1 /*****************************************************************************
2 * macroblock.c: h264 encoder library
3 *****************************************************************************
4 * Copyright (C) 2003-2008 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.
23 *****************************************************************************/
26 #include "encoder/me.h"
28 static NOINLINE void x264_mb_mc_0xywh( x264_t *h, int x, int y, int width, int height )
30 int i8 = x264_scan8[0]+x+8*y;
31 int i_ref = h->mb.cache.ref[0][i8];
32 int mvx = x264_clip3( h->mb.cache.mv[0][i8][0], h->mb.mv_min[0], h->mb.mv_max[0] ) + 4*4*x;
33 int mvy = x264_clip3( h->mb.cache.mv[0][i8][1], h->mb.mv_min[1], h->mb.mv_max[1] ) + 4*4*y;
35 h->mc.mc_luma( &h->mb.pic.p_fdec[0][4*y*FDEC_STRIDE+4*x], FDEC_STRIDE,
36 h->mb.pic.p_fref[0][i_ref], h->mb.pic.i_stride[0],
37 mvx, mvy, 4*width, 4*height, &h->sh.weight[i_ref][0] );
39 // chroma is offset if MCing from a field of opposite parity
40 if( h->mb.b_interlaced & i_ref )
41 mvy += (h->mb.i_mb_y & 1)*4 - 2;
43 h->mc.mc_chroma( &h->mb.pic.p_fdec[1][2*y*FDEC_STRIDE+2*x], FDEC_STRIDE,
44 h->mb.pic.p_fref[0][i_ref][4], h->mb.pic.i_stride[1],
45 mvx, mvy, 2*width, 2*height );
47 if( h->sh.weight[i_ref][1].weightfn )
48 h->sh.weight[i_ref][1].weightfn[width>>1]( &h->mb.pic.p_fdec[1][2*y*FDEC_STRIDE+2*x], FDEC_STRIDE,
49 &h->mb.pic.p_fdec[1][2*y*FDEC_STRIDE+2*x], FDEC_STRIDE,
50 &h->sh.weight[i_ref][1], height*2 );
52 h->mc.mc_chroma( &h->mb.pic.p_fdec[2][2*y*FDEC_STRIDE+2*x], FDEC_STRIDE,
53 h->mb.pic.p_fref[0][i_ref][5], h->mb.pic.i_stride[2],
54 mvx, mvy, 2*width, 2*height );
56 if( h->sh.weight[i_ref][2].weightfn )
57 h->sh.weight[i_ref][2].weightfn[width>>1]( &h->mb.pic.p_fdec[2][2*y*FDEC_STRIDE+2*x], FDEC_STRIDE,
58 &h->mb.pic.p_fdec[2][2*y*FDEC_STRIDE+2*x], FDEC_STRIDE,
59 &h->sh.weight[i_ref][2],height*2 );
62 static NOINLINE void x264_mb_mc_1xywh( x264_t *h, int x, int y, int width, int height )
64 int i8 = x264_scan8[0]+x+8*y;
65 int i_ref = h->mb.cache.ref[1][i8];
66 int mvx = x264_clip3( h->mb.cache.mv[1][i8][0], h->mb.mv_min[0], h->mb.mv_max[0] ) + 4*4*x;
67 int mvy = x264_clip3( h->mb.cache.mv[1][i8][1], h->mb.mv_min[1], h->mb.mv_max[1] ) + 4*4*y;
69 h->mc.mc_luma( &h->mb.pic.p_fdec[0][4*y*FDEC_STRIDE+4*x], FDEC_STRIDE,
70 h->mb.pic.p_fref[1][i_ref], h->mb.pic.i_stride[0],
71 mvx, mvy, 4*width, 4*height, weight_none );
73 if( h->mb.b_interlaced & i_ref )
74 mvy += (h->mb.i_mb_y & 1)*4 - 2;
76 h->mc.mc_chroma( &h->mb.pic.p_fdec[1][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 );
80 h->mc.mc_chroma( &h->mb.pic.p_fdec[2][2*y*FDEC_STRIDE+2*x], FDEC_STRIDE,
81 h->mb.pic.p_fref[1][i_ref][5], h->mb.pic.i_stride[2],
82 mvx, mvy, 2*width, 2*height );
85 static NOINLINE void x264_mb_mc_01xywh( x264_t *h, int x, int y, int width, int height )
87 int i8 = x264_scan8[0]+x+8*y;
88 int i_ref0 = h->mb.cache.ref[0][i8];
89 int i_ref1 = h->mb.cache.ref[1][i8];
90 int weight = h->mb.bipred_weight[i_ref0][i_ref1];
91 int mvx0 = x264_clip3( h->mb.cache.mv[0][i8][0], h->mb.mv_min[0], h->mb.mv_max[0] ) + 4*4*x;
92 int mvx1 = x264_clip3( h->mb.cache.mv[1][i8][0], h->mb.mv_min[0], h->mb.mv_max[0] ) + 4*4*x;
93 int mvy0 = x264_clip3( h->mb.cache.mv[0][i8][1], h->mb.mv_min[1], h->mb.mv_max[1] ) + 4*4*y;
94 int mvy1 = x264_clip3( h->mb.cache.mv[1][i8][1], h->mb.mv_min[1], h->mb.mv_max[1] ) + 4*4*y;
95 int i_mode = x264_size2pixel[height][width];
96 int i_stride0 = 16, i_stride1 = 16;
97 ALIGNED_ARRAY_16( pixel, tmp0,[16*16] );
98 ALIGNED_ARRAY_16( pixel, tmp1,[16*16] );
101 src0 = h->mc.get_ref( tmp0, &i_stride0, h->mb.pic.p_fref[0][i_ref0], h->mb.pic.i_stride[0],
102 mvx0, mvy0, 4*width, 4*height, weight_none );
103 src1 = h->mc.get_ref( tmp1, &i_stride1, h->mb.pic.p_fref[1][i_ref1], h->mb.pic.i_stride[0],
104 mvx1, mvy1, 4*width, 4*height, weight_none );
105 h->mc.avg[i_mode]( &h->mb.pic.p_fdec[0][4*y*FDEC_STRIDE+4*x], FDEC_STRIDE,
106 src0, i_stride0, src1, i_stride1, weight );
108 if( h->mb.b_interlaced & i_ref0 )
109 mvy0 += (h->mb.i_mb_y & 1)*4 - 2;
110 if( h->mb.b_interlaced & i_ref1 )
111 mvy1 += (h->mb.i_mb_y & 1)*4 - 2;
113 h->mc.mc_chroma( tmp0, 16, h->mb.pic.p_fref[0][i_ref0][4], h->mb.pic.i_stride[1],
114 mvx0, mvy0, 2*width, 2*height );
115 h->mc.mc_chroma( tmp1, 16, h->mb.pic.p_fref[1][i_ref1][4], h->mb.pic.i_stride[1],
116 mvx1, mvy1, 2*width, 2*height );
117 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 );
118 h->mc.mc_chroma( tmp0, 16, h->mb.pic.p_fref[0][i_ref0][5], h->mb.pic.i_stride[2],
119 mvx0, mvy0, 2*width, 2*height );
120 h->mc.mc_chroma( tmp1, 16, h->mb.pic.p_fref[1][i_ref1][5], h->mb.pic.i_stride[2],
121 mvx1, mvy1, 2*width, 2*height );
122 h->mc.avg[i_mode+3]( &h->mb.pic.p_fdec[2][2*y*FDEC_STRIDE+2*x], FDEC_STRIDE, tmp0, 16, tmp1, 16, weight );
125 void x264_mb_mc_8x8( x264_t *h, int i8 )
130 if( h->sh.i_type == SLICE_TYPE_P )
132 switch( h->mb.i_sub_partition[i8] )
135 x264_mb_mc_0xywh( h, x, y, 2, 2 );
138 x264_mb_mc_0xywh( h, x, y+0, 2, 1 );
139 x264_mb_mc_0xywh( h, x, y+1, 2, 1 );
142 x264_mb_mc_0xywh( h, x+0, y, 1, 2 );
143 x264_mb_mc_0xywh( h, x+1, y, 1, 2 );
146 x264_mb_mc_0xywh( h, x+0, y+0, 1, 1 );
147 x264_mb_mc_0xywh( h, x+1, y+0, 1, 1 );
148 x264_mb_mc_0xywh( h, x+0, y+1, 1, 1 );
149 x264_mb_mc_0xywh( h, x+1, y+1, 1, 1 );
155 int scan8 = x264_scan8[0] + x + 8*y;
157 if( h->mb.cache.ref[0][scan8] >= 0 )
158 if( h->mb.cache.ref[1][scan8] >= 0 )
159 x264_mb_mc_01xywh( h, x, y, 2, 2 );
161 x264_mb_mc_0xywh( h, x, y, 2, 2 );
163 x264_mb_mc_1xywh( h, x, y, 2, 2 );
167 void x264_mb_mc( x264_t *h )
169 if( h->mb.i_partition == D_8x8 )
171 for( int i = 0; i < 4; i++ )
172 x264_mb_mc_8x8( h, i );
176 int ref0a = h->mb.cache.ref[0][x264_scan8[ 0]];
177 int ref0b = h->mb.cache.ref[0][x264_scan8[12]];
178 int ref1a = h->mb.cache.ref[1][x264_scan8[ 0]];
179 int ref1b = h->mb.cache.ref[1][x264_scan8[12]];
181 if( h->mb.i_partition == D_16x16 )
184 if( ref1a >= 0 ) x264_mb_mc_01xywh( h, 0, 0, 4, 4 );
185 else x264_mb_mc_0xywh ( h, 0, 0, 4, 4 );
186 else x264_mb_mc_1xywh ( h, 0, 0, 4, 4 );
188 else if( h->mb.i_partition == D_16x8 )
191 if( ref1a >= 0 ) x264_mb_mc_01xywh( h, 0, 0, 4, 2 );
192 else x264_mb_mc_0xywh ( h, 0, 0, 4, 2 );
193 else x264_mb_mc_1xywh ( h, 0, 0, 4, 2 );
196 if( ref1b >= 0 ) x264_mb_mc_01xywh( h, 0, 2, 4, 2 );
197 else x264_mb_mc_0xywh ( h, 0, 2, 4, 2 );
198 else x264_mb_mc_1xywh ( h, 0, 2, 4, 2 );
200 else if( h->mb.i_partition == D_8x16 )
203 if( ref1a >= 0 ) x264_mb_mc_01xywh( h, 0, 0, 2, 4 );
204 else x264_mb_mc_0xywh ( h, 0, 0, 2, 4 );
205 else x264_mb_mc_1xywh ( h, 0, 0, 2, 4 );
208 if( ref1b >= 0 ) x264_mb_mc_01xywh( h, 2, 0, 2, 4 );
209 else x264_mb_mc_0xywh ( h, 2, 0, 2, 4 );
210 else x264_mb_mc_1xywh ( h, 2, 0, 2, 4 );
215 int x264_macroblock_cache_allocate( x264_t *h )
217 int i_mb_count = h->mb.i_mb_count;
219 h->mb.i_mb_stride = h->mb.i_mb_width;
220 h->mb.i_b8_stride = h->mb.i_mb_width * 2;
221 h->mb.i_b4_stride = h->mb.i_mb_width * 4;
223 h->mb.b_interlaced = h->param.b_interlaced;
225 CHECKED_MALLOC( h->mb.qp, i_mb_count * sizeof(int8_t) );
226 CHECKED_MALLOC( h->mb.cbp, i_mb_count * sizeof(int16_t) );
227 CHECKED_MALLOC( h->mb.skipbp, i_mb_count * sizeof(int8_t) );
228 CHECKED_MALLOC( h->mb.mb_transform_size, i_mb_count * sizeof(int8_t) );
229 CHECKED_MALLOC( h->mb.slice_table, i_mb_count * sizeof(uint16_t) );
230 memset( h->mb.slice_table, -1, i_mb_count * sizeof(uint16_t) );
232 /* 0 -> 3 top(4), 4 -> 6 : left(3) */
233 CHECKED_MALLOC( h->mb.intra4x4_pred_mode, i_mb_count * 8 * sizeof(int8_t) );
236 CHECKED_MALLOC( h->mb.non_zero_count, i_mb_count * 24 * sizeof(uint8_t) );
238 if( h->param.b_cabac )
240 CHECKED_MALLOC( h->mb.chroma_pred_mode, i_mb_count * sizeof(int8_t) );
241 CHECKED_MALLOC( h->mb.mvd[0], i_mb_count * sizeof( **h->mb.mvd ) );
242 CHECKED_MALLOC( h->mb.mvd[1], i_mb_count * sizeof( **h->mb.mvd ) );
245 for( int i = 0; i < 2; i++ )
247 int i_refs = X264_MIN(16, (i ? 1 + !!h->param.i_bframe_pyramid : h->param.i_frame_reference) ) << h->param.b_interlaced;
248 if( h->param.analyse.i_weighted_pred == X264_WEIGHTP_SMART )
249 i_refs = X264_MIN(16, i_refs + 2); //smart weights add two duplicate frames
250 else if( h->param.analyse.i_weighted_pred == X264_WEIGHTP_BLIND )
251 i_refs = X264_MIN(16, i_refs + 1); //blind weights add one duplicate frame
253 for( int j = !i; j < i_refs; j++ )
255 CHECKED_MALLOC( h->mb.mvr[i][j], 2 * (i_mb_count + 1) * sizeof(int16_t) );
256 M32( h->mb.mvr[i][j][0] ) = 0;
261 if( h->param.analyse.i_weighted_pred )
263 int i_padv = PADV << h->param.b_interlaced;
264 int align = h->param.cpu&X264_CPU_CACHELINE_64 ? 64 : h->param.cpu&X264_CPU_CACHELINE_32 ? 32 : 16;
265 int i_stride, luma_plane_size = 0;
268 if( h->param.analyse.i_weighted_pred == X264_WEIGHTP_FAKE )
270 // only need buffer for lookahead
271 if( !h->param.i_sync_lookahead || h == h->thread[h->param.i_threads] )
273 // Fake analysis only works on lowres
274 i_stride = ALIGN( h->mb.i_mb_width*8 + 2*PADH, align );
275 luma_plane_size = i_stride * (h->mb.i_mb_height*8+2*i_padv);
276 // Only need 1 buffer for analysis
284 i_stride = ALIGN( h->mb.i_mb_width*16 + 2*PADH, align );
285 luma_plane_size = i_stride * (h->mb.i_mb_height*16+2*i_padv);
287 if( h->param.analyse.i_weighted_pred == X264_WEIGHTP_SMART )
288 //SMART can weight one ref and one offset -1
291 //blind only has one weighted copy (offset -1)
295 for( int i = 0; i < numweightbuf; i++ )
296 CHECKED_MALLOC( h->mb.p_weight_buf[i], luma_plane_size * sizeof(pixel) );
303 void x264_macroblock_cache_free( x264_t *h )
305 for( int i = 0; i < 2; i++ )
306 for( int j = !i; j < 32; j++ )
307 if( h->mb.mvr[i][j] )
308 x264_free( h->mb.mvr[i][j]-1 );
309 for( int i = 0; i < 16; i++ )
310 x264_free( h->mb.p_weight_buf[i] );
312 if( h->param.b_cabac )
314 x264_free( h->mb.chroma_pred_mode );
315 x264_free( h->mb.mvd[0] );
316 x264_free( h->mb.mvd[1] );
318 x264_free( h->mb.slice_table );
319 x264_free( h->mb.intra4x4_pred_mode );
320 x264_free( h->mb.non_zero_count );
321 x264_free( h->mb.mb_transform_size );
322 x264_free( h->mb.skipbp );
323 x264_free( h->mb.cbp );
324 x264_free( h->mb.qp );
327 int x264_macroblock_thread_allocate( x264_t *h, int b_lookahead )
330 for( int i = 0; i <= h->param.b_interlaced; i++ )
332 for( int j = 0; j < 3; j++ )
334 /* shouldn't really be initialized, just silences a valgrind false-positive in predict_8x8_filter_mmx */
335 CHECKED_MALLOCZERO( h->intra_border_backup[i][j], ((h->mb.i_mb_width*16+32)>>!!j) * sizeof(pixel) );
336 h->intra_border_backup[i][j] += 8;
338 CHECKED_MALLOC( h->deblock_strength[i], sizeof(**h->deblock_strength) * h->mb.i_mb_width );
341 /* Allocate scratch buffer */
342 int scratch_size = 0;
345 int buf_hpel = (h->thread[0]->fdec->i_width[0]+48) * sizeof(dctcoef);
346 int buf_ssim = h->param.analyse.b_ssim * 8 * (h->param.i_width/4+3) * sizeof(int);
347 int me_range = X264_MIN(h->param.analyse.i_me_range, h->param.analyse.i_mv_range);
348 int buf_tesa = (h->param.analyse.i_me_method >= X264_ME_ESA) *
349 ((me_range*2+24) * sizeof(int16_t) + (me_range+4) * (me_range+1) * 4 * sizeof(mvsad_t));
350 scratch_size = X264_MAX3( buf_hpel, buf_ssim, buf_tesa );
352 int buf_mbtree = h->param.rc.b_mb_tree * ((h->mb.i_mb_width+3)&~3) * sizeof(int);
353 scratch_size = X264_MAX( scratch_size, buf_mbtree );
354 CHECKED_MALLOC( h->scratch_buffer, scratch_size );
361 void x264_macroblock_thread_free( x264_t *h, int b_lookahead )
364 for( int i = 0; i <= h->param.b_interlaced; i++ )
366 x264_free( h->deblock_strength[i] );
367 for( int j = 0; j < 3; j++ )
368 x264_free( h->intra_border_backup[i][j] - 8 );
370 x264_free( h->scratch_buffer );
373 void x264_macroblock_slice_init( x264_t *h )
375 h->mb.mv[0] = h->fdec->mv[0];
376 h->mb.mv[1] = h->fdec->mv[1];
377 h->mb.mvr[0][0] = h->fdec->mv16x16;
378 h->mb.ref[0] = h->fdec->ref[0];
379 h->mb.ref[1] = h->fdec->ref[1];
380 h->mb.type = h->fdec->mb_type;
381 h->mb.partition = h->fdec->mb_partition;
383 h->fdec->i_ref[0] = h->i_ref0;
384 h->fdec->i_ref[1] = h->i_ref1;
385 for( int i = 0; i < h->i_ref0; i++ )
386 h->fdec->ref_poc[0][i] = h->fref0[i]->i_poc;
387 if( h->sh.i_type == SLICE_TYPE_B )
389 for( int i = 0; i < h->i_ref1; i++ )
390 h->fdec->ref_poc[1][i] = h->fref1[i]->i_poc;
392 map_col_to_list0(-1) = -1;
393 map_col_to_list0(-2) = -2;
394 for( int i = 0; i < h->fref1[0]->i_ref[0]; i++ )
396 int poc = h->fref1[0]->ref_poc[0][i];
397 map_col_to_list0(i) = -2;
398 for( int j = 0; j < h->i_ref0; j++ )
399 if( h->fref0[j]->i_poc == poc )
401 map_col_to_list0(i) = j;
406 else if( h->sh.i_type == SLICE_TYPE_P )
408 memset( h->mb.cache.skip, 0, sizeof( h->mb.cache.skip ) );
410 if( h->sh.i_disable_deblocking_filter_idc != 1 && h->param.analyse.i_weighted_pred )
412 deblock_ref_table(-2) = -2;
413 deblock_ref_table(-1) = -1;
414 for( int i = 0; i < h->i_ref0 << h->sh.b_mbaff; i++ )
416 /* Mask off high bits to avoid frame num collisions with -1/-2.
417 * In current x264 frame num values don't cover a range of more
418 * than 32, so 6 bits is enough for uniqueness. */
419 if( !h->mb.b_interlaced )
420 deblock_ref_table(i) = h->fref0[i]->i_frame_num&63;
422 deblock_ref_table(i) = ((h->fref0[i>>1]->i_frame_num&63)<<1) + (i&1);
427 /* init with not available (for top right idx=7,15) */
428 memset( h->mb.cache.ref, -2, sizeof( h->mb.cache.ref ) );
431 for( int field = 0; field <= h->sh.b_mbaff; field++ )
433 int curpoc = h->fdec->i_poc + field*h->sh.i_delta_poc_bottom;
434 int refpoc = h->fref0[0]->i_poc;
435 if( h->sh.b_mbaff && field )
436 refpoc += h->sh.i_delta_poc_bottom;
437 int delta = curpoc - refpoc;
439 h->fdec->inv_ref_poc[field] = (256 + delta/2) / delta;
442 h->mb.i_neighbour4[6] =
443 h->mb.i_neighbour4[9] =
444 h->mb.i_neighbour4[12] =
445 h->mb.i_neighbour4[14] = MB_LEFT|MB_TOP|MB_TOPLEFT|MB_TOPRIGHT;
446 h->mb.i_neighbour4[3] =
447 h->mb.i_neighbour4[7] =
448 h->mb.i_neighbour4[11] =
449 h->mb.i_neighbour4[13] =
450 h->mb.i_neighbour4[15] =
451 h->mb.i_neighbour8[3] = MB_LEFT|MB_TOP|MB_TOPLEFT;
454 void x264_macroblock_thread_init( x264_t *h )
456 h->mb.i_me_method = h->param.analyse.i_me_method;
457 h->mb.i_subpel_refine = h->param.analyse.i_subpel_refine;
458 if( h->sh.i_type == SLICE_TYPE_B && (h->mb.i_subpel_refine == 6 || h->mb.i_subpel_refine == 8) )
459 h->mb.i_subpel_refine--;
460 h->mb.b_chroma_me = h->param.analyse.b_chroma_me && h->sh.i_type == SLICE_TYPE_P
461 && h->mb.i_subpel_refine >= 5;
462 h->mb.b_dct_decimate = h->sh.i_type == SLICE_TYPE_B ||
463 (h->param.analyse.b_dct_decimate && h->sh.i_type != SLICE_TYPE_I);
476 h->mb.pic.p_fenc[0] = h->mb.pic.fenc_buf;
477 h->mb.pic.p_fenc[1] = h->mb.pic.fenc_buf + 16*FENC_STRIDE;
478 h->mb.pic.p_fenc[2] = h->mb.pic.fenc_buf + 16*FENC_STRIDE + 8;
479 h->mb.pic.p_fdec[0] = h->mb.pic.fdec_buf + 2*FDEC_STRIDE;
480 h->mb.pic.p_fdec[1] = h->mb.pic.fdec_buf + 19*FDEC_STRIDE;
481 h->mb.pic.p_fdec[2] = h->mb.pic.fdec_buf + 19*FDEC_STRIDE + 16;
484 void x264_prefetch_fenc( x264_t *h, x264_frame_t *fenc, int i_mb_x, int i_mb_y )
486 int stride_y = fenc->i_stride[0];
487 int stride_uv = fenc->i_stride[1];
488 int off_y = 16 * (i_mb_x + i_mb_y * stride_y);
489 int off_uv = 8 * (i_mb_x + i_mb_y * stride_uv);
490 h->mc.prefetch_fenc( fenc->plane[0]+off_y, stride_y,
491 fenc->plane[1+(i_mb_x&1)]+off_uv, stride_uv, i_mb_x );
494 static NOINLINE void copy_column8( pixel *dst, pixel *src )
496 // input pointers are offset by 4 rows because that's faster (smaller instruction size on x86)
497 for( int i = -4; i < 4; i++ )
498 dst[i*FDEC_STRIDE] = src[i*FDEC_STRIDE];
501 static void ALWAYS_INLINE x264_macroblock_load_pic_pointers( x264_t *h, int mb_x, int mb_y, int i, int b_interlaced )
503 const int w = (i == 0 ? 16 : 8);
504 const int i_stride = h->fdec->i_stride[!!i];
505 const int i_stride2 = i_stride << b_interlaced;
506 const int i_pix_offset = b_interlaced
507 ? w * (mb_x + (mb_y&~1) * i_stride) + (mb_y&1) * i_stride
508 : w * (mb_x + mb_y * i_stride);
509 const pixel *plane_fdec = &h->fdec->plane[i][i_pix_offset];
510 const pixel *intra_fdec = &h->intra_border_backup[mb_y & h->sh.b_mbaff][i][mb_x*16>>!!i];
511 int ref_pix_offset[2] = { i_pix_offset, i_pix_offset };
512 x264_frame_t **fref[2] = { h->fref0, h->fref1 };
514 ref_pix_offset[1] += (1-2*(mb_y&1)) * i_stride;
515 h->mb.pic.i_stride[i] = i_stride2;
516 h->mb.pic.p_fenc_plane[i] = &h->fenc->plane[i][i_pix_offset];
517 h->mc.copy[i?PIXEL_8x8:PIXEL_16x16]( h->mb.pic.p_fenc[i], FENC_STRIDE,
518 h->mb.pic.p_fenc_plane[i], i_stride2, w );
519 memcpy( &h->mb.pic.p_fdec[i][-1-FDEC_STRIDE], intra_fdec-1, (w*3/2+1) * sizeof(pixel) );
521 for( int j = 0; j < w; j++ )
522 h->mb.pic.p_fdec[i][-1+j*FDEC_STRIDE] = plane_fdec[-1+j*i_stride2];
523 for( int j = 0; j < h->mb.pic.i_fref[0]; j++ )
525 h->mb.pic.p_fref[0][j][i==0 ? 0:i+3] = &fref[0][j >> b_interlaced]->plane[i][ref_pix_offset[j&1]];
528 for( int k = 1; k < 4; k++ )
529 h->mb.pic.p_fref[0][j][k] = &fref[0][j >> b_interlaced]->filtered[k][ref_pix_offset[j&1]];
530 if( h->sh.weight[j][0].weightfn )
531 h->mb.pic.p_fref_w[j] = &h->fenc->weighted[j >> b_interlaced][ref_pix_offset[j&1]];
533 h->mb.pic.p_fref_w[j] = h->mb.pic.p_fref[0][j][0];
536 if( h->sh.i_type == SLICE_TYPE_B )
537 for( int j = 0; j < h->mb.pic.i_fref[1]; j++ )
539 h->mb.pic.p_fref[1][j][i==0 ? 0:i+3] = &fref[1][j >> b_interlaced]->plane[i][ref_pix_offset[j&1]];
541 for( int k = 1; k < 4; k++ )
542 h->mb.pic.p_fref[1][j][k] = &fref[1][j >> b_interlaced]->filtered[k][ref_pix_offset[j&1]];
546 static void inline x264_macroblock_cache_load_neighbours( x264_t *h, int mb_x, int mb_y )
548 int top = (mb_y - (1 << h->mb.b_interlaced)) * h->mb.i_mb_stride + mb_x;
552 h->mb.i_mb_xy = mb_y * h->mb.i_mb_stride + mb_x;
553 h->mb.i_b8_xy = 2*(mb_y * h->mb.i_b8_stride + mb_x);
554 h->mb.i_b4_xy = 4*(mb_y * h->mb.i_b4_stride + mb_x);
555 h->mb.i_neighbour = 0;
556 h->mb.i_neighbour_intra = 0;
557 h->mb.i_neighbour_frame = 0;
558 h->mb.i_mb_top_xy = -1;
559 h->mb.i_mb_left_xy = -1;
560 h->mb.i_mb_topleft_xy = -1;
561 h->mb.i_mb_topright_xy = -1;
562 h->mb.i_mb_type_top = -1;
563 h->mb.i_mb_type_left = -1;
564 h->mb.i_mb_type_topleft = -1;
565 h->mb.i_mb_type_topright = -1;
569 h->mb.i_neighbour_frame |= MB_LEFT;
570 h->mb.i_mb_left_xy = h->mb.i_mb_xy - 1;
571 h->mb.i_mb_type_left = h->mb.type[h->mb.i_mb_left_xy];
572 if( h->mb.i_mb_xy > h->sh.i_first_mb )
574 h->mb.i_neighbour |= MB_LEFT;
576 if( !h->param.b_constrained_intra || IS_INTRA( h->mb.i_mb_type_left ) )
577 h->mb.i_neighbour_intra |= MB_LEFT;
581 /* We can't predict from the previous threadslice since it hasn't been encoded yet. */
582 if( (h->i_threadslice_start >> h->mb.b_interlaced) != (mb_y >> h->mb.b_interlaced) )
586 h->mb.i_neighbour_frame |= MB_TOP;
587 h->mb.i_mb_top_xy = top;
588 h->mb.i_mb_type_top = h->mb.type[h->mb.i_mb_top_xy];
589 if( top >= h->sh.i_first_mb )
591 h->mb.i_neighbour |= MB_TOP;
593 if( !h->param.b_constrained_intra || IS_INTRA( h->mb.i_mb_type_top ) )
594 h->mb.i_neighbour_intra |= MB_TOP;
596 /* We only need to prefetch the top blocks because the left was just written
597 * to as part of the previous cache_save. Since most target CPUs use write-allocate
598 * caches, left blocks are near-guaranteed to be in L1 cache. Top--not so much. */
599 x264_prefetch( &h->mb.cbp[top] );
600 x264_prefetch( h->mb.intra4x4_pred_mode[top] );
601 x264_prefetch( &h->mb.non_zero_count[top][12] );
602 /* These aren't always allocated, but prefetching an invalid address can't hurt. */
603 x264_prefetch( &h->mb.mb_transform_size[top] );
604 x264_prefetch( &h->mb.skipbp[top] );
608 if( mb_x > 0 && top - 1 >= 0 )
610 h->mb.i_neighbour_frame |= MB_TOPLEFT;
611 h->mb.i_mb_topleft_xy = top - 1;
612 h->mb.i_mb_type_topleft = h->mb.type[h->mb.i_mb_topleft_xy];
613 if( top - 1 >= h->sh.i_first_mb )
615 h->mb.i_neighbour |= MB_TOPLEFT;
617 if( !h->param.b_constrained_intra || IS_INTRA( h->mb.i_mb_type_topleft ) )
618 h->mb.i_neighbour_intra |= MB_TOPLEFT;
622 if( mb_x < h->mb.i_mb_width - 1 && top + 1 >= 0 )
624 h->mb.i_neighbour_frame |= MB_TOPRIGHT;
625 h->mb.i_mb_topright_xy = top + 1;
626 h->mb.i_mb_type_topright = h->mb.type[h->mb.i_mb_topright_xy];
627 if( top + 1 >= h->sh.i_first_mb )
629 h->mb.i_neighbour |= MB_TOPRIGHT;
631 if( !h->param.b_constrained_intra || IS_INTRA( h->mb.i_mb_type_topright ) )
632 h->mb.i_neighbour_intra |= MB_TOPRIGHT;
638 void x264_macroblock_cache_load( x264_t *h, int mb_x, int mb_y )
640 x264_macroblock_cache_load_neighbours( h, mb_x, mb_y );
642 int left = h->mb.i_mb_left_xy;
643 int top = h->mb.i_mb_top_xy;
644 int top_y = mb_y - (1 << h->mb.b_interlaced);
645 int s8x8 = h->mb.i_b8_stride;
646 int s4x4 = h->mb.i_b4_stride;
647 int top_8x8 = (2*top_y+1) * s8x8 + 2*mb_x;
648 int top_4x4 = (4*top_y+3) * s4x4 + 4*mb_x;
649 int lists = (1 << h->sh.i_type) & 3;
651 /* GCC pessimizes direct loads from heap-allocated arrays due to aliasing. */
652 /* By only dereferencing them once, we avoid this issue. */
653 int8_t (*i4x4)[8] = h->mb.intra4x4_pred_mode;
654 uint8_t (*nnz)[24] = h->mb.non_zero_count;
655 int16_t *cbp = h->mb.cbp;
658 if( h->mb.i_neighbour & MB_TOP )
660 h->mb.cache.i_cbp_top = cbp[top];
662 CP32( &h->mb.cache.intra4x4_pred_mode[x264_scan8[0] - 8], &i4x4[top][0] );
664 /* load non_zero_count */
665 CP32( &h->mb.cache.non_zero_count[x264_scan8[0] - 8], &nnz[top][12] );
666 /* shift because x264_scan8[16] is misaligned */
667 M32( &h->mb.cache.non_zero_count[x264_scan8[16+0] - 9] ) = M16( &nnz[top][18] ) << 8;
668 M32( &h->mb.cache.non_zero_count[x264_scan8[16+4] - 9] ) = M16( &nnz[top][22] ) << 8;
672 h->mb.cache.i_cbp_top = -1;
675 M32( &h->mb.cache.intra4x4_pred_mode[x264_scan8[0] - 8] ) = 0xFFFFFFFFU;
677 /* load non_zero_count */
678 M32( &h->mb.cache.non_zero_count[x264_scan8[ 0] - 8] ) = 0x80808080U;
679 M32( &h->mb.cache.non_zero_count[x264_scan8[16+0] - 9] ) = 0x80808080U;
680 M32( &h->mb.cache.non_zero_count[x264_scan8[16+4] - 9] ) = 0x80808080U;
683 if( h->mb.i_neighbour & MB_LEFT )
685 h->mb.cache.i_cbp_left = cbp[left];
688 h->mb.cache.intra4x4_pred_mode[x264_scan8[0 ] - 1] = i4x4[left][4];
689 h->mb.cache.intra4x4_pred_mode[x264_scan8[2 ] - 1] = i4x4[left][5];
690 h->mb.cache.intra4x4_pred_mode[x264_scan8[8 ] - 1] = i4x4[left][6];
691 h->mb.cache.intra4x4_pred_mode[x264_scan8[10] - 1] = i4x4[left][3];
693 /* load non_zero_count */
694 h->mb.cache.non_zero_count[x264_scan8[0 ] - 1] = nnz[left][3];
695 h->mb.cache.non_zero_count[x264_scan8[2 ] - 1] = nnz[left][7];
696 h->mb.cache.non_zero_count[x264_scan8[8 ] - 1] = nnz[left][11];
697 h->mb.cache.non_zero_count[x264_scan8[10] - 1] = nnz[left][15];
699 h->mb.cache.non_zero_count[x264_scan8[16+0] - 1] = nnz[left][16+1];
700 h->mb.cache.non_zero_count[x264_scan8[16+2] - 1] = nnz[left][16+3];
702 h->mb.cache.non_zero_count[x264_scan8[16+4+0] - 1] = nnz[left][16+4+1];
703 h->mb.cache.non_zero_count[x264_scan8[16+4+2] - 1] = nnz[left][16+4+3];
705 /* Finish the prefetching */
706 for( int l = 0; l < lists; l++ )
708 x264_prefetch( &h->mb.mv[l][top_4x4-1] );
709 /* Top right being not in the same cacheline as top left will happen
710 * once every 4 MBs, so one extra prefetch is worthwhile */
711 x264_prefetch( &h->mb.mv[l][top_4x4+4] );
712 x264_prefetch( &h->mb.ref[l][top_8x8-1] );
713 x264_prefetch( &h->mb.mvd[l][top] );
718 h->mb.cache.i_cbp_left = -1;
720 h->mb.cache.intra4x4_pred_mode[x264_scan8[0 ] - 1] =
721 h->mb.cache.intra4x4_pred_mode[x264_scan8[2 ] - 1] =
722 h->mb.cache.intra4x4_pred_mode[x264_scan8[8 ] - 1] =
723 h->mb.cache.intra4x4_pred_mode[x264_scan8[10] - 1] = -1;
725 /* load non_zero_count */
726 h->mb.cache.non_zero_count[x264_scan8[0 ] - 1] =
727 h->mb.cache.non_zero_count[x264_scan8[2 ] - 1] =
728 h->mb.cache.non_zero_count[x264_scan8[8 ] - 1] =
729 h->mb.cache.non_zero_count[x264_scan8[10] - 1] =
730 h->mb.cache.non_zero_count[x264_scan8[16+0] - 1] =
731 h->mb.cache.non_zero_count[x264_scan8[16+2] - 1] =
732 h->mb.cache.non_zero_count[x264_scan8[16+4+0] - 1] =
733 h->mb.cache.non_zero_count[x264_scan8[16+4+2] - 1] = 0x80;
736 if( h->pps->b_transform_8x8_mode )
738 h->mb.cache.i_neighbour_transform_size =
739 ( (h->mb.i_neighbour & MB_LEFT) && h->mb.mb_transform_size[left] )
740 + ( (h->mb.i_neighbour & MB_TOP) && h->mb.mb_transform_size[top] );
745 h->mb.pic.i_fref[0] = h->i_ref0 << h->mb.b_interlaced;
746 h->mb.pic.i_fref[1] = h->i_ref1 << h->mb.b_interlaced;
747 h->mb.cache.i_neighbour_interlaced =
748 !!(h->mb.i_neighbour & MB_LEFT)
749 + !!(h->mb.i_neighbour & MB_TOP);
752 if( !h->mb.b_interlaced )
754 copy_column8( h->mb.pic.p_fdec[0]-1+ 4*FDEC_STRIDE, h->mb.pic.p_fdec[0]+15+ 4*FDEC_STRIDE );
755 copy_column8( h->mb.pic.p_fdec[0]-1+12*FDEC_STRIDE, h->mb.pic.p_fdec[0]+15+12*FDEC_STRIDE );
756 copy_column8( h->mb.pic.p_fdec[1]-1+ 4*FDEC_STRIDE, h->mb.pic.p_fdec[1]+ 7+ 4*FDEC_STRIDE );
757 copy_column8( h->mb.pic.p_fdec[2]-1+ 4*FDEC_STRIDE, h->mb.pic.p_fdec[2]+ 7+ 4*FDEC_STRIDE );
758 x264_macroblock_load_pic_pointers( h, mb_x, mb_y, 0, 0 );
759 x264_macroblock_load_pic_pointers( h, mb_x, mb_y, 1, 0 );
760 x264_macroblock_load_pic_pointers( h, mb_x, mb_y, 2, 0 );
764 x264_macroblock_load_pic_pointers( h, mb_x, mb_y, 0, 1 );
765 x264_macroblock_load_pic_pointers( h, mb_x, mb_y, 1, 1 );
766 x264_macroblock_load_pic_pointers( h, mb_x, mb_y, 2, 1 );
769 if( h->fdec->integral )
771 int offset = 16 * (mb_x + mb_y * h->fdec->i_stride[0]);
772 for( int i = 0; i < h->mb.pic.i_fref[0]; i++ )
773 h->mb.pic.p_integral[0][i] = &h->fref0[i]->integral[offset];
774 for( int i = 0; i < h->mb.pic.i_fref[1]; i++ )
775 h->mb.pic.p_integral[1][i] = &h->fref1[i]->integral[offset];
778 x264_prefetch_fenc( h, h->fenc, mb_x, mb_y );
780 /* load ref/mv/mvd */
781 for( int l = 0; l < lists; l++ )
783 int16_t (*mv)[2] = h->mb.mv[l];
784 int8_t *ref = h->mb.ref[l];
786 int i8 = x264_scan8[0] - 1 - 1*8;
787 if( h->mb.i_neighbour & MB_TOPLEFT )
789 h->mb.cache.ref[l][i8] = ref[top_8x8 - 1];
790 CP32( h->mb.cache.mv[l][i8], mv[top_4x4 - 1] );
794 h->mb.cache.ref[l][i8] = -2;
795 M32( h->mb.cache.mv[l][i8] ) = 0;
798 i8 = x264_scan8[0] - 8;
799 if( h->mb.i_neighbour & MB_TOP )
801 h->mb.cache.ref[l][i8+0] =
802 h->mb.cache.ref[l][i8+1] = ref[top_8x8 + 0];
803 h->mb.cache.ref[l][i8+2] =
804 h->mb.cache.ref[l][i8+3] = ref[top_8x8 + 1];
805 CP128( h->mb.cache.mv[l][i8], mv[top_4x4] );
809 M128( h->mb.cache.mv[l][i8] ) = M128_ZERO;
810 M32( &h->mb.cache.ref[l][i8] ) = (uint8_t)(-2) * 0x01010101U;
813 i8 = x264_scan8[0] + 4 - 1*8;
814 if( h->mb.i_neighbour & MB_TOPRIGHT )
816 h->mb.cache.ref[l][i8] = ref[top_8x8 + 2];
817 CP32( h->mb.cache.mv[l][i8], mv[top_4x4 + 4] );
820 h->mb.cache.ref[l][i8] = -2;
822 i8 = x264_scan8[0] - 1;
823 if( h->mb.i_neighbour & MB_LEFT )
825 const int ir = h->mb.i_b8_xy - 1;
826 const int iv = h->mb.i_b4_xy - 1;
827 h->mb.cache.ref[l][i8+0*8] =
828 h->mb.cache.ref[l][i8+1*8] = ref[ir + 0*s8x8];
829 h->mb.cache.ref[l][i8+2*8] =
830 h->mb.cache.ref[l][i8+3*8] = ref[ir + 1*s8x8];
832 CP32( h->mb.cache.mv[l][i8+0*8], mv[iv + 0*s4x4] );
833 CP32( h->mb.cache.mv[l][i8+1*8], mv[iv + 1*s4x4] );
834 CP32( h->mb.cache.mv[l][i8+2*8], mv[iv + 2*s4x4] );
835 CP32( h->mb.cache.mv[l][i8+3*8], mv[iv + 3*s4x4] );
839 for( int i = 0; i < 4; i++ )
841 h->mb.cache.ref[l][i8+i*8] = -2;
842 M32( h->mb.cache.mv[l][i8+i*8] ) = 0;
846 if( h->param.b_cabac )
848 uint8_t (*mvd)[8][2] = h->mb.mvd[l];
849 if( h->mb.i_neighbour & MB_TOP )
850 CP64( h->mb.cache.mvd[l][x264_scan8[0] - 8], mvd[top][0] );
852 M64( h->mb.cache.mvd[l][x264_scan8[0] - 8] ) = 0;
854 if( h->mb.i_neighbour & MB_LEFT )
856 CP16( h->mb.cache.mvd[l][x264_scan8[0 ] - 1], mvd[left][4] );
857 CP16( h->mb.cache.mvd[l][x264_scan8[2 ] - 1], mvd[left][5] );
858 CP16( h->mb.cache.mvd[l][x264_scan8[8 ] - 1], mvd[left][6] );
859 CP16( h->mb.cache.mvd[l][x264_scan8[10] - 1], mvd[left][3] );
862 for( int i = 0; i < 4; i++ )
863 M16( h->mb.cache.mvd[l][x264_scan8[0]-1+i*8] ) = 0;
868 if( h->sh.i_type == SLICE_TYPE_B )
870 h->mb.bipred_weight = h->mb.bipred_weight_buf[h->mb.b_interlaced&(mb_y&1)];
871 h->mb.dist_scale_factor = h->mb.dist_scale_factor_buf[h->mb.b_interlaced&(mb_y&1)];
872 if( h->param.b_cabac )
875 x264_macroblock_cache_skip( h, 0, 0, 4, 4, 0 );
876 skipbp = (h->mb.i_neighbour & MB_LEFT) ? h->mb.skipbp[left] : 0;
877 h->mb.cache.skip[x264_scan8[0] - 1] = skipbp & 0x2;
878 h->mb.cache.skip[x264_scan8[8] - 1] = skipbp & 0x8;
879 skipbp = (h->mb.i_neighbour & MB_TOP) ? h->mb.skipbp[top] : 0;
880 h->mb.cache.skip[x264_scan8[0] - 8] = skipbp & 0x4;
881 h->mb.cache.skip[x264_scan8[4] - 8] = skipbp & 0x8;
885 if( h->sh.i_type == SLICE_TYPE_P )
886 x264_mb_predict_mv_pskip( h, h->mb.cache.pskip_mv );
888 h->mb.i_neighbour4[0] =
889 h->mb.i_neighbour8[0] = (h->mb.i_neighbour_intra & (MB_TOP|MB_LEFT|MB_TOPLEFT))
890 | ((h->mb.i_neighbour_intra & MB_TOP) ? MB_TOPRIGHT : 0);
891 h->mb.i_neighbour4[4] =
892 h->mb.i_neighbour4[1] = MB_LEFT | ((h->mb.i_neighbour_intra & MB_TOP) ? (MB_TOP|MB_TOPLEFT|MB_TOPRIGHT) : 0);
893 h->mb.i_neighbour4[2] =
894 h->mb.i_neighbour4[8] =
895 h->mb.i_neighbour4[10] =
896 h->mb.i_neighbour8[2] = MB_TOP|MB_TOPRIGHT | ((h->mb.i_neighbour_intra & MB_LEFT) ? (MB_LEFT|MB_TOPLEFT) : 0);
897 h->mb.i_neighbour4[5] =
898 h->mb.i_neighbour8[1] = MB_LEFT | (h->mb.i_neighbour_intra & MB_TOPRIGHT)
899 | ((h->mb.i_neighbour_intra & MB_TOP) ? MB_TOP|MB_TOPLEFT : 0);
902 void x264_macroblock_cache_load_neighbours_deblock( x264_t *h, int mb_x, int mb_y )
904 int deblock_on_slice_edges = h->sh.i_disable_deblocking_filter_idc != 2;
906 h->mb.i_neighbour = 0;
907 h->mb.i_mb_xy = mb_y * h->mb.i_mb_stride + mb_x;
911 h->mb.i_mb_left_xy = h->mb.i_mb_xy - 1;
912 if( deblock_on_slice_edges || h->mb.slice_table[h->mb.i_mb_left_xy] == h->mb.slice_table[h->mb.i_mb_xy] )
913 h->mb.i_neighbour |= MB_LEFT;
916 if( mb_y > h->mb.b_interlaced )
918 h->mb.i_mb_top_xy = h->mb.i_mb_xy - (h->mb.i_mb_stride << h->mb.b_interlaced);
919 if( deblock_on_slice_edges || h->mb.slice_table[h->mb.i_mb_top_xy] == h->mb.slice_table[h->mb.i_mb_xy] )
920 h->mb.i_neighbour |= MB_TOP;
924 void x264_macroblock_cache_load_deblock( x264_t *h )
926 if( IS_INTRA( h->mb.type[h->mb.i_mb_xy] ) )
929 /* If we have multiple slices and we're deblocking on slice edges, we
930 * have to reload neighbour data. */
931 if( h->sh.i_first_mb && h->sh.i_disable_deblocking_filter_idc != 2 )
933 int old_neighbour = h->mb.i_neighbour;
934 int mb_x = h->mb.i_mb_x;
935 int mb_y = h->mb.i_mb_y;
936 x264_macroblock_cache_load_neighbours_deblock( h, mb_x, mb_y );
937 int new_neighbour = h->mb.i_neighbour;
938 h->mb.i_neighbour &= ~old_neighbour;
939 if( h->mb.i_neighbour )
941 int top_y = mb_y - (1 << h->mb.b_interlaced);
942 int top_8x8 = (2*top_y+1) * h->mb.i_b8_stride + 2*mb_x;
943 int top_4x4 = (4*top_y+3) * h->mb.i_b4_stride + 4*mb_x;
944 int s8x8 = h->mb.i_b8_stride;
945 int s4x4 = h->mb.i_b4_stride;
947 uint8_t (*nnz)[24] = h->mb.non_zero_count;
949 if( h->mb.i_neighbour & MB_TOP )
950 CP32( &h->mb.cache.non_zero_count[x264_scan8[0] - 8], &nnz[h->mb.i_mb_top_xy][12] );
952 if( h->mb.i_neighbour & MB_LEFT )
954 int left = h->mb.i_mb_left_xy;
955 h->mb.cache.non_zero_count[x264_scan8[0 ] - 1] = nnz[left][3];
956 h->mb.cache.non_zero_count[x264_scan8[2 ] - 1] = nnz[left][7];
957 h->mb.cache.non_zero_count[x264_scan8[8 ] - 1] = nnz[left][11];
958 h->mb.cache.non_zero_count[x264_scan8[10] - 1] = nnz[left][15];
961 for( int l = 0; l <= (h->sh.i_type == SLICE_TYPE_B); l++ )
963 int16_t (*mv)[2] = h->mb.mv[l];
964 int8_t *ref = h->mb.ref[l];
966 int i8 = x264_scan8[0] - 8;
967 if( h->mb.i_neighbour & MB_TOP )
969 h->mb.cache.ref[l][i8+0] =
970 h->mb.cache.ref[l][i8+1] = ref[top_8x8 + 0];
971 h->mb.cache.ref[l][i8+2] =
972 h->mb.cache.ref[l][i8+3] = ref[top_8x8 + 1];
973 CP128( h->mb.cache.mv[l][i8], mv[top_4x4] );
976 i8 = x264_scan8[0] - 1;
977 if( h->mb.i_neighbour & MB_LEFT )
979 int ir = h->mb.i_b8_xy - 1;
980 int iv = h->mb.i_b4_xy - 1;
981 h->mb.cache.ref[l][i8+0*8] =
982 h->mb.cache.ref[l][i8+1*8] = ref[ir + 0*s8x8];
983 h->mb.cache.ref[l][i8+2*8] =
984 h->mb.cache.ref[l][i8+3*8] = ref[ir + 1*s8x8];
986 CP32( h->mb.cache.mv[l][i8+0*8], mv[iv + 0*s4x4] );
987 CP32( h->mb.cache.mv[l][i8+1*8], mv[iv + 1*s4x4] );
988 CP32( h->mb.cache.mv[l][i8+2*8], mv[iv + 2*s4x4] );
989 CP32( h->mb.cache.mv[l][i8+3*8], mv[iv + 3*s4x4] );
993 h->mb.i_neighbour = new_neighbour;
996 if( h->param.analyse.i_weighted_pred && h->sh.i_type == SLICE_TYPE_P )
998 /* Handle reference frame duplicates */
999 int i8 = x264_scan8[0] - 8;
1000 h->mb.cache.ref[0][i8+0] =
1001 h->mb.cache.ref[0][i8+1] = deblock_ref_table(h->mb.cache.ref[0][i8+0]);
1002 h->mb.cache.ref[0][i8+2] =
1003 h->mb.cache.ref[0][i8+3] = deblock_ref_table(h->mb.cache.ref[0][i8+2]);
1005 i8 = x264_scan8[0] - 1;
1006 h->mb.cache.ref[0][i8+0*8] =
1007 h->mb.cache.ref[0][i8+1*8] = deblock_ref_table(h->mb.cache.ref[0][i8+0*8]);
1008 h->mb.cache.ref[0][i8+2*8] =
1009 h->mb.cache.ref[0][i8+3*8] = deblock_ref_table(h->mb.cache.ref[0][i8+2*8]);
1011 int ref0 = deblock_ref_table(h->mb.cache.ref[0][x264_scan8[ 0]]);
1012 int ref1 = deblock_ref_table(h->mb.cache.ref[0][x264_scan8[ 4]]);
1013 int ref2 = deblock_ref_table(h->mb.cache.ref[0][x264_scan8[ 8]]);
1014 int ref3 = deblock_ref_table(h->mb.cache.ref[0][x264_scan8[12]]);
1015 uint32_t reftop = pack16to32( (uint8_t)ref0, (uint8_t)ref1 ) * 0x0101;
1016 uint32_t refbot = pack16to32( (uint8_t)ref2, (uint8_t)ref3 ) * 0x0101;
1018 M32( &h->mb.cache.ref[0][x264_scan8[0]+8*0] ) = reftop;
1019 M32( &h->mb.cache.ref[0][x264_scan8[0]+8*1] ) = reftop;
1020 M32( &h->mb.cache.ref[0][x264_scan8[0]+8*2] ) = refbot;
1021 M32( &h->mb.cache.ref[0][x264_scan8[0]+8*3] ) = refbot;
1024 /* Munge NNZ for cavlc + 8x8dct */
1025 if( !h->param.b_cabac && h->pps->b_transform_8x8_mode )
1027 uint8_t (*nnz)[24] = h->mb.non_zero_count;
1028 int top = h->mb.i_mb_top_xy;
1029 int left = h->mb.i_mb_left_xy;
1031 if( (h->mb.i_neighbour & MB_TOP) && h->mb.mb_transform_size[top] )
1033 int i8 = x264_scan8[0] - 8;
1034 int nnz_top0 = M16( &nnz[top][8] ) | M16( &nnz[top][12] );
1035 int nnz_top1 = M16( &nnz[top][10] ) | M16( &nnz[top][14] );
1036 M16( &h->mb.cache.non_zero_count[i8+0] ) = nnz_top0 ? 0x0101 : 0;
1037 M16( &h->mb.cache.non_zero_count[i8+2] ) = nnz_top1 ? 0x0101 : 0;
1040 if( (h->mb.i_neighbour & MB_LEFT) && h->mb.mb_transform_size[left] )
1042 int i8 = x264_scan8[0] - 1;
1043 int nnz_left0 = M16( &nnz[left][2] ) | M16( &nnz[left][6] );
1044 int nnz_left1 = M16( &nnz[left][10] ) | M16( &nnz[left][14] );
1045 h->mb.cache.non_zero_count[i8+8*0] = !!nnz_left0;
1046 h->mb.cache.non_zero_count[i8+8*1] = !!nnz_left0;
1047 h->mb.cache.non_zero_count[i8+8*2] = !!nnz_left1;
1048 h->mb.cache.non_zero_count[i8+8*3] = !!nnz_left1;
1051 if( h->mb.mb_transform_size[h->mb.i_mb_xy] )
1053 int nnz0 = M16( &h->mb.cache.non_zero_count[x264_scan8[ 0]] ) | M16( &h->mb.cache.non_zero_count[x264_scan8[ 2]] );
1054 int nnz1 = M16( &h->mb.cache.non_zero_count[x264_scan8[ 4]] ) | M16( &h->mb.cache.non_zero_count[x264_scan8[ 6]] );
1055 int nnz2 = M16( &h->mb.cache.non_zero_count[x264_scan8[ 8]] ) | M16( &h->mb.cache.non_zero_count[x264_scan8[10]] );
1056 int nnz3 = M16( &h->mb.cache.non_zero_count[x264_scan8[12]] ) | M16( &h->mb.cache.non_zero_count[x264_scan8[14]] );
1057 uint32_t nnztop = pack16to32( !!nnz0, !!nnz1 ) * 0x0101;
1058 uint32_t nnzbot = pack16to32( !!nnz2, !!nnz3 ) * 0x0101;
1060 M32( &h->mb.cache.non_zero_count[x264_scan8[0]+8*0] ) = nnztop;
1061 M32( &h->mb.cache.non_zero_count[x264_scan8[0]+8*1] ) = nnztop;
1062 M32( &h->mb.cache.non_zero_count[x264_scan8[0]+8*2] ) = nnzbot;
1063 M32( &h->mb.cache.non_zero_count[x264_scan8[0]+8*3] ) = nnzbot;
1068 static void ALWAYS_INLINE x264_macroblock_store_pic( x264_t *h, int i )
1071 int i_stride = h->fdec->i_stride[!!i];
1072 int i_stride2 = i_stride << h->mb.b_interlaced;
1073 int i_pix_offset = h->mb.b_interlaced
1074 ? w * (h->mb.i_mb_x + (h->mb.i_mb_y&~1) * i_stride) + (h->mb.i_mb_y&1) * i_stride
1075 : w * (h->mb.i_mb_x + h->mb.i_mb_y * i_stride);
1076 h->mc.copy[i?PIXEL_8x8:PIXEL_16x16]( &h->fdec->plane[i][i_pix_offset], i_stride2,
1077 h->mb.pic.p_fdec[i], FDEC_STRIDE, w );
1080 void x264_macroblock_cache_save( x264_t *h )
1082 const int i_mb_xy = h->mb.i_mb_xy;
1083 const int i_mb_type = x264_mb_type_fix[h->mb.i_type];
1084 const int s8x8 = h->mb.i_b8_stride;
1085 const int s4x4 = h->mb.i_b4_stride;
1086 const int i_mb_4x4 = h->mb.i_b4_xy;
1087 const int i_mb_8x8 = h->mb.i_b8_xy;
1089 /* GCC pessimizes direct stores to heap-allocated arrays due to aliasing. */
1090 /* By only dereferencing them once, we avoid this issue. */
1091 int8_t *i4x4 = h->mb.intra4x4_pred_mode[i_mb_xy];
1092 uint8_t *nnz = h->mb.non_zero_count[i_mb_xy];
1094 x264_macroblock_store_pic( h, 0 );
1095 x264_macroblock_store_pic( h, 1 );
1096 x264_macroblock_store_pic( h, 2 );
1098 x264_prefetch_fenc( h, h->fdec, h->mb.i_mb_x, h->mb.i_mb_y );
1100 h->mb.type[i_mb_xy] = i_mb_type;
1101 h->mb.slice_table[i_mb_xy] = h->sh.i_first_mb;
1102 h->mb.partition[i_mb_xy] = IS_INTRA( i_mb_type ) ? D_16x16 : h->mb.i_partition;
1103 h->mb.i_mb_prev_xy = i_mb_xy;
1106 if( i_mb_type == I_4x4 )
1108 CP32( &i4x4[0], &h->mb.cache.intra4x4_pred_mode[x264_scan8[10]] );
1109 M32( &i4x4[4] ) = pack8to32( h->mb.cache.intra4x4_pred_mode[x264_scan8[5] ],
1110 h->mb.cache.intra4x4_pred_mode[x264_scan8[7] ],
1111 h->mb.cache.intra4x4_pred_mode[x264_scan8[13] ], 0);
1113 else if( !h->param.b_constrained_intra || IS_INTRA(i_mb_type) )
1114 M64( i4x4 ) = I_PRED_4x4_DC * 0x0101010101010101ULL;
1116 M64( i4x4 ) = (uint8_t)(-1) * 0x0101010101010101ULL;
1119 if( i_mb_type == I_PCM )
1121 h->mb.qp[i_mb_xy] = 0;
1122 h->mb.i_last_dqp = 0;
1123 h->mb.i_cbp_chroma = 2;
1124 h->mb.i_cbp_luma = 0xf;
1125 h->mb.cbp[i_mb_xy] = 0x72f; /* all set */
1126 h->mb.b_transform_8x8 = 0;
1127 memset( nnz, 16, sizeof( *h->mb.non_zero_count ) );
1128 for( int i = 0; i < 24; i++ )
1129 h->mb.cache.non_zero_count[x264_scan8[i]] = 16;
1133 /* save non zero count */
1134 CP32( &nnz[0*4], &h->mb.cache.non_zero_count[x264_scan8[0]+0*8] );
1135 CP32( &nnz[1*4], &h->mb.cache.non_zero_count[x264_scan8[0]+1*8] );
1136 CP32( &nnz[2*4], &h->mb.cache.non_zero_count[x264_scan8[0]+2*8] );
1137 CP32( &nnz[3*4], &h->mb.cache.non_zero_count[x264_scan8[0]+3*8] );
1138 M16( &nnz[16+0*2] ) = M32( &h->mb.cache.non_zero_count[x264_scan8[16+0*2]-1] ) >> 8;
1139 M16( &nnz[16+1*2] ) = M32( &h->mb.cache.non_zero_count[x264_scan8[16+1*2]-1] ) >> 8;
1140 M16( &nnz[16+2*2] ) = M32( &h->mb.cache.non_zero_count[x264_scan8[16+2*2]-1] ) >> 8;
1141 M16( &nnz[16+3*2] ) = M32( &h->mb.cache.non_zero_count[x264_scan8[16+3*2]-1] ) >> 8;
1143 if( h->mb.i_type != I_16x16 && h->mb.i_cbp_luma == 0 && h->mb.i_cbp_chroma == 0 )
1144 h->mb.i_qp = h->mb.i_last_qp;
1145 h->mb.qp[i_mb_xy] = h->mb.i_qp;
1146 h->mb.i_last_dqp = h->mb.i_qp - h->mb.i_last_qp;
1147 h->mb.i_last_qp = h->mb.i_qp;
1150 if( h->mb.i_cbp_luma == 0 && h->mb.i_type != I_8x8 )
1151 h->mb.b_transform_8x8 = 0;
1152 h->mb.mb_transform_size[i_mb_xy] = h->mb.b_transform_8x8;
1154 if( h->sh.i_type != SLICE_TYPE_I )
1156 int16_t (*mv0)[2] = &h->mb.mv[0][i_mb_4x4];
1157 int16_t (*mv1)[2] = &h->mb.mv[1][i_mb_4x4];
1158 int8_t *ref0 = &h->mb.ref[0][i_mb_8x8];
1159 int8_t *ref1 = &h->mb.ref[1][i_mb_8x8];
1160 if( !IS_INTRA( i_mb_type ) )
1162 ref0[0+0*s8x8] = h->mb.cache.ref[0][x264_scan8[0]];
1163 ref0[1+0*s8x8] = h->mb.cache.ref[0][x264_scan8[4]];
1164 ref0[0+1*s8x8] = h->mb.cache.ref[0][x264_scan8[8]];
1165 ref0[1+1*s8x8] = h->mb.cache.ref[0][x264_scan8[12]];
1166 CP128( &mv0[0*s4x4], h->mb.cache.mv[0][x264_scan8[0]+8*0] );
1167 CP128( &mv0[1*s4x4], h->mb.cache.mv[0][x264_scan8[0]+8*1] );
1168 CP128( &mv0[2*s4x4], h->mb.cache.mv[0][x264_scan8[0]+8*2] );
1169 CP128( &mv0[3*s4x4], h->mb.cache.mv[0][x264_scan8[0]+8*3] );
1170 if( h->sh.i_type == SLICE_TYPE_B )
1172 ref1[0+0*s8x8] = h->mb.cache.ref[1][x264_scan8[0]];
1173 ref1[1+0*s8x8] = h->mb.cache.ref[1][x264_scan8[4]];
1174 ref1[0+1*s8x8] = h->mb.cache.ref[1][x264_scan8[8]];
1175 ref1[1+1*s8x8] = h->mb.cache.ref[1][x264_scan8[12]];
1176 CP128( &mv1[0*s4x4], h->mb.cache.mv[1][x264_scan8[0]+8*0] );
1177 CP128( &mv1[1*s4x4], h->mb.cache.mv[1][x264_scan8[0]+8*1] );
1178 CP128( &mv1[2*s4x4], h->mb.cache.mv[1][x264_scan8[0]+8*2] );
1179 CP128( &mv1[3*s4x4], h->mb.cache.mv[1][x264_scan8[0]+8*3] );
1184 M16( &ref0[0*s8x8] ) = (uint8_t)(-1) * 0x0101;
1185 M16( &ref0[1*s8x8] ) = (uint8_t)(-1) * 0x0101;
1186 M128( &mv0[0*s4x4] ) = M128_ZERO;
1187 M128( &mv0[1*s4x4] ) = M128_ZERO;
1188 M128( &mv0[2*s4x4] ) = M128_ZERO;
1189 M128( &mv0[3*s4x4] ) = M128_ZERO;
1190 if( h->sh.i_type == SLICE_TYPE_B )
1192 M16( &ref1[0*s8x8] ) = (uint8_t)(-1) * 0x0101;
1193 M16( &ref1[1*s8x8] ) = (uint8_t)(-1) * 0x0101;
1194 M128( &mv1[0*s4x4] ) = M128_ZERO;
1195 M128( &mv1[1*s4x4] ) = M128_ZERO;
1196 M128( &mv1[2*s4x4] ) = M128_ZERO;
1197 M128( &mv1[3*s4x4] ) = M128_ZERO;
1202 if( h->param.b_cabac )
1204 uint8_t (*mvd0)[2] = h->mb.mvd[0][i_mb_xy];
1205 uint8_t (*mvd1)[2] = h->mb.mvd[1][i_mb_xy];
1206 if( IS_INTRA(i_mb_type) && i_mb_type != I_PCM )
1207 h->mb.chroma_pred_mode[i_mb_xy] = x264_mb_pred_mode8x8c_fix[ h->mb.i_chroma_pred_mode ];
1209 h->mb.chroma_pred_mode[i_mb_xy] = I_PRED_CHROMA_DC;
1211 if( (0x3FF30 >> i_mb_type) & 1 ) /* !INTRA && !SKIP && !DIRECT */
1213 CP64( mvd0[0], h->mb.cache.mvd[0][x264_scan8[10]] );
1214 CP16( mvd0[4], h->mb.cache.mvd[0][x264_scan8[5 ]] );
1215 CP16( mvd0[5], h->mb.cache.mvd[0][x264_scan8[7 ]] );
1216 CP16( mvd0[6], h->mb.cache.mvd[0][x264_scan8[13]] );
1217 if( h->sh.i_type == SLICE_TYPE_B )
1219 CP64( mvd1[0], h->mb.cache.mvd[1][x264_scan8[10]] );
1220 CP16( mvd1[4], h->mb.cache.mvd[1][x264_scan8[5 ]] );
1221 CP16( mvd1[5], h->mb.cache.mvd[1][x264_scan8[7 ]] );
1222 CP16( mvd1[6], h->mb.cache.mvd[1][x264_scan8[13]] );
1227 M128( mvd0[0] ) = M128_ZERO;
1228 if( h->sh.i_type == SLICE_TYPE_B )
1229 M128( mvd1[0] ) = M128_ZERO;
1232 if( h->sh.i_type == SLICE_TYPE_B )
1234 if( i_mb_type == B_SKIP || i_mb_type == B_DIRECT )
1235 h->mb.skipbp[i_mb_xy] = 0xf;
1236 else if( i_mb_type == B_8x8 )
1238 int skipbp = ( h->mb.i_sub_partition[0] == D_DIRECT_8x8 ) << 0;
1239 skipbp |= ( h->mb.i_sub_partition[1] == D_DIRECT_8x8 ) << 1;
1240 skipbp |= ( h->mb.i_sub_partition[2] == D_DIRECT_8x8 ) << 2;
1241 skipbp |= ( h->mb.i_sub_partition[3] == D_DIRECT_8x8 ) << 3;
1242 h->mb.skipbp[i_mb_xy] = skipbp;
1245 h->mb.skipbp[i_mb_xy] = 0;
1251 void x264_macroblock_bipred_init( x264_t *h )
1253 for( int field = 0; field <= h->sh.b_mbaff; field++ )
1254 for( int i_ref0 = 0; i_ref0 < (h->i_ref0<<h->sh.b_mbaff); i_ref0++ )
1256 int poc0 = h->fref0[i_ref0>>h->sh.b_mbaff]->i_poc;
1257 if( h->sh.b_mbaff && field^(i_ref0&1) )
1258 poc0 += h->sh.i_delta_poc_bottom;
1259 for( int i_ref1 = 0; i_ref1 < (h->i_ref1<<h->sh.b_mbaff); i_ref1++ )
1261 int dist_scale_factor;
1262 int poc1 = h->fref1[i_ref1>>h->sh.b_mbaff]->i_poc;
1263 if( h->sh.b_mbaff && field^(i_ref1&1) )
1264 poc1 += h->sh.i_delta_poc_bottom;
1265 int cur_poc = h->fdec->i_poc + field*h->sh.i_delta_poc_bottom;
1266 int td = x264_clip3( poc1 - poc0, -128, 127 );
1267 if( td == 0 /* || pic0 is a long-term ref */ )
1268 dist_scale_factor = 256;
1271 int tb = x264_clip3( cur_poc - poc0, -128, 127 );
1272 int tx = (16384 + (abs(td) >> 1)) / td;
1273 dist_scale_factor = x264_clip3( (tb * tx + 32) >> 6, -1024, 1023 );
1276 h->mb.dist_scale_factor_buf[field][i_ref0][i_ref1] = dist_scale_factor;
1278 dist_scale_factor >>= 2;
1279 if( h->param.analyse.b_weighted_bipred
1280 && dist_scale_factor >= -64
1281 && dist_scale_factor <= 128 )
1283 h->mb.bipred_weight_buf[field][i_ref0][i_ref1] = 64 - dist_scale_factor;
1284 // ssse3 implementation of biweight doesn't support the extrema.
1285 // if we ever generate them, we'll have to drop that optimization.
1286 assert( dist_scale_factor >= -63 && dist_scale_factor <= 127 );
1289 h->mb.bipred_weight_buf[field][i_ref0][i_ref1] = 32;