index_table = &left_indices[3];
632 h->mb.topleft_partition = 0;
634 int topleft_y = top_y;
635 int topright_y = top_y;
638 left[0] = left[1] = h->mb.i_mb_xy - 1;
639 h->mb.left_b8[0] = h->mb.left_b8[1] = h->mb.i_b8_xy - 2;
640 h->mb.left_b4[0] = h->mb.left_b4[1] = h->mb.i_b4_xy - 4;
646 if( mb_x && h->mb.b_interlaced != h->mb.field[h->mb.i_mb_xy-1] )
648 left[0] = left[1] = h->mb.i_mb_xy - 1 - h->mb.i_mb_stride;
649 h->mb.left_b8[0] = h->mb.left_b8[1] = h->mb.i_b8_xy - 2 - 2*h->mb.i_b8_stride;
650 h->mb.left_b4[0] = h->mb.left_b4[1] = h->mb.i_b4_xy - 4 - 4*h->mb.i_b4_stride;
652 if( h->mb.b_interlaced )
654 h->mb.left_index_table = &left_indices[2];
655 left[1] += h->mb.i_mb_stride;
656 h->mb.left_b8[1] += 2*h->mb.i_b8_stride;
657 h->mb.left_b4[1] += 4*h->mb.i_b4_stride;
661 h->mb.left_index_table = &left_indices[1];
663 h->mb.topleft_partition = 1;
666 if( !h->mb.b_interlaced )
671 if( h->mb.b_interlaced && top >= 0 )
673 if( !h->mb.field[top] )
675 top += h->mb.i_mb_stride;
679 topleft_y += !h->mb.field[h->mb.i_mb_stride*topleft_y + mb_x - 1];
680 if( mb_x < h->mb.i_mb_width-1 )
681 topright_y += !h->mb.field[h->mb.i_mb_stride*topright_y + mb_x + 1];
683 if( mb_x && h->mb.b_interlaced != h->mb.field[h->mb.i_mb_xy-1] )
685 if( h->mb.b_interlaced )
687 h->mb.left_index_table = &left_indices[2];
688 left[1] += h->mb.i_mb_stride;
689 h->mb.left_b8[1] += 2*h->mb.i_b8_stride;
690 h->mb.left_b4[1] += 4*h->mb.i_b4_stride;
693 h->mb.left_index_table = &left_indices[0];
700 h->mb.i_neighbour_frame |= MB_LEFT;
701 h->mb.i_mb_left_xy[0] = left[0];
702 h->mb.i_mb_left_xy[1] = left[1];
703 h->mb.i_mb_type_left[0] = h->mb.type[h->mb.i_mb_left_xy[0]];
704 h->mb.i_mb_type_left[1] = h->mb.type[h->mb.i_mb_left_xy[1]];
705 if( h->mb.i_mb_xy > h->sh.i_first_mb )
707 h->mb.i_neighbour |= MB_LEFT;
709 // FIXME: We don't currently support constrained intra + mbaff.
710 if( !h->param.b_constrained_intra || IS_INTRA( h->mb.i_mb_type_left[0] ) )
711 h->mb.i_neighbour_intra |= MB_LEFT;
715 /* We can't predict from the previous threadslice since it hasn't been encoded yet. */
716 if( (h->i_threadslice_start >> h->mb.b_interlaced) != (mb_y >> h->mb.b_interlaced) )
720 h->mb.i_neighbour_frame |= MB_TOP;
721 h->mb.i_mb_top_xy = top;
722 h->mb.i_mb_top_y = top_y;
723 h->mb.i_mb_type_top = h->mb.type[h->mb.i_mb_top_xy];
724 if( top >= h->sh.i_first_mb )
726 h->mb.i_neighbour |= MB_TOP;
728 if( !h->param.b_constrained_intra || IS_INTRA( h->mb.i_mb_type_top ) )
729 h->mb.i_neighbour_intra |= MB_TOP;
731 /* We only need to prefetch the top blocks because the left was just written
732 * to as part of the previous cache_save. Since most target CPUs use write-allocate
733 * caches, left blocks are near-guaranteed to be in L1 cache. Top--not so much. */
734 x264_prefetch( &h->mb.cbp[top] );
735 x264_prefetch( h->mb.intra4x4_pred_mode[top] );
736 x264_prefetch( &h->mb.non_zero_count[top][12] );
737 /* These aren't always allocated, but prefetching an invalid address can't hurt. */
738 x264_prefetch( &h->mb.mb_transform_size[top] );
739 x264_prefetch( &h->mb.skipbp[top] );
743 if( mb_x > 0 && topleft_y >= 0 )
745 h->mb.i_neighbour_frame |= MB_TOPLEFT;
746 h->mb.i_mb_topleft_xy = h->mb.i_mb_stride*topleft_y + mb_x - 1;
747 h->mb.i_mb_topleft_y = topleft_y;
748 h->mb.i_mb_type_topleft = h->mb.type[h->mb.i_mb_topleft_xy];
749 if( h->mb.i_mb_topleft_xy >= h->sh.i_first_mb )
751 h->mb.i_neighbour |= MB_TOPLEFT;
753 if( !h->param.b_constrained_intra || IS_INTRA( h->mb.i_mb_type_topleft ) )
754 h->mb.i_neighbour_intra |= MB_TOPLEFT;
758 if( mb_x < h->mb.i_mb_width - 1 && topright_y >= 0 )
760 h->mb.i_neighbour_frame |= MB_TOPRIGHT;
761 h->mb.i_mb_topright_xy = h->mb.i_mb_stride*topright_y + mb_x + 1;
762 h->mb.i_mb_topright_y = topright_y;
763 h->mb.i_mb_type_topright = h->mb.type[h->mb.i_mb_topright_xy];
764 if( h->mb.i_mb_topright_xy >= h->sh.i_first_mb )
766 h->mb.i_neighbour |= MB_TOPRIGHT;
768 if( !h->param.b_constrained_intra || IS_INTRA( h->mb.i_mb_type_topright ) )
769 h->mb.i_neighbour_intra |= MB_TOPRIGHT;
775 void x264_macroblock_cache_load( x264_t *h, int mb_x, int mb_y )
777 x264_macroblock_cache_load_neighbours( h, mb_x, mb_y );
779 int *left = h->mb.i_mb_left_xy;
780 int top = h->mb.i_mb_top_xy;
781 int top_y = h->mb.i_mb_top_y;
782 int s8x8 = h->mb.i_b8_stride;
783 int s4x4 = h->mb.i_b4_stride;
784 int top_8x8 = (2*top_y+1) * s8x8 + 2*mb_x;
785 int top_4x4 = (4*top_y+3) * s4x4 + 4*mb_x;
786 int lists = (1 << h->sh.i_type) & 3;
788 /* GCC pessimizes direct loads from heap-allocated arrays due to aliasing. */
789 /* By only dereferencing them once, we avoid this issue. */
790 int8_t (*i4x4)[8] = h->mb.intra4x4_pred_mode;
791 uint8_t (*nnz)[24] = h->mb.non_zero_count;
792 int16_t *cbp = h->mb.cbp;
794 x264_left_table_t *left_index_table = h->mb.left_index_table;
797 if( h->mb.i_neighbour & MB_TOP )
799 h->mb.cache.i_cbp_top = cbp[top];
801 CP32( &h->mb.cache.intra4x4_pred_mode[x264_scan8[0] - 8], &i4x4[top][0] );
803 /* load non_zero_count */
804 CP32( &h->mb.cache.non_zero_count[x264_scan8[0] - 8], &nnz[top][12] );
805 /* shift because x264_scan8[16] is misaligned */
806 M32( &h->mb.cache.non_zero_count[x264_scan8[16+0] - 9] ) = M16( &nnz[top][18] ) << 8;
807 M32( &h->mb.cache.non_zero_count[x264_scan8[16+4] - 9] ) = M16( &nnz[top][22] ) << 8;
809 /* Finish the prefetching */
810 for( int l = 0; l < lists; l++ )
812 x264_prefetch( &h->mb.mv[l][top_4x4-1] );
813 /* Top right being not in the same cacheline as top left will happen
814 * once every 4 MBs, so one extra prefetch is worthwhile */
815 x264_prefetch( &h->mb.mv[l][top_4x4+4] );
816 x264_prefetch( &h->mb.ref[l][top_8x8-1] );
817 x264_prefetch( &h->mb.mvd[l][top] );
822 h->mb.cache.i_cbp_top = -1;
825 M32( &h->mb.cache.intra4x4_pred_mode[x264_scan8[0] - 8] ) = 0xFFFFFFFFU;
827 /* load non_zero_count */
828 M32( &h->mb.cache.non_zero_count[x264_scan8[ 0] - 8] ) = 0x80808080U;
829 M32( &h->mb.cache.non_zero_count[x264_scan8[16+0] - 9] ) = 0x80808080U;
830 M32( &h->mb.cache.non_zero_count[x264_scan8[16+4] - 9] ) = 0x80808080U;
833 if( h->mb.i_neighbour & MB_LEFT )
835 const int16_t top_luma = (cbp[left[0]] >> (left_index_table->mv[0]&(~1))) & 2;
836 const int16_t bot_luma = (cbp[left[1]] >> (left_index_table->mv[2]&(~1))) & 2;
837 h->mb.cache.i_cbp_left = (cbp[left[0]] & 0xfff0) | (bot_luma<<2) | top_luma;
840 h->mb.cache.intra4x4_pred_mode[x264_scan8[0 ] - 1] = i4x4[left[0]][left_index_table->intra[0]];
841 h->mb.cache.intra4x4_pred_mode[x264_scan8[2 ] - 1] = i4x4[left[0]][left_index_table->intra[1]];
842 h->mb.cache.intra4x4_pred_mode[x264_scan8[8 ] - 1] = i4x4[left[1]][left_index_table->intra[2]];
843 h->mb.cache.intra4x4_pred_mode[x264_scan8[10] - 1] = i4x4[left[1]][left_index_table->intra[3]];
845 /* load non_zero_count */
846 h->mb.cache.non_zero_count[x264_scan8[0 ] - 1] = nnz[left[0]][left_index_table->nnz[0]];
847 h->mb.cache.non_zero_count[x264_scan8[2 ] - 1] = nnz[left[0]][left_index_table->nnz[1]];
848 h->mb.cache.non_zero_count[x264_scan8[8 ] - 1] = nnz[left[1]][left_index_table->nnz[2]];
849 h->mb.cache.non_zero_count[x264_scan8[10] - 1] = nnz[left[1]][left_index_table->nnz[3]];
851 h->mb.cache.non_zero_count[x264_scan8[16+0] - 1] = nnz[left[0]][left_index_table->nnz_chroma[0]];
852 h->mb.cache.non_zero_count[x264_scan8[16+2] - 1] = nnz[left[1]][left_index_table->nnz_chroma[1]];
854 h->mb.cache.non_zero_count[x264_scan8[16+4+0] - 1] = nnz[left[0]][left_index_table->nnz_chroma[2]];
855 h->mb.cache.non_zero_count[x264_scan8[16+4+2] - 1] = nnz[left[1]][left_index_table->nnz_chroma[3]];
859 h->mb.cache.i_cbp_left = -1;
861 h->mb.cache.intra4x4_pred_mode[x264_scan8[0 ] - 1] =
862 h->mb.cache.intra4x4_pred_mode[x264_scan8[2 ] - 1] =
863 h->mb.cache.intra4x4_pred_mode[x264_scan8[8 ] - 1] =
864 h->mb.cache.intra4x4_pred_mode[x264_scan8[10] - 1] = -1;
866 /* load non_zero_count */
867 h->mb.cache.non_zero_count[x264_scan8[0 ] - 1] =
868 h->mb.cache.non_zero_count[x264_scan8[2 ] - 1] =
869 h->mb.cache.non_zero_count[x264_scan8[8 ] - 1] =
870 h->mb.cache.non_zero_count[x264_scan8[10] - 1] =
871 h->mb.cache.non_zero_count[x264_scan8[16+0] - 1] =
872 h->mb.cache.non_zero_count[x264_scan8[16+2] - 1] =
873 h->mb.cache.non_zero_count[x264_scan8[16+4+0] - 1] =
874 h->mb.cache.non_zero_count[x264_scan8[16+4+2] - 1] = 0x80;
877 if( h->pps->b_transform_8x8_mode )
879 h->mb.cache.i_neighbour_transform_size =
880 ( (h->mb.i_neighbour & MB_LEFT) && h->mb.mb_transform_size[left[0]] )
881 + ( (h->mb.i_neighbour & MB_TOP) && h->mb.mb_transform_size[top] );
884 if( h->param.b_interlaced )
886 h->mb.pic.i_fref[0] = h->i_ref[0] << h->mb.b_interlaced;
887 h->mb.pic.i_fref[1] = h->i_ref[1] << h->mb.b_interlaced;
890 if( !h->param.b_interlaced )
892 x264_copy_column8( h->mb.pic.p_fdec[0]-1+ 4*FDEC_STRIDE, h->mb.pic.p_fdec[0]+15+ 4*FDEC_STRIDE );
893 x264_copy_column8( h->mb.pic.p_fdec[0]-1+12*FDEC_STRIDE, h->mb.pic.p_fdec[0]+15+12*FDEC_STRIDE );
894 x264_copy_column8( h->mb.pic.p_fdec[1]-1+ 4*FDEC_STRIDE, h->mb.pic.p_fdec[1]+ 7+ 4*FDEC_STRIDE );
895 x264_copy_column8( h->mb.pic.p_fdec[2]-1+ 4*FDEC_STRIDE, h->mb.pic.p_fdec[2]+ 7+ 4*FDEC_STRIDE );
896 x264_macroblock_load_pic_pointers( h, mb_x, mb_y, 0, 0 );
897 x264_macroblock_load_pic_pointers( h, mb_x, mb_y, 1, 0 );
901 x264_macroblock_load_pic_pointers( h, mb_x, mb_y, 0, 1 );
902 x264_macroblock_load_pic_pointers( h, mb_x, mb_y, 1, 1 );
905 if( h->fdec->integral )
907 int offset = 16 * (mb_x + mb_y * h->fdec->i_stride[0]);
908 for( int list = 0; list < 2; list++ )
909 for( int i = 0; i < h->mb.pic.i_fref[list]; i++ )
910 h->mb.pic.p_integral[list][i] = &h->fref[list][i]->integral[offset];
913 x264_prefetch_fenc( h, h->fenc, mb_x, mb_y );
915 /* load ref/mv/mvd */
916 for( int l = 0; l < lists; l++ )
918 int16_t (*mv)[2] = h->mb.mv[l];
919 int8_t *ref = h->mb.ref[l];
921 int i8 = x264_scan8[0] - 1 - 1*8;
922 if( h->mb.i_neighbour & MB_TOPLEFT )
924 int ir = 2*(s8x8*h->mb.i_mb_topleft_y + mb_x-1)+1+s8x8;
925 int iv = 4*(s4x4*h->mb.i_mb_topleft_y + mb_x-1)+3+3*s4x4;
926 if( h->mb.topleft_partition )
928 /* Take motion vector from the middle of macroblock instead of
929 * the bottom right as usual. */
933 h->mb.cache.ref[l][i8] = ref[ir];
934 CP32( h->mb.cache.mv[l][i8], mv[iv] );
938 h->mb.cache.ref[l][i8] = -2;
939 M32( h->mb.cache.mv[l][i8] ) = 0;
942 i8 = x264_scan8[0] - 8;
943 if( h->mb.i_neighbour & MB_TOP )
945 h->mb.cache.ref[l][i8+0] =
946 h->mb.cache.ref[l][i8+1] = ref[top_8x8 + 0];
947 h->mb.cache.ref[l][i8+2] =
948 h->mb.cache.ref[l][i8+3] = ref[top_8x8 + 1];
949 CP128( h->mb.cache.mv[l][i8], mv[top_4x4] );
953 M128( h->mb.cache.mv[l][i8] ) = M128_ZERO;
954 M32( &h->mb.cache.ref[l][i8] ) = (uint8_t)(-2) * 0x01010101U;
957 i8 = x264_scan8[0] + 4 - 1*8;
958 if( h->mb.i_neighbour & MB_TOPRIGHT )
960 h->mb.cache.ref[l][i8] = ref[2*(s8x8*h->mb.i_mb_topright_y + (mb_x+1))+s8x8];
961 CP32( h->mb.cache.mv[l][i8], mv[4*(s4x4*h->mb.i_mb_topright_y + (mb_x+1))+3*s4x4] );
964 h->mb.cache.ref[l][i8] = -2;
966 i8 = x264_scan8[0] - 1;
967 if( h->mb.i_neighbour & MB_LEFT )
969 h->mb.cache.ref[l][i8+0*8] = ref[h->mb.left_b8[0] + 1 + s8x8*left_index_table->ref[0]];
970 h->mb.cache.ref[l][i8+1*8] = ref[h->mb.left_b8[0] + 1 + s8x8*left_index_table->ref[1]];
971 h->mb.cache.ref[l][i8+2*8] = ref[h->mb.left_b8[1] + 1 + s8x8*left_index_table->ref[2]];
972 h->mb.cache.ref[l][i8+3*8] = ref[h->mb.left_b8[1] + 1 + s8x8*left_index_table->ref[3]];
974 CP32( h->mb.cache.mv[l][i8+0*8], mv[h->mb.left_b4[0] + 3 + s4x4*left_index_table->mv[0]] );
975 CP32( h->mb.cache.mv[l][i8+1*8], mv[h->mb.left_b4[0] + 3 + s4x4*left_index_table->mv[1]] );
976 CP32( h->mb.cache.mv[l][i8+2*8], mv[h->mb.left_b4[1] + 3 + s4x4*left_index_table->mv[2]] );
977 CP32( h->mb.cache.mv[l][i8+3*8], mv[h->mb.left_b4[1] + 3 + s4x4*left_index_table->mv[3]] );
981 for( int i = 0; i < 4; i++ )
983 h->mb.cache.ref[l][i8+i*8] = -2;
984 M32( h->mb.cache.mv[l][i8+i*8] ) = 0;
988 /* Extra logic for top right mv in mbaff.
994 * If the top right of the 4x4 partitions labeled a, b and c in the
995 * above diagram do not exist, but the entries d, e and f exist (in
996 * the macroblock to the left) then use those instead.
998 if( h->sh.b_mbaff && (h->mb.i_neighbour & MB_LEFT) )
1000 if( h->mb.b_interlaced && !h->mb.field[h->mb.i_mb_xy-1] )
1002 h->mb.cache.topright_ref[l][0] = ref[h->mb.left_b8[0] + 1 + s8x8*0];
1003 h->mb.cache.topright_ref[l][1] = ref[h->mb.left_b8[0] + 1 + s8x8*1];
1004 h->mb.cache.topright_ref[l][2] = ref[h->mb.left_b8[1] + 1 + s8x8*0];
1005 CP32( h->mb.cache.topright_mv[l][0], mv[h->mb.left_b4[0] + 3 + s4x4*(left_index_table->mv[0]+1)] );
1006 CP32( h->mb.cache.topright_mv[l][1], mv[h->mb.left_b4[0] + 3 + s4x4*(left_index_table->mv[1]+1)] );
1007 CP32( h->mb.cache.topright_mv[l][2], mv[h->mb.left_b4[1] + 3 + s4x4*(left_index_table->mv[2]+1)] );
1009 else if( !h->mb.b_interlaced && h->mb.field[h->mb.i_mb_xy-1] )
1011 // Looking at the bottom field so always take the bottom macroblock of the pair.
1012 h->mb.cache.topright_ref[l][0] = ref[h->mb.left_b8[0] + 1 + s8x8*2 + s8x8*left_index_table->ref[0]];
1013 h->mb.cache.topright_ref[l][1] = ref[h->mb.left_b8[0] + 1 + s8x8*2 + s8x8*left_index_table->ref[0]];
1014 h->mb.cache.topright_ref[l][2] = ref[h->mb.left_b8[0] + 1 + s8x8*2 + s8x8*left_index_table->ref[2]];
1015 CP32( h->mb.cache.topright_mv[l][0], mv[h->mb.left_b4[0] + 3 + s4x4*4 + s4x4*left_index_table->mv[0]] );
1016 CP32( h->mb.cache.topright_mv[l][1], mv[h->mb.left_b4[0] + 3 + s4x4*4 + s4x4*left_index_table->mv[1]] );
1017 CP32( h->mb.cache.topright_mv[l][2], mv[h->mb.left_b4[0] + 3 + s4x4*4 + s4x4*left_index_table->mv[2]] );
1021 if( h->param.b_cabac )
1023 uint8_t (*mvd)[8][2] = h->mb.mvd[l];
1024 if( h->mb.i_neighbour & MB_TOP )
1025 CP64( h->mb.cache.mvd[l][x264_scan8[0] - 8], mvd[top][0] );
1027 M64( h->mb.cache.mvd[l][x264_scan8[0] - 8] ) = 0;
1029 if( h->mb.cache.ref[l][x264_scan8[0]-1] >= 0 )
1031 CP16( h->mb.cache.mvd[l][x264_scan8[0 ] - 1], mvd[left[0]][left_index_table->intra[0]] );
1032 CP16( h->mb.cache.mvd[l][x264_scan8[2 ] - 1], mvd[left[0]][left_index_table->intra[1]] );
1036 M16( h->mb.cache.mvd[l][x264_scan8[0]-1+0*8] ) = 0;
1037 M16( h->mb.cache.mvd[l][x264_scan8[0]-1+1*8] ) = 0;
1039 if( h->mb.cache.ref[l][x264_scan8[0]-1+2*8] >=0 )
1041 CP16( h->mb.cache.mvd[l][x264_scan8[8 ] - 1], mvd[left[1]][left_index_table->intra[2]] );
1042 CP16( h->mb.cache.mvd[l][x264_scan8[10] - 1], mvd[left[1]][left_index_table->intra[3]] );
1046 M16( h->mb.cache.mvd[l][x264_scan8[0]-1+2*8] ) = 0;
1047 M16( h->mb.cache.mvd[l][x264_scan8[0]-1+3*8] ) = 0;
1051 /* If motion vectors are cached from frame macroblocks but this
1052 * macroblock is a field macroblock then the motion vector must be
1053 * halved. Similarly, motion vectors from field macroblocks are doubled. */
1057 if( FIELD_DIFFERENT(h->mb.i_mb_topleft_xy) )\
1058 MAP_F2F(mv, ref, x264_scan8[0] - 1 - 1*8)\
1059 if( FIELD_DIFFERENT(top) )\
1061 MAP_F2F(mv, ref, x264_scan8[0] + 0 - 1*8)\
1062 MAP_F2F(mv, ref, x264_scan8[0] + 1 - 1*8)\
1063 MAP_F2F(mv, ref, x264_scan8[0] + 2 - 1*8)\
1064 MAP_F2F(mv, ref, x264_scan8[0] + 3 - 1*8)\
1066 if( FIELD_DIFFERENT(h->mb.i_mb_topright_xy) )\
1067 MAP_F2F(mv, ref, x264_scan8[0] + 4 - 1*8)\
1068 if( FIELD_DIFFERENT(left[0]) )\
1070 MAP_F2F(mv, ref, x264_scan8[0] - 1 + 0*8)\
1071 MAP_F2F(mv, ref, x264_scan8[0] - 1 + 1*8)\
1072 MAP_F2F(mv, ref, x264_scan8[0] - 1 + 2*8)\
1073 MAP_F2F(mv, ref, x264_scan8[0] - 1 + 3*8)\
1074 MAP_F2F(topright_mv, topright_ref, 0)\
1075 MAP_F2F(topright_mv, topright_ref, 1)\
1076 MAP_F2F(topright_mv, topright_ref, 2)\
1079 if( h->mb.b_interlaced )
1081 #define FIELD_DIFFERENT(macroblock) (macroblock >= 0 && !h->mb.field[macroblock])
1082 #define MAP_F2F(varmv, varref, index)\
1083 if( h->mb.cache.varref[l][index] >= 0 )\
1085 h->mb.cache.varref[l][index] <<= 1;\
1086 h->mb.cache.varmv[l][index][1] /= 2;\
1087 h->mb.cache.mvd[l][index][1] >>= 1;\
1091 #undef FIELD_DIFFERENT
1095 #define FIELD_DIFFERENT(macroblock) (macroblock >= 0 && h->mb.field[macroblock])
1096 #define MAP_F2F(varmv, varref, index)\
1097 if( h->mb.cache.varref[l][index] >= 0 )\
1099 h->mb.cache.varref[l][index] >>= 1;\
1100 h->mb.cache.varmv[l][index][1] <<= 1;\
1101 h->mb.cache.mvd[l][index][1] <<= 1;\
1105 #undef FIELD_DIFFERENT
1110 if( h->sh.b_mbaff && mb_x == 0 && !(mb_y&1) && mb_y > 0 )
1111 h->mb.field_decoding_flag = h->mb.field[h->mb.i_mb_xy - h->mb.i_mb_stride];
1113 /* Check whether skip here would cause decoder to predict interlace mode incorrectly.
1114 * FIXME: It might be better to change the interlace type rather than forcing a skip to be non-skip. */
1115 h->mb.b_allow_skip = 1;
1118 if( h->mb.b_interlaced != h->mb.field_decoding_flag &&
1119 h->mb.i_mb_prev_xy >= 0 && IS_SKIP(h->mb.type[h->mb.i_mb_prev_xy]) )
1120 h->mb.b_allow_skip = 0;
1121 if( (mb_y&1) && IS_SKIP(h->mb.type[h->mb.i_mb_xy - h->mb.i_mb_stride]) )
1123 if( h->mb.i_neighbour & MB_LEFT )
1125 if( h->mb.field[h->mb.i_mb_xy - 1] != h->mb.b_interlaced )
1126 h->mb.b_allow_skip = 0;
1128 else if( h->mb.i_neighbour & MB_TOP )
1130 if( h->mb.field[h->mb.i_mb_top_xy] != h->mb.b_interlaced )
1131 h->mb.b_allow_skip = 0;
1133 else // Frame mb pair is predicted
1135 if( h->mb.b_interlaced )
1136 h->mb.b_allow_skip = 0;
1141 if( h->param.b_cabac )
1143 int left_xy, top_xy;
1146 /* Neighbours here are calculated based on field_decoding_flag */
1147 int mb_xy = mb_x + (mb_y&~1)*h->mb.i_mb_stride;
1148 left_xy = mb_xy - 1;
1149 if( (mb_y&1) && mb_x > 0 && h->mb.field_decoding_flag == h->mb.field[left_xy] )
1150 left_xy += h->mb.i_mb_stride;
1151 if( h->mb.field_decoding_flag )
1153 top_xy = mb_xy - h->mb.i_mb_stride;
1154 if( !(mb_y&1) && top_xy >= 0 && h->mb.slice_table[top_xy] == h->sh.i_first_mb && h->mb.field[top_xy] )
1155 top_xy -= h->mb.i_mb_stride;
1158 top_xy = mb_x + (mb_y-1)*h->mb.i_mb_stride;
1162 left_xy = h->mb.i_mb_left_xy[0];
1163 top_xy = h->mb.i_mb_top_xy;
1166 h->mb.cache.i_neighbour_skip = (mb_x > 0 && h->mb.slice_table[left_xy] == h->sh.i_first_mb && !IS_SKIP( h->mb.type[left_xy] ))
1167 + (top_xy >= 0 && h->mb.slice_table[top_xy] == h->sh.i_first_mb && !IS_SKIP( h->mb.type[top_xy] ));
1171 if( h->sh.i_type == SLICE_TYPE_B )
1173 h->mb.bipred_weight = h->mb.bipred_weight_buf[h->mb.b_interlaced][h->mb.b_interlaced&(mb_y&1)];
1174 h->mb.dist_scale_factor = h->mb.dist_scale_factor_buf[h->mb.b_interlaced][h->mb.b_interlaced&(mb_y&1)];
1175 if( h->param.b_cabac )
1178 x264_macroblock_cache_skip( h, 0, 0, 4, 4, 0 );
1179 skipbp = (h->mb.i_neighbour & MB_LEFT) ? h->mb.skipbp[left[0]] : 0;
1180 h->mb.cache.skip[x264_scan8[0] - 1] = (skipbp >> (1+(left_index_table->mv[0]&~1))) & 1;
1181 skipbp = (h->mb.i_neighbour & MB_LEFT) ? h->mb.skipbp[left[1]] : 0;
1182 h->mb.cache.skip[x264_scan8[8] - 1] = (skipbp >> (1+(left_index_table->mv[2]&~1))) & 1;
1183 skipbp = (h->mb.i_neighbour & MB_TOP) ? h->mb.skipbp[top] : 0;
1184 h->mb.cache.skip[x264_scan8[0] - 8] = skipbp & 0x4;
1185 h->mb.cache.skip[x264_scan8[4] - 8] = skipbp & 0x8;
1189 if( h->sh.i_type == SLICE_TYPE_P )
1190 x264_mb_predict_mv_pskip( h, h->mb.cache.pskip_mv );
1192 h->mb.i_neighbour4[0] =
1193 h->mb.i_neighbour8[0] = (h->mb.i_neighbour_intra & (MB_TOP|MB_LEFT|MB_TOPLEFT))
1194 | ((h->mb.i_neighbour_intra & MB_TOP) ? MB_TOPRIGHT : 0);
1195 h->mb.i_neighbour4[4] =
1196 h->mb.i_neighbour4[1] = MB_LEFT | ((h->mb.i_neighbour_intra & MB_TOP) ? (MB_TOP|MB_TOPLEFT|MB_TOPRIGHT) : 0);
1197 h->mb.i_neighbour4[2] =
1198 h->mb.i_neighbour4[8] =
1199 h->mb.i_neighbour4[10] =
1200 h->mb.i_neighbour8[2] = MB_TOP|MB_TOPRIGHT | ((h->mb.i_neighbour_intra & MB_LEFT) ? (MB_LEFT|MB_TOPLEFT) : 0);
1201 h->mb.i_neighbour4[5] =
1202 h->mb.i_neighbour8[1] = MB_LEFT | (h->mb.i_neighbour_intra & MB_TOPRIGHT)
1203 | ((h->mb.i_neighbour_intra & MB_TOP) ? MB_TOP|MB_TOPLEFT : 0);
1206 void x264_macroblock_cache_load_neighbours_deblock( x264_t *h, int mb_x, int mb_y )
1208 int deblock_on_slice_edges = h->sh.i_disable_deblocking_filter_idc != 2;
1210 h->mb.i_neighbour = 0;
1211 h->mb.i_mb_xy = mb_y * h->mb.i_mb_stride + mb_x;
1212 h->mb.b_interlaced = h->param.b_interlaced && h->mb.field[h->mb.i_mb_xy];
1213 h->mb.i_mb_top_y = mb_y - (1 << h->mb.b_interlaced);
1214 h->mb.i_mb_top_xy = mb_x + h->mb.i_mb_stride*h->mb.i_mb_top_y;
1215 h->mb.i_mb_left_xy[1] =
1216 h->mb.i_mb_left_xy[0] = h->mb.i_mb_xy - 1;
1221 if( mb_x && h->mb.field[h->mb.i_mb_xy - 1] != h->mb.b_interlaced )
1222 h->mb.i_mb_left_xy[0] -= h->mb.i_mb_stride;
1226 if( h->mb.i_mb_top_xy >= 0 && h->mb.b_interlaced && !h->mb.field[h->mb.i_mb_top_xy] )
1228 h->mb.i_mb_top_xy += h->mb.i_mb_stride;
1231 if( mb_x && h->mb.field[h->mb.i_mb_xy - 1] != h->mb.b_interlaced )
1232 h->mb.i_mb_left_xy[1] += h->mb.i_mb_stride;
1236 if( mb_x > 0 && (deblock_on_slice_edges ||
1237 h->mb.slice_table[h->mb.i_mb_left_xy[0]] == h->mb.slice_table[h->mb.i_mb_xy]) )
1238 h->mb.i_neighbour |= MB_LEFT;
1239 if( mb_y > h->mb.b_interlaced && (deblock_on_slice_edges
1240 || h->mb.slice_table[h->mb.i_mb_top_xy] == h->mb.slice_table[h->mb.i_mb_xy]) )
1241 h->mb.i_neighbour |= MB_TOP;
1244 void x264_macroblock_deblock_strength( x264_t *h )
1246 uint8_t (*bs)[8][4] = h->deblock_strength[h->mb.i_mb_y&1][h->mb.i_mb_x];
1247 if( IS_INTRA( h->mb.type[h->mb.i_mb_xy] ) )
1249 memset( bs[0], 3, 4*4*sizeof(uint8_t) );
1250 memset( bs[1], 3, 4*4*sizeof(uint8_t) );
1251 if( !h->sh.b_mbaff ) return;
1254 /* If we have multiple slices and we're deblocking on slice edges, we
1255 * have to reload neighbour data. */
1256 if( h->sh.b_mbaff || (h->sh.i_first_mb && h->sh.i_disable_deblocking_filter_idc != 2) )
1258 int old_neighbour = h->mb.i_neighbour;
1259 int mb_x = h->mb.i_mb_x;
1260 int mb_y = h->mb.i_mb_y;
1261 x264_macroblock_cache_load_neighbours_deblock( h, mb_x, mb_y );
1262 int new_neighbour = h->mb.i_neighbour;
1263 h->mb.i_neighbour &= ~old_neighbour;
1264 if( h->mb.i_neighbour )
1266 int top_y = h->mb.i_mb_top_y;
1267 int top_8x8 = (2*top_y+1) * h->mb.i_b8_stride + 2*mb_x;
1268 int top_4x4 = (4*top_y+3) * h->mb.i_b4_stride + 4*mb_x;
1269 int s8x8 = h->mb.i_b8_stride;
1270 int s4x4 = h->mb.i_b4_stride;
1272 uint8_t (*nnz)[24] = h->mb.non_zero_count;
1273 x264_left_table_t *left_index_table = h->mb.left_index_table;
1275 if( h->mb.i_neighbour & MB_TOP )
1276 CP32( &h->mb.cache.non_zero_count[x264_scan8[0] - 8], &nnz[h->mb.i_mb_top_xy][12] );
1278 if( h->mb.i_neighbour & MB_LEFT )
1280 int *left = h->mb.i_mb_left_xy;
1281 h->mb.cache.non_zero_count[x264_scan8[0 ] - 1] = nnz[left[0]][left_index_table->nnz[0]];
1282 h->mb.cache.non_zero_count[x264_scan8[2 ] - 1] = nnz[left[0]][left_index_table->nnz[1]];
1283 h->mb.cache.non_zero_count[x264_scan8[8 ] - 1] = nnz[left[1]][left_index_table->nnz[2]];
1284 h->mb.cache.non_zero_count[x264_scan8[10] - 1] = nnz[left[1]][left_index_table->nnz[3]];
1287 for( int l = 0; l <= (h->sh.i_type == SLICE_TYPE_B); l++ )
1289 int16_t (*mv)[2] = h->mb.mv[l];
1290 int8_t *ref = h->mb.ref[l];
1292 int i8 = x264_scan8[0] - 8;
1293 if( h->mb.i_neighbour & MB_TOP )
1295 h->mb.cache.ref[l][i8+0] =
1296 h->mb.cache.ref[l][i8+1] = ref[top_8x8 + 0];
1297 h->mb.cache.ref[l][i8+2] =
1298 h->mb.cache.ref[l][i8+3] = ref[top_8x8 + 1];
1299 CP128( h->mb.cache.mv[l][i8], mv[top_4x4] );
1302 i8 = x264_scan8[0] - 1;
1303 if( h->mb.i_neighbour & MB_LEFT )
1305 h->mb.cache.ref[l][i8+0*8] =
1306 h->mb.cache.ref[l][i8+1*8] = ref[h->mb.left_b8[0] + 1 + s8x8*left_index_table->ref[0]];
1307 h->mb.cache.ref[l][i8+2*8] =
1308 h->mb.cache.ref[l][i8+3*8] = ref[h->mb.left_b8[1] + 1 + s8x8*left_index_table->ref[2]];
1310 CP32( h->mb.cache.mv[l][i8+0*8], mv[h->mb.left_b4[0] + 3 + s4x4*left_index_table->mv[0]] );
1311 CP32( h->mb.cache.mv[l][i8+1*8], mv[h->mb.left_b4[0] + 3 + s4x4*left_index_table->mv[1]] );
1312 CP32( h->mb.cache.mv[l][i8+2*8], mv[h->mb.left_b4[1] + 3 + s4x4*left_index_table->mv[2]] );
1313 CP32( h->mb.cache.mv[l][i8+3*8], mv[h->mb.left_b4[1] + 3 + s4x4*left_index_table->mv[3]] );
1317 h->mb.i_neighbour = new_neighbour;
1320 if( h->param.analyse.i_weighted_pred == X264_WEIGHTP_SMART && h->sh.i_type == SLICE_TYPE_P )
1322 /* Handle reference frame duplicates */
1323 int i8 = x264_scan8[0] - 8;
1324 h->mb.cache.ref[0][i8+0] =
1325 h->mb.cache.ref[0][i8+1] = deblock_ref_table(h->mb.cache.ref[0][i8+0]);
1326 h->mb.cache.ref[0][i8+2] =
1327 h->mb.cache.ref[0][i8+3] = deblock_ref_table(h->mb.cache.ref[0][i8+2]);
1329 i8 = x264_scan8[0] - 1;
1330 h->mb.cache.ref[0][i8+0*8] =
1331 h->mb.cache.ref[0][i8+1*8] = deblock_ref_table(h->mb.cache.ref[0][i8+0*8]);
1332 h->mb.cache.ref[0][i8+2*8] =
1333 h->mb.cache.ref[0][i8+3*8] = deblock_ref_table(h->mb.cache.ref[0][i8+2*8]);
1335 int ref0 = deblock_ref_table(h->mb.cache.ref[0][x264_scan8[ 0]]);
1336 int ref1 = deblock_ref_table(h->mb.cache.ref[0][x264_scan8[ 4]]);
1337 int ref2 = deblock_ref_table(h->mb.cache.ref[0][x264_scan8[ 8]]);
1338 int ref3 = deblock_ref_table(h->mb.cache.ref[0][x264_scan8[12]]);
1339 uint32_t reftop = pack16to32( (uint8_t)ref0, (uint8_t)ref1 ) * 0x0101;
1340 uint32_t refbot = pack16to32( (uint8_t)ref2, (uint8_t)ref3 ) * 0x0101;
1342 M32( &h->mb.cache.ref[0][x264_scan8[0]+8*0] ) = reftop;
1343 M32( &h->mb.cache.ref[0][x264_scan8[0]+8*1] ) = reftop;
1344 M32( &h->mb.cache.ref[0][x264_scan8[0]+8*2] ) = refbot;
1345 M32( &h->mb.cache.ref[0][x264_scan8[0]+8*3] ) = refbot;
1348 /* Munge NNZ for cavlc + 8x8dct */
1349 if( !h->param.b_cabac && h->pps->b_transform_8x8_mode )
1351 uint8_t (*nnz)[24] = h->mb.non_zero_count;
1352 int top = h->mb.i_mb_top_xy;
1353 int *left = h->mb.i_mb_left_xy;
1355 if( (h->mb.i_neighbour & MB_TOP) && h->mb.mb_transform_size[top] )
1357 int i8 = x264_scan8[0] - 8;
1358 int nnz_top0 = M16( &nnz[top][8] ) | M16( &nnz[top][12] );
1359 int nnz_top1 = M16( &nnz[top][10] ) | M16( &nnz[top][14] );
1360 M16( &h->mb.cache.non_zero_count[i8+0] ) = nnz_top0 ? 0x0101 : 0;
1361 M16( &h->mb.cache.non_zero_count[i8+2] ) = nnz_top1 ? 0x0101 : 0;
1364 if( h->mb.i_neighbour & MB_LEFT )
1366 int i8 = x264_scan8[0] - 1;
1367 if( h->mb.mb_transform_size[left[0]] )
1369 int nnz_left0 = M16( &nnz[left[0]][2] ) | M16( &nnz[left[0]][6] );
1370 h->mb.cache.non_zero_count[i8+8*0] = !!nnz_left0;
1371 h->mb.cache.non_zero_count[i8+8*1] = !!nnz_left0;
1373 if( h->mb.mb_transform_size[left[1]] )
1375 int nnz_left1 = M16( &nnz[left[1]][10] ) | M16( &nnz[left[1]][14] );
1376 h->mb.cache.non_zero_count[i8+8*2] = !!nnz_left1;
1377 h->mb.cache.non_zero_count[i8+8*3] = !!nnz_left1;
1381 if( h->mb.mb_transform_size[h->mb.i_mb_xy] )
1383 int nnz0 = M16( &h->mb.cache.non_zero_count[x264_scan8[ 0]] ) | M16( &h->mb.cache.non_zero_count[x264_scan8[ 2]] );
1384 int nnz1 = M16( &h->mb.cache.non_zero_count[x264_scan8[ 4]] ) | M16( &h->mb.cache.non_zero_count[x264_scan8[ 6]] );
1385 int nnz2 = M16( &h->mb.cache.non_zero_count[x264_scan8[ 8]] ) | M16( &h->mb.cache.non_zero_count[x264_scan8[10]] );
1386 int nnz3 = M16( &h->mb.cache.non_zero_count[x264_scan8[12]] ) | M16( &h->mb.cache.non_zero_count[x264_scan8[14]] );
1387 uint32_t nnztop = pack16to32( !!nnz0, !!nnz1 ) * 0x0101;
1388 uint32_t nnzbot = pack16to32( !!nnz2, !!nnz3 ) * 0x0101;
1390 M32( &h->mb.cache.non_zero_count[x264_scan8[0]+8*0] ) = nnztop;
1391 M32( &h->mb.cache.non_zero_count[x264_scan8[0]+8*1] ) = nnztop;
1392 M32( &h->mb.cache.non_zero_count[x264_scan8[0]+8*2] ) = nnzbot;
1393 M32( &h->mb.cache.non_zero_count[x264_scan8[0]+8*3] ) = nnzbot;
1397 int mvy_limit = 4 >> h->mb.b_interlaced;
1398 h->loopf.deblock_strength( h->mb.cache.non_zero_count, h->mb.cache.ref, h->mb.cache.mv,
1399 bs, mvy_limit, h->sh.i_type == SLICE_TYPE_B, h );
1402 static void ALWAYS_INLINE x264_macroblock_store_pic( x264_t *h, int mb_x, int mb_y, int i, int b_mbaff )
1405 int i_stride = h->fdec->i_stride[i];
1406 int i_stride2 = i_stride << (b_mbaff && h->mb.b_interlaced);
1407 int i_pix_offset = (b_mbaff && h->mb.b_interlaced)
1408 ? 16 * mb_x + w * (mb_y&~1) * i_stride + (mb_y&1) * i_stride
1409 : 16 * mb_x + w * mb_y * i_stride;
1411 h->mc.store_interleave_8x8x2( &h->fdec->plane[1][i_pix_offset], i_stride2, h->mb.pic.p_fdec[1], h->mb.pic.p_fdec[2] );
1413 h->mc.copy[PIXEL_16x16]( &h->fdec->plane[0][i_pix_offset], i_stride2, h->mb.pic.p_fdec[0], FDEC_STRIDE, 16 );
1416 static void ALWAYS_INLINE x264_macroblock_backup_intra( x264_t *h, int mb_x, int mb_y, int b_mbaff )
1418 /* In MBAFF we store the last two rows in intra_border_backup[0] and [1].
1419 * For progressive mbs this is the bottom two rows, and for interlaced the
1420 * bottom row of each field. We also store samples needed for the next
1421 * mbpair in intra_border_backup[2]. */
1422 int backup_dst = !b_mbaff ? 0 : (mb_y&1) ? 1 : h->mb.b_interlaced ? 0 : 2;
1423 memcpy( &h->intra_border_backup[backup_dst][0][mb_x*16 ], h->mb.pic.p_fdec[0]+FDEC_STRIDE*15, 16*sizeof(pixel) );
1424 memcpy( &h->intra_border_backup[backup_dst][1][mb_x*16 ], h->mb.pic.p_fdec[1]+FDEC_STRIDE*7, 8*sizeof(pixel) );
1425 memcpy( &h->intra_border_backup[backup_dst][1][mb_x*16+8], h->mb.pic.p_fdec[2]+FDEC_STRIDE*7, 8*sizeof(pixel) );
1430 int backup_src = (h->mb.b_interlaced ? 7 : 14) * FDEC_STRIDE;
1431 backup_dst = h->mb.b_interlaced ? 2 : 0;
1432 memcpy( &h->intra_border_backup[backup_dst][0][mb_x*16 ], h->mb.pic.p_fdec[0]+backup_src, 16*sizeof(pixel) );
1433 backup_src = (h->mb.b_interlaced ? 3 : 6) * FDEC_STRIDE;
1434 memcpy( &h->intra_border_backup[backup_dst][1][mb_x*16 ], h->mb.pic.p_fdec[1]+backup_src, 8*sizeof(pixel) );
1435 memcpy( &h->intra_border_backup[backup_dst][1][mb_x*16+8], h->mb.pic.p_fdec[2]+backup_src, 8*sizeof(pixel) );
1440 /* In progressive we update intra_border_backup in-place, so the topleft neighbor will
1441 * no longer exist there when load_pic_pointers wants it. Move it within p_fdec instead. */
1442 h->mb.pic.p_fdec[0][-FDEC_STRIDE-1] = h->mb.pic.p_fdec[0][-FDEC_STRIDE+15];
1443 h->mb.pic.p_fdec[1][-FDEC_STRIDE-1] = h->mb.pic.p_fdec[1][-FDEC_STRIDE+7];
1444 h->mb.pic.p_fdec[2][-FDEC_STRIDE-1] = h->mb.pic.p_fdec[2][-FDEC_STRIDE+7];
1448 void x264_macroblock_cache_save( x264_t *h )
1450 const int i_mb_xy = h->mb.i_mb_xy;
1451 const int i_mb_type = x264_mb_type_fix[h->mb.i_type];
1452 const int s8x8 = h->mb.i_b8_stride;
1453 const int s4x4 = h->mb.i_b4_stride;
1454 const int i_mb_4x4 = h->mb.i_b4_xy;
1455 const int i_mb_8x8 = h->mb.i_b8_xy;
1457 /* GCC pessimizes direct stores to heap-allocated arrays due to aliasing. */
1458 /* By only dereferencing them once, we avoid this issue. */
1459 int8_t *i4x4 = h->mb.intra4x4_pred_mode[i_mb_xy];
1460 uint8_t *nnz = h->mb.non_zero_count[i_mb_xy];
1464 x264_macroblock_backup_intra( h, h->mb.i_mb_x, h->mb.i_mb_y, 1 );
1465 x264_macroblock_store_pic( h, h->mb.i_mb_x, h->mb.i_mb_y, 0, 1 );
1466 x264_macroblock_store_pic( h, h->mb.i_mb_x, h->mb.i_mb_y, 1, 1 );
1470 x264_macroblock_backup_intra( h, h->mb.i_mb_x, h->mb.i_mb_y, 0 );
1471 x264_macroblock_store_pic( h, h->mb.i_mb_x, h->mb.i_mb_y, 0, 0 );
1472 x264_macroblock_store_pic( h, h->mb.i_mb_x, h->mb.i_mb_y, 1, 0 );
1475 x264_prefetch_fenc( h, h->fdec, h->mb.i_mb_x, h->mb.i_mb_y );
1477 h->mb.type[i_mb_xy] = i_mb_type;
1478 h->mb.slice_table[i_mb_xy] = h->sh.i_first_mb;
1479 h->mb.partition[i_mb_xy] = IS_INTRA( i_mb_type ) ? D_16x16 : h->mb.i_partition;
1480 h->mb.i_mb_prev_xy = i_mb_xy;
1483 if( i_mb_type == I_4x4 )
1485 CP32( &i4x4[0], &h->mb.cache.intra4x4_pred_mode[x264_scan8[10]] );
1486 M32( &i4x4[4] ) = pack8to32( h->mb.cache.intra4x4_pred_mode[x264_scan8[5] ],
1487 h->mb.cache.intra4x4_pred_mode[x264_scan8[7] ],
1488 h->mb.cache.intra4x4_pred_mode[x264_scan8[13] ], 0);
1490 else if( !h->param.b_constrained_intra || IS_INTRA(i_mb_type) )
1491 M64( i4x4 ) = I_PRED_4x4_DC * 0x0101010101010101ULL;
1493 M64( i4x4 ) = (uint8_t)(-1) * 0x0101010101010101ULL;
1496 if( i_mb_type == I_PCM )
1498 h->mb.qp[i_mb_xy] = 0;
1499 h->mb.i_last_dqp = 0;
1500 h->mb.i_cbp_chroma = 2;
1501 h->mb.i_cbp_luma = 0xf;
1502 h->mb.cbp[i_mb_xy] = 0x72f; /* all set */
1503 h->mb.b_transform_8x8 = 0;
1504 for( int i = 0; i < 24; i++ )
1505 h->mb.cache.non_zero_count[x264_scan8[i]] = h->param.b_cabac ? 1 : 16;
1509 if( h->mb.i_type != I_16x16 && h->mb.i_cbp_luma == 0 && h->mb.i_cbp_chroma == 0 )
1510 h->mb.i_qp = h->mb.i_last_qp;
1511 h->mb.qp[i_mb_xy] = h->mb.i_qp;
1512 h->mb.i_last_dqp = h->mb.i_qp - h->mb.i_last_qp;
1513 h->mb.i_last_qp = h->mb.i_qp;
1516 /* save non zero count */
1517 CP32( &nnz[0*4], &h->mb.cache.non_zero_count[x264_scan8[0]+0*8] );
1518 CP32( &nnz[1*4], &h->mb.cache.non_zero_count[x264_scan8[0]+1*8] );
1519 CP32( &nnz[2*4], &h->mb.cache.non_zero_count[x264_scan8[0]+2*8] );
1520 CP32( &nnz[3*4], &h->mb.cache.non_zero_count[x264_scan8[0]+3*8] );
1521 M16( &nnz[16+0*2] ) = M32( &h->mb.cache.non_zero_count[x264_scan8[16+0*2]-1] ) >> 8;
1522 M16( &nnz[16+1*2] ) = M32( &h->mb.cache.non_zero_count[x264_scan8[16+1*2]-1] ) >> 8;
1523 M16( &nnz[16+2*2] ) = M32( &h->mb.cache.non_zero_count[x264_scan8[16+2*2]-1] ) >> 8;
1524 M16( &nnz[16+3*2] ) = M32( &h->mb.cache.non_zero_count[x264_scan8[16+3*2]-1] ) >> 8;
1526 if( h->mb.i_cbp_luma == 0 && h->mb.i_type != I_8x8 )
1527 h->mb.b_transform_8x8 = 0;
1528 h->mb.mb_transform_size[i_mb_xy] = h->mb.b_transform_8x8;
1530 if( h->sh.i_type != SLICE_TYPE_I )
1532 int16_t (*mv0)[2] = &h->mb.mv[0][i_mb_4x4];
1533 int16_t (*mv1)[2] = &h->mb.mv[1][i_mb_4x4];
1534 int8_t *ref0 = &h->mb.ref[0][i_mb_8x8];
1535 int8_t *ref1 = &h->mb.ref[1][i_mb_8x8];
1536 if( !IS_INTRA( i_mb_type ) )
1538 ref0[0+0*s8x8] = h->mb.cache.ref[0][x264_scan8[0]];
1539 ref0[1+0*s8x8] = h->mb.cache.ref[0][x264_scan8[4]];
1540 ref0[0+1*s8x8] = h->mb.cache.ref[0][x264_scan8[8]];
1541 ref0[1+1*s8x8] = h->mb.cache.ref[0][x264_scan8[12]];
1542 CP128( &mv0[0*s4x4], h->mb.cache.mv[0][x264_scan8[0]+8*0] );
1543 CP128( &mv0[1*s4x4], h->mb.cache.mv[0][x264_scan8[0]+8*1] );
1544 CP128( &mv0[2*s4x4], h->mb.cache.mv[0][x264_scan8[0]+8*2] );
1545 CP128( &mv0[3*s4x4], h->mb.cache.mv[0][x264_scan8[0]+8*3] );
1546 if( h->sh.i_type == SLICE_TYPE_B )
1548 ref1[0+0*s8x8] = h->mb.cache.ref[1][x264_scan8[0]];
1549 ref1[1+0*s8x8] = h->mb.cache.ref[1][x264_scan8[4]];
1550 ref1[0+1*s8x8] = h->mb.cache.ref[1][x264_scan8[8]];
1551 ref1[1+1*s8x8] = h->mb.cache.ref[1][x264_scan8[12]];
1552 CP128( &mv1[0*s4x4], h->mb.cache.mv[1][x264_scan8[0]+8*0] );
1553 CP128( &mv1[1*s4x4], h->mb.cache.mv[1][x264_scan8[0]+8*1] );
1554 CP128( &mv1[2*s4x4], h->mb.cache.mv[1][x264_scan8[0]+8*2] );
1555 CP128( &mv1[3*s4x4], h->mb.cache.mv[1][x264_scan8[0]+8*3] );
1560 M16( &ref0[0*s8x8] ) = (uint8_t)(-1) * 0x0101;
1561 M16( &ref0[1*s8x8] ) = (uint8_t)(-1) * 0x0101;
1562 M128( &mv0[0*s4x4] ) = M128_ZERO;
1563 M128( &mv0[1*s4x4] ) = M128_ZERO;
1564 M128( &mv0[2*s4x4] ) = M128_ZERO;
1565 M128( &mv0[3*s4x4] ) = M128_ZERO;
1566 if( h->sh.i_type == SLICE_TYPE_B )
1568 M16( &ref1[0*s8x8] ) = (uint8_t)(-1) * 0x0101;
1569 M16( &ref1[1*s8x8] ) = (uint8_t)(-1) * 0x0101;
1570 M128( &mv1[0*s4x4] ) = M128_ZERO;
1571 M128( &mv1[1*s4x4] ) = M128_ZERO;
1572 M128( &mv1[2*s4x4] ) = M128_ZERO;
1573 M128( &mv1[3*s4x4] ) = M128_ZERO;
1578 if( h->param.b_cabac )
1580 uint8_t (*mvd0)[2] = h->mb.mvd[0][i_mb_xy];
1581 uint8_t (*mvd1)[2] = h->mb.mvd[1][i_mb_xy];
1582 if( IS_INTRA(i_mb_type) && i_mb_type != I_PCM )
1583 h->mb.chroma_pred_mode[i_mb_xy] = x264_mb_pred_mode8x8c_fix[ h->mb.i_chroma_pred_mode ];
1585 h->mb.chroma_pred_mode[i_mb_xy] = I_PRED_CHROMA_DC;
1587 if( (0x3FF30 >> i_mb_type) & 1 ) /* !INTRA && !SKIP && !DIRECT */
1589 CP64( mvd0[0], h->mb.cache.mvd[0][x264_scan8[10]] );
1590 CP16( mvd0[4], h->mb.cache.mvd[0][x264_scan8[5 ]] );
1591 CP16( mvd0[5], h->mb.cache.mvd[0][x264_scan8[7 ]] );
1592 CP16( mvd0[6], h->mb.cache.mvd[0][x264_scan8[13]] );
1593 if( h->sh.i_type == SLICE_TYPE_B )
1595 CP64( mvd1[0], h->mb.cache.mvd[1][x264_scan8[10]] );
1596 CP16( mvd1[4], h->mb.cache.mvd[1][x264_scan8[5 ]] );
1597 CP16( mvd1[5], h->mb.cache.mvd[1][x264_scan8[7 ]] );
1598 CP16( mvd1[6], h->mb.cache.mvd[1][x264_scan8[13]] );
1603 M128( mvd0[0] ) = M128_ZERO;
1604 if( h->sh.i_type == SLICE_TYPE_B )
1605 M128( mvd1[0] ) = M128_ZERO;
1608 if( h->sh.i_type == SLICE_TYPE_B )
1610 if( i_mb_type == B_SKIP || i_mb_type == B_DIRECT )
1611 h->mb.skipbp[i_mb_xy] = 0xf;
1612 else if( i_mb_type == B_8x8 )
1614 int skipbp = ( h->mb.i_sub_partition[0] == D_DIRECT_8x8 ) << 0;
1615 skipbp |= ( h->mb.i_sub_partition[1] == D_DIRECT_8x8 ) << 1;
1616 skipbp |= ( h->mb.i_sub_partition[2] == D_DIRECT_8x8 ) << 2;
1617 skipbp |= ( h->mb.i_sub_partition[3] == D_DIRECT_8x8 ) << 3;
1618 h->mb.skipbp[i_mb_xy] = skipbp;
1621 h->mb.skipbp[i_mb_xy] = 0;
1627 void x264_macroblock_bipred_init( x264_t *h )
1629 for( int mbfield = 0; mbfield <= h->sh.b_mbaff; mbfield++ )
1630 for( int field = 0; field <= h->sh.b_mbaff; field++ )
1631 for( int i_ref0 = 0; i_ref0 < (h->i_ref[0]<<mbfield); i_ref0++ )
1633 x264_frame_t *l0 = h->fref[0][i_ref0>>mbfield];
1634 int poc0 = l0->i_poc + mbfield*l0->i_delta_poc[field^(i_ref0&1)];
1635 for( int i_ref1 = 0; i_ref1 < (h->i_ref[1]<<mbfield); i_ref1++ )
1637 int dist_scale_factor;
1638 x264_frame_t *l1 = h->fref[1][i_ref1>>mbfield];
1639 int cur_poc = h->fdec->i_poc + mbfield*h->fdec->i_delta_poc[field];
1640 int poc1 = l1->i_poc + mbfield*l1->i_delta_poc[field^(i_ref1&1)];
1641 int td = x264_clip3( poc1 - poc0, -128, 127 );
1642 if( td == 0 /* || pic0 is a long-term ref */ )
1643 dist_scale_factor = 256;
1646 int tb = x264_clip3( cur_poc - poc0, -128, 127 );
1647 int tx = (16384 + (abs(td) >> 1)) / td;
1648 dist_scale_factor = x264_clip3( (tb * tx + 32) >> 6, -1024, 1023 );
1651 h->mb.dist_scale_factor_buf[mbfield][field][i_ref0][i_ref1] = dist_scale_factor;
1653 dist_scale_factor >>= 2;
1654 if( h->param.analyse.b_weighted_bipred
1655 && dist_scale_factor >= -64
1656 && dist_scale_factor <= 128 )
1658 h->mb.bipred_weight_buf[mbfield][field][i_ref0][i_ref1] = 64 - dist_scale_factor;
1659 // ssse3 implementation of biweight doesn't support the extrema.
1660 // if we ever generate them, we'll have to drop that optimization.
1661 assert( dist_scale_factor >= -63 && dist_scale_factor <= 127 );
1664 h->mb.bipred_weight_buf[mbfield][field][i_ref0][i_ref1] = 32;