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( uint8_t, tmp0,[16*16] );
98 ALIGNED_ARRAY_16( uint8_t, 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->sps->i_mb_width;
220 h->mb.i_b8_stride = h->sps->i_mb_width * 2;
221 h->mb.i_b4_stride = h->sps->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++ )
254 CHECKED_MALLOC( h->mb.mvr[i][j], 2 * i_mb_count * sizeof(int16_t) );
257 if( h->param.analyse.i_weighted_pred )
259 int i_padv = PADV << h->param.b_interlaced;
260 #define ALIGN(x,a) (((x)+((a)-1))&~((a)-1))
261 int align = h->param.cpu&X264_CPU_CACHELINE_64 ? 64 : h->param.cpu&X264_CPU_CACHELINE_32 ? 32 : 16;
262 int i_stride, luma_plane_size = 0;
265 if( h->param.analyse.i_weighted_pred == X264_WEIGHTP_FAKE )
267 // only need buffer for lookahead
268 if( !h->param.i_sync_lookahead || h == h->thread[h->param.i_threads] )
270 // Fake analysis only works on lowres
271 i_stride = ALIGN( h->sps->i_mb_width*8 + 2*PADH, align );
272 luma_plane_size = i_stride * (h->sps->i_mb_height*8+2*i_padv);
273 // Only need 1 buffer for analysis
281 i_stride = ALIGN( h->sps->i_mb_width*16 + 2*PADH, align );
282 luma_plane_size = i_stride * (h->sps->i_mb_height*16+2*i_padv);
284 if( h->param.analyse.i_weighted_pred == X264_WEIGHTP_SMART )
285 //SMART can weight one ref and one offset -1
288 //blind only has one weighted copy (offset -1)
292 for( int i = 0; i < numweightbuf; i++ )
293 CHECKED_MALLOC( h->mb.p_weight_buf[i], luma_plane_size );
301 void x264_macroblock_cache_free( x264_t *h )
303 for( int i = 0; i < 2; i++ )
304 for( int j = !i; j < 32; j++ )
305 x264_free( h->mb.mvr[i][j] );
306 for( int i = 0; i < 16; i++ )
307 x264_free( h->mb.p_weight_buf[i] );
309 if( h->param.b_cabac )
311 x264_free( h->mb.chroma_pred_mode );
312 x264_free( h->mb.mvd[0] );
313 x264_free( h->mb.mvd[1] );
315 x264_free( h->mb.slice_table );
316 x264_free( h->mb.intra4x4_pred_mode );
317 x264_free( h->mb.non_zero_count );
318 x264_free( h->mb.mb_transform_size );
319 x264_free( h->mb.skipbp );
320 x264_free( h->mb.cbp );
321 x264_free( h->mb.qp );
324 int x264_macroblock_thread_allocate( x264_t *h, int b_lookahead )
327 for( int i = 0; i <= h->param.b_interlaced; i++ )
329 for( int j = 0; j < 3; j++ )
331 /* shouldn't really be initialized, just silences a valgrind false-positive in predict_8x8_filter_mmx */
332 CHECKED_MALLOCZERO( h->intra_border_backup[i][j], (h->sps->i_mb_width*16+32)>>!!j );
333 h->intra_border_backup[i][j] += 8;
335 CHECKED_MALLOC( h->deblock_strength[i], sizeof(**h->deblock_strength) * h->sps->i_mb_width );
338 /* Allocate scratch buffer */
339 int scratch_size = 0;
342 int buf_hpel = (h->thread[0]->fdec->i_width[0]+48) * sizeof(int16_t);
343 int buf_ssim = h->param.analyse.b_ssim * 8 * (h->param.i_width/4+3) * sizeof(int);
344 int me_range = X264_MIN(h->param.analyse.i_me_range, h->param.analyse.i_mv_range);
345 int buf_tesa = (h->param.analyse.i_me_method >= X264_ME_ESA) *
346 ((me_range*2+18) * sizeof(int16_t) + (me_range+4) * (me_range+1) * 4 * sizeof(mvsad_t));
347 scratch_size = X264_MAX3( buf_hpel, buf_ssim, buf_tesa );
349 int buf_mbtree = h->param.rc.b_mb_tree * ((h->sps->i_mb_width+3)&~3) * sizeof(int);
350 scratch_size = X264_MAX( scratch_size, buf_mbtree );
351 CHECKED_MALLOC( h->scratch_buffer, scratch_size );
358 void x264_macroblock_thread_free( x264_t *h, int b_lookahead )
361 for( int i = 0; i <= h->param.b_interlaced; i++ )
363 x264_free( h->deblock_strength[i] );
364 for( int j = 0; j < 3; j++ )
365 x264_free( h->intra_border_backup[i][j] - 8 );
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;
380 h->fdec->i_ref[0] = h->i_ref0;
381 h->fdec->i_ref[1] = h->i_ref1;
382 for( int i = 0; i < h->i_ref0; i++ )
383 h->fdec->ref_poc[0][i] = h->fref0[i]->i_poc;
384 if( h->sh.i_type == SLICE_TYPE_B )
386 for( int i = 0; i < h->i_ref1; i++ )
387 h->fdec->ref_poc[1][i] = h->fref1[i]->i_poc;
389 map_col_to_list0(-1) = -1;
390 map_col_to_list0(-2) = -2;
391 for( int i = 0; i < h->fref1[0]->i_ref[0]; i++ )
393 int poc = h->fref1[0]->ref_poc[0][i];
394 map_col_to_list0(i) = -2;
395 for( int j = 0; j < h->i_ref0; j++ )
396 if( h->fref0[j]->i_poc == poc )
398 map_col_to_list0(i) = j;
403 if( h->sh.i_type == SLICE_TYPE_P )
404 memset( h->mb.cache.skip, 0, sizeof( h->mb.cache.skip ) );
406 /* init with not available (for top right idx=7,15) */
407 memset( h->mb.cache.ref, -2, sizeof( h->mb.cache.ref ) );
410 for( int field = 0; field <= h->sh.b_mbaff; field++ )
412 int curpoc = h->fdec->i_poc + field*h->sh.i_delta_poc_bottom;
413 int refpoc = h->fref0[0]->i_poc;
414 if( h->sh.b_mbaff && field )
415 refpoc += h->sh.i_delta_poc_bottom;
416 int delta = curpoc - refpoc;
418 h->fdec->inv_ref_poc[field] = (256 + delta/2) / delta;
421 deblock_ref_table(-2) = -2;
422 deblock_ref_table(-1) = -1;
423 for( int i = 0; i < h->i_ref0 << h->sh.b_mbaff; i++ )
425 /* Mask off high bits to avoid frame num collisions with -1/-2.
426 * In current x264 frame num values don't cover a range of more
427 * than 32, so 6 bits is enough for uniqueness. */
428 if( !h->mb.b_interlaced )
429 deblock_ref_table(i) = h->fref0[i]->i_frame_num&63;
431 deblock_ref_table(i) = ((h->fref0[i>>1]->i_frame_num&63)<<1) + (i&1);
434 h->mb.i_neighbour4[6] =
435 h->mb.i_neighbour4[9] =
436 h->mb.i_neighbour4[12] =
437 h->mb.i_neighbour4[14] = MB_LEFT|MB_TOP|MB_TOPLEFT|MB_TOPRIGHT;
438 h->mb.i_neighbour4[3] =
439 h->mb.i_neighbour4[7] =
440 h->mb.i_neighbour4[11] =
441 h->mb.i_neighbour4[13] =
442 h->mb.i_neighbour4[15] =
443 h->mb.i_neighbour8[3] = MB_LEFT|MB_TOP|MB_TOPLEFT;
446 void x264_macroblock_thread_init( x264_t *h )
448 h->mb.i_me_method = h->param.analyse.i_me_method;
449 h->mb.i_subpel_refine = h->param.analyse.i_subpel_refine;
450 if( h->sh.i_type == SLICE_TYPE_B && (h->mb.i_subpel_refine == 6 || h->mb.i_subpel_refine == 8) )
451 h->mb.i_subpel_refine--;
452 h->mb.b_chroma_me = h->param.analyse.b_chroma_me && h->sh.i_type == SLICE_TYPE_P
453 && h->mb.i_subpel_refine >= 5;
454 h->mb.b_dct_decimate = h->sh.i_type == SLICE_TYPE_B ||
455 (h->param.analyse.b_dct_decimate && h->sh.i_type != SLICE_TYPE_I);
468 h->mb.pic.p_fenc[0] = h->mb.pic.fenc_buf;
469 h->mb.pic.p_fenc[1] = h->mb.pic.fenc_buf + 16*FENC_STRIDE;
470 h->mb.pic.p_fenc[2] = h->mb.pic.fenc_buf + 16*FENC_STRIDE + 8;
471 h->mb.pic.p_fdec[0] = h->mb.pic.fdec_buf + 2*FDEC_STRIDE;
472 h->mb.pic.p_fdec[1] = h->mb.pic.fdec_buf + 19*FDEC_STRIDE;
473 h->mb.pic.p_fdec[2] = h->mb.pic.fdec_buf + 19*FDEC_STRIDE + 16;
476 void x264_prefetch_fenc( x264_t *h, x264_frame_t *fenc, int i_mb_x, int i_mb_y )
478 int stride_y = fenc->i_stride[0];
479 int stride_uv = fenc->i_stride[1];
480 int off_y = 16 * (i_mb_x + i_mb_y * stride_y);
481 int off_uv = 8 * (i_mb_x + i_mb_y * stride_uv);
482 h->mc.prefetch_fenc( fenc->plane[0]+off_y, stride_y,
483 fenc->plane[1+(i_mb_x&1)]+off_uv, stride_uv, i_mb_x );
486 static NOINLINE void copy_column8( uint8_t *dst, uint8_t *src )
488 // input pointers are offset by 4 rows because that's faster (smaller instruction size on x86)
489 for( int i = -4; i < 4; i++ )
490 dst[i*FDEC_STRIDE] = src[i*FDEC_STRIDE];
493 static void ALWAYS_INLINE x264_macroblock_load_pic_pointers( x264_t *h, int mb_x, int mb_y, int i )
495 const int w = (i == 0 ? 16 : 8);
496 const int i_stride = h->fdec->i_stride[!!i];
497 const int i_stride2 = i_stride << h->mb.b_interlaced;
498 const int i_pix_offset = h->mb.b_interlaced
499 ? w * (mb_x + (mb_y&~1) * i_stride) + (mb_y&1) * i_stride
500 : w * (mb_x + mb_y * i_stride);
501 const uint8_t *plane_fdec = &h->fdec->plane[i][i_pix_offset];
502 const uint8_t *intra_fdec = &h->intra_border_backup[mb_y & h->sh.b_mbaff][i][mb_x*16>>!!i];
503 int ref_pix_offset[2] = { i_pix_offset, i_pix_offset };
504 x264_frame_t **fref[2] = { h->fref0, h->fref1 };
505 if( h->mb.b_interlaced )
506 ref_pix_offset[1] += (1-2*(mb_y&1)) * i_stride;
507 h->mb.pic.i_stride[i] = i_stride2;
508 h->mb.pic.p_fenc_plane[i] = &h->fenc->plane[i][i_pix_offset];
509 h->mc.copy[i?PIXEL_8x8:PIXEL_16x16]( h->mb.pic.p_fenc[i], FENC_STRIDE,
510 h->mb.pic.p_fenc_plane[i], i_stride2, w );
511 memcpy( &h->mb.pic.p_fdec[i][-1-FDEC_STRIDE], intra_fdec-1, w*3/2+1 );
512 if( h->mb.b_interlaced )
513 for( int j = 0; j < w; j++ )
514 h->mb.pic.p_fdec[i][-1+j*FDEC_STRIDE] = plane_fdec[-1+j*i_stride2];
515 for( int j = 0; j < h->mb.pic.i_fref[0]; j++ )
517 h->mb.pic.p_fref[0][j][i==0 ? 0:i+3] = &fref[0][j >> h->mb.b_interlaced]->plane[i][ref_pix_offset[j&1]];
520 for( int k = 1; k < 4; k++ )
521 h->mb.pic.p_fref[0][j][k] = &fref[0][j >> h->mb.b_interlaced]->filtered[k][ref_pix_offset[j&1]];
522 if( h->sh.weight[j][0].weightfn )
523 h->mb.pic.p_fref_w[j] = &h->fenc->weighted[j >> h->mb.b_interlaced][ref_pix_offset[j&1]];
525 h->mb.pic.p_fref_w[j] = h->mb.pic.p_fref[0][j][0];
528 if( h->sh.i_type == SLICE_TYPE_B )
529 for( int j = 0; j < h->mb.pic.i_fref[1]; j++ )
531 h->mb.pic.p_fref[1][j][i==0 ? 0:i+3] = &fref[1][j >> h->mb.b_interlaced]->plane[i][ref_pix_offset[j&1]];
533 for( int k = 1; k < 4; k++ )
534 h->mb.pic.p_fref[1][j][k] = &fref[1][j >> h->mb.b_interlaced]->filtered[k][ref_pix_offset[j&1]];
538 static void inline x264_macroblock_cache_load_neighbours( x264_t *h, int mb_x, int mb_y )
540 int top = (mb_y - (1 << h->mb.b_interlaced)) * h->mb.i_mb_stride + mb_x;
544 h->mb.i_mb_xy = mb_y * h->mb.i_mb_stride + mb_x;
545 h->mb.i_b8_xy = 2*(mb_y * h->mb.i_b8_stride + mb_x);
546 h->mb.i_b4_xy = 4*(mb_y * h->mb.i_b4_stride + mb_x);
547 h->mb.i_neighbour = 0;
548 h->mb.i_neighbour_intra = 0;
549 h->mb.i_neighbour_frame = 0;
550 h->mb.i_mb_top_xy = -1;
551 h->mb.i_mb_left_xy = -1;
552 h->mb.i_mb_topleft_xy = -1;
553 h->mb.i_mb_topright_xy = -1;
554 h->mb.i_mb_type_top = -1;
555 h->mb.i_mb_type_left = -1;
556 h->mb.i_mb_type_topleft = -1;
557 h->mb.i_mb_type_topright = -1;
561 h->mb.i_neighbour_frame |= MB_LEFT;
562 h->mb.i_mb_left_xy = h->mb.i_mb_xy - 1;
563 h->mb.i_mb_type_left = h->mb.type[h->mb.i_mb_left_xy];
564 if( h->mb.i_mb_xy > h->sh.i_first_mb )
566 h->mb.i_neighbour |= MB_LEFT;
568 if( !h->param.b_constrained_intra || IS_INTRA( h->mb.i_mb_type_left ) )
569 h->mb.i_neighbour_intra |= MB_LEFT;
573 /* We can't predict from the previous threadslice since it hasn't been encoded yet. */
574 if( (h->i_threadslice_start >> h->mb.b_interlaced) != (mb_y >> h->mb.b_interlaced) )
578 h->mb.i_neighbour_frame |= MB_TOP;
579 h->mb.i_mb_top_xy = top;
580 h->mb.i_mb_type_top = h->mb.type[h->mb.i_mb_top_xy];
581 if( top >= h->sh.i_first_mb )
583 h->mb.i_neighbour |= MB_TOP;
585 if( !h->param.b_constrained_intra || IS_INTRA( h->mb.i_mb_type_top ) )
586 h->mb.i_neighbour_intra |= MB_TOP;
588 /* We only need to prefetch the top blocks because the left was just written
589 * to as part of the previous cache_save. Since most target CPUs use write-allocate
590 * caches, left blocks are near-guaranteed to be in L1 cache. Top--not so much. */
591 x264_prefetch( &h->mb.cbp[top] );
592 x264_prefetch( h->mb.intra4x4_pred_mode[top] );
593 x264_prefetch( &h->mb.non_zero_count[top][12] );
594 /* These aren't always allocated, but prefetching an invalid address can't hurt. */
595 x264_prefetch( &h->mb.mb_transform_size[top] );
596 x264_prefetch( &h->mb.skipbp[top] );
600 if( mb_x > 0 && top - 1 >= 0 )
602 h->mb.i_neighbour_frame |= MB_TOPLEFT;
603 h->mb.i_mb_topleft_xy = top - 1;
604 h->mb.i_mb_type_topleft = h->mb.type[h->mb.i_mb_topleft_xy];
605 if( top - 1 >= h->sh.i_first_mb )
607 h->mb.i_neighbour |= MB_TOPLEFT;
609 if( !h->param.b_constrained_intra || IS_INTRA( h->mb.i_mb_type_topleft ) )
610 h->mb.i_neighbour_intra |= MB_TOPLEFT;
614 if( mb_x < h->sps->i_mb_width - 1 && top + 1 >= 0 )
616 h->mb.i_neighbour_frame |= MB_TOPRIGHT;
617 h->mb.i_mb_topright_xy = top + 1;
618 h->mb.i_mb_type_topright = h->mb.type[h->mb.i_mb_topright_xy];
619 if( top + 1 >= h->sh.i_first_mb )
621 h->mb.i_neighbour |= MB_TOPRIGHT;
623 if( !h->param.b_constrained_intra || IS_INTRA( h->mb.i_mb_type_topright ) )
624 h->mb.i_neighbour_intra |= MB_TOPRIGHT;
630 void x264_macroblock_cache_load( x264_t *h, int mb_x, int mb_y )
632 x264_macroblock_cache_load_neighbours( h, mb_x, mb_y );
634 int left = h->mb.i_mb_left_xy;
635 int top = h->mb.i_mb_top_xy;
636 int top_y = mb_y - (1 << h->mb.b_interlaced);
637 int top_8x8 = (2*top_y+1) * h->mb.i_b8_stride + 2*mb_x;
638 int top_4x4 = (4*top_y+3) * h->mb.i_b4_stride + 4*mb_x;
640 /* GCC pessimizes direct loads from heap-allocated arrays due to aliasing. */
641 /* By only dereferencing them once, we avoid this issue. */
642 int8_t (*i4x4)[8] = h->mb.intra4x4_pred_mode;
643 uint8_t (*nnz)[24] = h->mb.non_zero_count;
644 int16_t *cbp = h->mb.cbp;
647 if( h->mb.i_neighbour & MB_TOP )
649 h->mb.cache.i_cbp_top = cbp[top];
651 CP32( &h->mb.cache.intra4x4_pred_mode[x264_scan8[0] - 8], &i4x4[top][0] );
653 /* load non_zero_count */
654 CP32( &h->mb.cache.non_zero_count[x264_scan8[0] - 8], &nnz[top][12] );
655 /* shift because x264_scan8[16] is misaligned */
656 M32( &h->mb.cache.non_zero_count[x264_scan8[16+0] - 9] ) = M16( &nnz[top][18] ) << 8;
657 M32( &h->mb.cache.non_zero_count[x264_scan8[16+4] - 9] ) = M16( &nnz[top][22] ) << 8;
661 h->mb.cache.i_cbp_top = -1;
664 M32( &h->mb.cache.intra4x4_pred_mode[x264_scan8[0] - 8] ) = 0xFFFFFFFFU;
666 /* load non_zero_count */
667 M32( &h->mb.cache.non_zero_count[x264_scan8[ 0] - 8] ) = 0x80808080U;
668 M32( &h->mb.cache.non_zero_count[x264_scan8[16+0] - 9] ) = 0x80808080U;
669 M32( &h->mb.cache.non_zero_count[x264_scan8[16+4] - 9] ) = 0x80808080U;
672 if( h->mb.i_neighbour & MB_LEFT )
674 h->mb.cache.i_cbp_left = cbp[left];
677 h->mb.cache.intra4x4_pred_mode[x264_scan8[0 ] - 1] = i4x4[left][4];
678 h->mb.cache.intra4x4_pred_mode[x264_scan8[2 ] - 1] = i4x4[left][5];
679 h->mb.cache.intra4x4_pred_mode[x264_scan8[8 ] - 1] = i4x4[left][6];
680 h->mb.cache.intra4x4_pred_mode[x264_scan8[10] - 1] = i4x4[left][3];
682 /* load non_zero_count */
683 h->mb.cache.non_zero_count[x264_scan8[0 ] - 1] = nnz[left][3];
684 h->mb.cache.non_zero_count[x264_scan8[2 ] - 1] = nnz[left][7];
685 h->mb.cache.non_zero_count[x264_scan8[8 ] - 1] = nnz[left][11];
686 h->mb.cache.non_zero_count[x264_scan8[10] - 1] = nnz[left][15];
688 h->mb.cache.non_zero_count[x264_scan8[16+0] - 1] = nnz[left][16+1];
689 h->mb.cache.non_zero_count[x264_scan8[16+2] - 1] = nnz[left][16+3];
691 h->mb.cache.non_zero_count[x264_scan8[16+4+0] - 1] = nnz[left][16+4+1];
692 h->mb.cache.non_zero_count[x264_scan8[16+4+2] - 1] = nnz[left][16+4+3];
694 /* Finish the prefetching */
695 if( h->sh.i_type != SLICE_TYPE_I )
696 for( int l = 0; l < (h->sh.i_type == SLICE_TYPE_B) + 1; l++ )
698 x264_prefetch( &h->mb.mv[l][top_4x4-1] );
699 /* Top right being not in the same cacheline as top left will happen
700 * once every 4 MBs, so one extra prefetch is worthwhile */
701 x264_prefetch( &h->mb.mv[l][top_4x4+4] );
702 x264_prefetch( &h->mb.ref[l][top_8x8-1] );
703 x264_prefetch( &h->mb.mvd[l][top] );
708 h->mb.cache.i_cbp_left = -1;
710 h->mb.cache.intra4x4_pred_mode[x264_scan8[0 ] - 1] =
711 h->mb.cache.intra4x4_pred_mode[x264_scan8[2 ] - 1] =
712 h->mb.cache.intra4x4_pred_mode[x264_scan8[8 ] - 1] =
713 h->mb.cache.intra4x4_pred_mode[x264_scan8[10] - 1] = -1;
715 /* load non_zero_count */
716 h->mb.cache.non_zero_count[x264_scan8[0 ] - 1] =
717 h->mb.cache.non_zero_count[x264_scan8[2 ] - 1] =
718 h->mb.cache.non_zero_count[x264_scan8[8 ] - 1] =
719 h->mb.cache.non_zero_count[x264_scan8[10] - 1] =
720 h->mb.cache.non_zero_count[x264_scan8[16+0] - 1] =
721 h->mb.cache.non_zero_count[x264_scan8[16+2] - 1] =
722 h->mb.cache.non_zero_count[x264_scan8[16+4+0] - 1] =
723 h->mb.cache.non_zero_count[x264_scan8[16+4+2] - 1] = 0x80;
726 if( h->pps->b_transform_8x8_mode )
728 h->mb.cache.i_neighbour_transform_size =
729 ( (h->mb.i_neighbour & MB_LEFT) && h->mb.mb_transform_size[left] )
730 + ( (h->mb.i_neighbour & MB_TOP) && h->mb.mb_transform_size[top] );
735 h->mb.pic.i_fref[0] = h->i_ref0 << h->mb.b_interlaced;
736 h->mb.pic.i_fref[1] = h->i_ref1 << h->mb.b_interlaced;
737 h->mb.cache.i_neighbour_interlaced =
738 !!(h->mb.i_neighbour & MB_LEFT)
739 + !!(h->mb.i_neighbour & MB_TOP);
742 if( !h->mb.b_interlaced )
744 copy_column8( h->mb.pic.p_fdec[0]-1+ 4*FDEC_STRIDE, h->mb.pic.p_fdec[0]+15+ 4*FDEC_STRIDE );
745 copy_column8( h->mb.pic.p_fdec[0]-1+12*FDEC_STRIDE, h->mb.pic.p_fdec[0]+15+12*FDEC_STRIDE );
746 copy_column8( h->mb.pic.p_fdec[1]-1+ 4*FDEC_STRIDE, h->mb.pic.p_fdec[1]+ 7+ 4*FDEC_STRIDE );
747 copy_column8( h->mb.pic.p_fdec[2]-1+ 4*FDEC_STRIDE, h->mb.pic.p_fdec[2]+ 7+ 4*FDEC_STRIDE );
750 /* load picture pointers */
751 x264_macroblock_load_pic_pointers( h, mb_x, mb_y, 0 );
752 x264_macroblock_load_pic_pointers( h, mb_x, mb_y, 1 );
753 x264_macroblock_load_pic_pointers( h, mb_x, mb_y, 2 );
755 if( h->fdec->integral )
757 int offset = 16 * (mb_x + mb_y * h->fdec->i_stride[0]);
758 for( int i = 0; i < h->mb.pic.i_fref[0]; i++ )
759 h->mb.pic.p_integral[0][i] = &h->fref0[i]->integral[offset];
760 for( int i = 0; i < h->mb.pic.i_fref[1]; i++ )
761 h->mb.pic.p_integral[1][i] = &h->fref1[i]->integral[offset];
764 x264_prefetch_fenc( h, h->fenc, mb_x, mb_y );
766 /* load ref/mv/mvd */
767 if( h->sh.i_type != SLICE_TYPE_I )
769 int s8x8 = h->mb.i_b8_stride;
770 int s4x4 = h->mb.i_b4_stride;
772 for( int l = 0; l < (h->sh.i_type == SLICE_TYPE_B) + 1; l++ )
774 int16_t (*mv)[2] = h->mb.mv[l];
775 int8_t *ref = h->mb.ref[l];
777 int i8 = x264_scan8[0] - 1 - 1*8;
778 if( h->mb.i_neighbour & MB_TOPLEFT )
780 h->mb.cache.ref[l][i8] = ref[top_8x8 - 1];
781 CP32( h->mb.cache.mv[l][i8], mv[top_4x4 - 1] );
785 h->mb.cache.ref[l][i8] = -2;
786 M32( h->mb.cache.mv[l][i8] ) = 0;
789 i8 = x264_scan8[0] - 8;
790 if( h->mb.i_neighbour & MB_TOP )
792 h->mb.cache.ref[l][i8+0] =
793 h->mb.cache.ref[l][i8+1] = ref[top_8x8 + 0];
794 h->mb.cache.ref[l][i8+2] =
795 h->mb.cache.ref[l][i8+3] = ref[top_8x8 + 1];
796 CP128( h->mb.cache.mv[l][i8], mv[top_4x4] );
800 M128( h->mb.cache.mv[l][i8] ) = M128_ZERO;
801 M32( &h->mb.cache.ref[l][i8] ) = (uint8_t)(-2) * 0x01010101U;
804 i8 = x264_scan8[0] + 4 - 1*8;
805 if( h->mb.i_neighbour & MB_TOPRIGHT )
807 h->mb.cache.ref[l][i8] = ref[top_8x8 + 2];
808 CP32( h->mb.cache.mv[l][i8], mv[top_4x4 + 4] );
811 h->mb.cache.ref[l][i8] = -2;
813 i8 = x264_scan8[0] - 1;
814 if( h->mb.i_neighbour & MB_LEFT )
816 const int ir = h->mb.i_b8_xy - 1;
817 const int iv = h->mb.i_b4_xy - 1;
818 h->mb.cache.ref[l][i8+0*8] =
819 h->mb.cache.ref[l][i8+1*8] = ref[ir + 0*s8x8];
820 h->mb.cache.ref[l][i8+2*8] =
821 h->mb.cache.ref[l][i8+3*8] = ref[ir + 1*s8x8];
823 CP32( h->mb.cache.mv[l][i8+0*8], mv[iv + 0*s4x4] );
824 CP32( h->mb.cache.mv[l][i8+1*8], mv[iv + 1*s4x4] );
825 CP32( h->mb.cache.mv[l][i8+2*8], mv[iv + 2*s4x4] );
826 CP32( h->mb.cache.mv[l][i8+3*8], mv[iv + 3*s4x4] );
830 for( int i = 0; i < 4; i++ )
832 h->mb.cache.ref[l][i8+i*8] = -2;
833 M32( h->mb.cache.mv[l][i8+i*8] ) = 0;
837 if( h->param.b_cabac )
839 uint8_t (*mvd)[8][2] = h->mb.mvd[l];
840 if( h->mb.i_neighbour & MB_TOP )
841 CP64( h->mb.cache.mvd[l][x264_scan8[0] - 8], mvd[top][0] );
843 M64( h->mb.cache.mvd[l][x264_scan8[0] - 8] ) = 0;
845 if( h->mb.i_neighbour & MB_LEFT )
847 CP16( h->mb.cache.mvd[l][x264_scan8[0 ] - 1], mvd[left][4] );
848 CP16( h->mb.cache.mvd[l][x264_scan8[2 ] - 1], mvd[left][5] );
849 CP16( h->mb.cache.mvd[l][x264_scan8[8 ] - 1], mvd[left][6] );
850 CP16( h->mb.cache.mvd[l][x264_scan8[10] - 1], mvd[left][3] );
853 for( int i = 0; i < 4; i++ )
854 M16( h->mb.cache.mvd[l][x264_scan8[0]-1+i*8] ) = 0;
859 if( h->sh.i_type == SLICE_TYPE_B )
861 h->mb.bipred_weight = h->mb.bipred_weight_buf[h->mb.b_interlaced&(mb_y&1)];
862 h->mb.dist_scale_factor = h->mb.dist_scale_factor_buf[h->mb.b_interlaced&(mb_y&1)];
863 if( h->param.b_cabac )
866 x264_macroblock_cache_skip( h, 0, 0, 4, 4, 0 );
867 skipbp = (h->mb.i_neighbour & MB_LEFT) ? h->mb.skipbp[left] : 0;
868 h->mb.cache.skip[x264_scan8[0] - 1] = skipbp & 0x2;
869 h->mb.cache.skip[x264_scan8[8] - 1] = skipbp & 0x8;
870 skipbp = (h->mb.i_neighbour & MB_TOP) ? h->mb.skipbp[top] : 0;
871 h->mb.cache.skip[x264_scan8[0] - 8] = skipbp & 0x4;
872 h->mb.cache.skip[x264_scan8[4] - 8] = skipbp & 0x8;
876 if( h->sh.i_type == SLICE_TYPE_P )
877 x264_mb_predict_mv_pskip( h, h->mb.cache.pskip_mv );
880 h->mb.i_neighbour4[0] =
881 h->mb.i_neighbour8[0] = (h->mb.i_neighbour_intra & (MB_TOP|MB_LEFT|MB_TOPLEFT))
882 | ((h->mb.i_neighbour_intra & MB_TOP) ? MB_TOPRIGHT : 0);
883 h->mb.i_neighbour4[4] =
884 h->mb.i_neighbour4[1] = MB_LEFT | ((h->mb.i_neighbour_intra & MB_TOP) ? (MB_TOP|MB_TOPLEFT|MB_TOPRIGHT) : 0);
885 h->mb.i_neighbour4[2] =
886 h->mb.i_neighbour4[8] =
887 h->mb.i_neighbour4[10] =
888 h->mb.i_neighbour8[2] = MB_TOP|MB_TOPRIGHT | ((h->mb.i_neighbour_intra & MB_LEFT) ? (MB_LEFT|MB_TOPLEFT) : 0);
889 h->mb.i_neighbour4[5] =
890 h->mb.i_neighbour8[1] = MB_LEFT | (h->mb.i_neighbour_intra & MB_TOPRIGHT)
891 | ((h->mb.i_neighbour_intra & MB_TOP) ? MB_TOP|MB_TOPLEFT : 0);
894 void x264_macroblock_cache_load_neighbours_deblock( x264_t *h, int mb_x, int mb_y )
896 int deblock_on_slice_edges = h->sh.i_disable_deblocking_filter_idc != 2;
897 int top = (mb_y - (1 << h->mb.b_interlaced)) * h->mb.i_mb_stride + mb_x;
899 h->mb.i_neighbour = 0;
900 h->mb.i_mb_xy = mb_y * h->mb.i_mb_stride + mb_x;
904 h->mb.i_mb_left_xy = h->mb.i_mb_xy - 1;
905 if( deblock_on_slice_edges || h->mb.slice_table[h->mb.i_mb_left_xy] == h->mb.slice_table[h->mb.i_mb_xy] )
906 h->mb.i_neighbour |= MB_LEFT;
911 h->mb.i_mb_top_xy = top;
912 if( deblock_on_slice_edges || h->mb.slice_table[h->mb.i_mb_top_xy] == h->mb.slice_table[h->mb.i_mb_xy] )
913 h->mb.i_neighbour |= MB_TOP;
917 void x264_macroblock_cache_load_deblock( x264_t *h )
919 if( IS_INTRA( h->mb.type[h->mb.i_mb_xy] ) )
922 /* If we have multiple slices and we're deblocking on slice edges, we
923 * have to reload neighbour data. */
924 if( h->sh.i_first_mb && h->sh.i_disable_deblocking_filter_idc != 2 )
926 int old_neighbour = h->mb.i_neighbour;
927 int mb_x = h->mb.i_mb_x;
928 int mb_y = h->mb.i_mb_y;
929 x264_macroblock_cache_load_neighbours_deblock( h, mb_x, mb_y );
930 h->mb.i_neighbour &= ~old_neighbour;
931 if( h->mb.i_neighbour )
933 int left = h->mb.i_mb_left_xy;
934 int top = h->mb.i_mb_top_xy;
935 int top_y = mb_y - (1 << h->mb.b_interlaced);
936 int top_8x8 = (2*top_y+1) * h->mb.i_b8_stride + 2*mb_x;
937 int top_4x4 = (4*top_y+3) * h->mb.i_b4_stride + 4*mb_x;
938 int s8x8 = h->mb.i_b8_stride;
939 int s4x4 = h->mb.i_b4_stride;
941 uint8_t (*nnz)[24] = h->mb.non_zero_count;
943 if( h->mb.i_neighbour & MB_TOP )
944 CP32( &h->mb.cache.non_zero_count[x264_scan8[0] - 8], &nnz[top][12] );
946 if( h->mb.i_neighbour & MB_LEFT )
948 h->mb.cache.non_zero_count[x264_scan8[0 ] - 1] = nnz[left][3];
949 h->mb.cache.non_zero_count[x264_scan8[2 ] - 1] = nnz[left][7];
950 h->mb.cache.non_zero_count[x264_scan8[8 ] - 1] = nnz[left][11];
951 h->mb.cache.non_zero_count[x264_scan8[10] - 1] = nnz[left][15];
954 for( int l = 0; l <= (h->sh.i_type == SLICE_TYPE_B); l++ )
956 int16_t (*mv)[2] = h->mb.mv[l];
957 int8_t *ref = h->mb.ref[l];
959 int i8 = x264_scan8[0] - 8;
960 if( h->mb.i_neighbour & MB_TOP )
962 h->mb.cache.ref[l][i8+0] =
963 h->mb.cache.ref[l][i8+1] = ref[top_8x8 + 0];
964 h->mb.cache.ref[l][i8+2] =
965 h->mb.cache.ref[l][i8+3] = ref[top_8x8 + 1];
966 CP128( h->mb.cache.mv[l][i8], mv[top_4x4] );
969 i8 = x264_scan8[0] - 1;
970 if( h->mb.i_neighbour & MB_LEFT )
972 int ir = h->mb.i_b8_xy - 1;
973 int iv = h->mb.i_b4_xy - 1;
974 h->mb.cache.ref[l][i8+0*8] =
975 h->mb.cache.ref[l][i8+1*8] = ref[ir + 0*s8x8];
976 h->mb.cache.ref[l][i8+2*8] =
977 h->mb.cache.ref[l][i8+3*8] = ref[ir + 1*s8x8];
979 CP32( h->mb.cache.mv[l][i8+0*8], mv[iv + 0*s4x4] );
980 CP32( h->mb.cache.mv[l][i8+1*8], mv[iv + 1*s4x4] );
981 CP32( h->mb.cache.mv[l][i8+2*8], mv[iv + 2*s4x4] );
982 CP32( h->mb.cache.mv[l][i8+3*8], mv[iv + 3*s4x4] );
988 if( h->param.analyse.i_weighted_pred && h->sh.i_type == SLICE_TYPE_P )
990 /* Handle reference frame duplicates */
991 int i8 = x264_scan8[0] - 8;
992 h->mb.cache.ref[0][i8+0] =
993 h->mb.cache.ref[0][i8+1] = deblock_ref_table(h->mb.cache.ref[0][i8+0]);
994 h->mb.cache.ref[0][i8+2] =
995 h->mb.cache.ref[0][i8+3] = deblock_ref_table(h->mb.cache.ref[0][i8+2]);
997 i8 = x264_scan8[0] - 1;
998 h->mb.cache.ref[0][i8+0*8] =
999 h->mb.cache.ref[0][i8+1*8] = deblock_ref_table(h->mb.cache.ref[0][i8+0*8]);
1000 h->mb.cache.ref[0][i8+2*8] =
1001 h->mb.cache.ref[0][i8+3*8] = deblock_ref_table(h->mb.cache.ref[0][i8+2*8]);
1003 int ref0 = deblock_ref_table(h->mb.cache.ref[0][x264_scan8[ 0]]);
1004 int ref1 = deblock_ref_table(h->mb.cache.ref[0][x264_scan8[ 4]]);
1005 int ref2 = deblock_ref_table(h->mb.cache.ref[0][x264_scan8[ 8]]);
1006 int ref3 = deblock_ref_table(h->mb.cache.ref[0][x264_scan8[12]]);
1007 uint32_t reftop = pack16to32( (uint8_t)ref0, (uint8_t)ref1 ) * 0x0101;
1008 uint32_t refbot = pack16to32( (uint8_t)ref2, (uint8_t)ref3 ) * 0x0101;
1010 M32( &h->mb.cache.ref[0][x264_scan8[0]+8*0] ) = reftop;
1011 M32( &h->mb.cache.ref[0][x264_scan8[0]+8*1] ) = reftop;
1012 M32( &h->mb.cache.ref[0][x264_scan8[0]+8*2] ) = refbot;
1013 M32( &h->mb.cache.ref[0][x264_scan8[0]+8*3] ) = refbot;
1016 /* Munge NNZ for cavlc + 8x8dct */
1017 if( !h->param.b_cabac && h->pps->b_transform_8x8_mode )
1019 uint8_t (*nnz)[24] = h->mb.non_zero_count;
1020 int top = h->mb.i_mb_top_xy;
1021 int left = h->mb.i_mb_left_xy;
1023 if( (h->mb.i_neighbour & MB_TOP) && h->mb.mb_transform_size[top] )
1025 int i8 = x264_scan8[0] - 8;
1026 int nnz_top0 = M16( &nnz[top][8] ) | M16( &nnz[top][12] );
1027 int nnz_top1 = M16( &nnz[top][10] ) | M16( &nnz[top][14] );
1028 M16( &h->mb.cache.non_zero_count[i8+0] ) = nnz_top0 ? 0x0101 : 0;
1029 M16( &h->mb.cache.non_zero_count[i8+2] ) = nnz_top1 ? 0x0101 : 0;
1032 if( (h->mb.i_neighbour & MB_LEFT) && h->mb.mb_transform_size[left] )
1034 int i8 = x264_scan8[0] - 1;
1035 int nnz_left0 = M16( &nnz[left][2] ) | M16( &nnz[left][6] );
1036 int nnz_left1 = M16( &nnz[left][10] ) | M16( &nnz[left][14] );
1037 h->mb.cache.non_zero_count[i8+8*0] = !!nnz_left0;
1038 h->mb.cache.non_zero_count[i8+8*1] = !!nnz_left0;
1039 h->mb.cache.non_zero_count[i8+8*2] = !!nnz_left1;
1040 h->mb.cache.non_zero_count[i8+8*3] = !!nnz_left1;
1043 if( h->mb.mb_transform_size[h->mb.i_mb_xy] )
1045 int nnz0 = M16( &h->mb.cache.non_zero_count[x264_scan8[ 0]] ) | M16( &h->mb.cache.non_zero_count[x264_scan8[ 2]] );
1046 int nnz1 = M16( &h->mb.cache.non_zero_count[x264_scan8[ 4]] ) | M16( &h->mb.cache.non_zero_count[x264_scan8[ 6]] );
1047 int nnz2 = M16( &h->mb.cache.non_zero_count[x264_scan8[ 8]] ) | M16( &h->mb.cache.non_zero_count[x264_scan8[10]] );
1048 int nnz3 = M16( &h->mb.cache.non_zero_count[x264_scan8[12]] ) | M16( &h->mb.cache.non_zero_count[x264_scan8[14]] );
1049 uint32_t nnztop = pack16to32( !!nnz0, !!nnz1 ) * 0x0101;
1050 uint32_t nnzbot = pack16to32( !!nnz2, !!nnz3 ) * 0x0101;
1052 M32( &h->mb.cache.non_zero_count[x264_scan8[0]+8*0] ) = nnztop;
1053 M32( &h->mb.cache.non_zero_count[x264_scan8[0]+8*1] ) = nnztop;
1054 M32( &h->mb.cache.non_zero_count[x264_scan8[0]+8*2] ) = nnzbot;
1055 M32( &h->mb.cache.non_zero_count[x264_scan8[0]+8*3] ) = nnzbot;
1060 static void ALWAYS_INLINE x264_macroblock_store_pic( x264_t *h, int i )
1063 int i_stride = h->fdec->i_stride[!!i];
1064 int i_stride2 = i_stride << h->mb.b_interlaced;
1065 int i_pix_offset = h->mb.b_interlaced
1066 ? w * (h->mb.i_mb_x + (h->mb.i_mb_y&~1) * i_stride) + (h->mb.i_mb_y&1) * i_stride
1067 : w * (h->mb.i_mb_x + h->mb.i_mb_y * i_stride);
1068 h->mc.copy[i?PIXEL_8x8:PIXEL_16x16]( &h->fdec->plane[i][i_pix_offset], i_stride2,
1069 h->mb.pic.p_fdec[i], FDEC_STRIDE, w );
1072 void x264_macroblock_cache_save( x264_t *h )
1074 const int i_mb_xy = h->mb.i_mb_xy;
1075 const int i_mb_type = x264_mb_type_fix[h->mb.i_type];
1076 const int s8x8 = h->mb.i_b8_stride;
1077 const int s4x4 = h->mb.i_b4_stride;
1078 const int i_mb_4x4 = h->mb.i_b4_xy;
1079 const int i_mb_8x8 = h->mb.i_b8_xy;
1081 /* GCC pessimizes direct stores to heap-allocated arrays due to aliasing. */
1082 /* By only dereferencing them once, we avoid this issue. */
1083 int8_t *i4x4 = h->mb.intra4x4_pred_mode[i_mb_xy];
1084 uint8_t *nnz = h->mb.non_zero_count[i_mb_xy];
1086 x264_macroblock_store_pic( h, 0 );
1087 x264_macroblock_store_pic( h, 1 );
1088 x264_macroblock_store_pic( h, 2 );
1090 x264_prefetch_fenc( h, h->fdec, h->mb.i_mb_x, h->mb.i_mb_y );
1092 h->mb.type[i_mb_xy] = i_mb_type;
1093 h->mb.slice_table[i_mb_xy] = h->sh.i_first_mb;
1094 h->mb.partition[i_mb_xy] = IS_INTRA( i_mb_type ) ? D_16x16 : h->mb.i_partition;
1095 h->mb.i_mb_prev_xy = i_mb_xy;
1098 if( i_mb_type == I_4x4 )
1100 CP32( &i4x4[0], &h->mb.cache.intra4x4_pred_mode[x264_scan8[10]] );
1101 M32( &i4x4[4] ) = pack8to32( h->mb.cache.intra4x4_pred_mode[x264_scan8[5] ],
1102 h->mb.cache.intra4x4_pred_mode[x264_scan8[7] ],
1103 h->mb.cache.intra4x4_pred_mode[x264_scan8[13] ], 0);
1105 else if( !h->param.b_constrained_intra || IS_INTRA(i_mb_type) )
1106 M64( i4x4 ) = I_PRED_4x4_DC * 0x0101010101010101ULL;
1108 M64( i4x4 ) = (uint8_t)(-1) * 0x0101010101010101ULL;
1111 if( i_mb_type == I_PCM )
1113 h->mb.qp[i_mb_xy] = 0;
1114 h->mb.i_last_dqp = 0;
1115 h->mb.i_cbp_chroma = 2;
1116 h->mb.i_cbp_luma = 0xf;
1117 h->mb.cbp[i_mb_xy] = 0x72f; /* all set */
1118 h->mb.b_transform_8x8 = 0;
1119 memset( nnz, 16, sizeof( *h->mb.non_zero_count ) );
1120 for( int i = 0; i < 24; i++ )
1121 h->mb.cache.non_zero_count[x264_scan8[i]] = 16;
1125 /* save non zero count */
1126 CP32( &nnz[0*4], &h->mb.cache.non_zero_count[x264_scan8[0]+0*8] );
1127 CP32( &nnz[1*4], &h->mb.cache.non_zero_count[x264_scan8[0]+1*8] );
1128 CP32( &nnz[2*4], &h->mb.cache.non_zero_count[x264_scan8[0]+2*8] );
1129 CP32( &nnz[3*4], &h->mb.cache.non_zero_count[x264_scan8[0]+3*8] );
1130 M16( &nnz[16+0*2] ) = M32( &h->mb.cache.non_zero_count[x264_scan8[16+0*2]-1] ) >> 8;
1131 M16( &nnz[16+1*2] ) = M32( &h->mb.cache.non_zero_count[x264_scan8[16+1*2]-1] ) >> 8;
1132 M16( &nnz[16+2*2] ) = M32( &h->mb.cache.non_zero_count[x264_scan8[16+2*2]-1] ) >> 8;
1133 M16( &nnz[16+3*2] ) = M32( &h->mb.cache.non_zero_count[x264_scan8[16+3*2]-1] ) >> 8;
1135 if( h->mb.i_type != I_16x16 && h->mb.i_cbp_luma == 0 && h->mb.i_cbp_chroma == 0 )
1136 h->mb.i_qp = h->mb.i_last_qp;
1137 h->mb.qp[i_mb_xy] = h->mb.i_qp;
1138 h->mb.i_last_dqp = h->mb.i_qp - h->mb.i_last_qp;
1139 h->mb.i_last_qp = h->mb.i_qp;
1142 if( h->mb.i_cbp_luma == 0 && h->mb.i_type != I_8x8 )
1143 h->mb.b_transform_8x8 = 0;
1144 h->mb.mb_transform_size[i_mb_xy] = h->mb.b_transform_8x8;
1146 if( h->sh.i_type != SLICE_TYPE_I )
1148 int16_t (*mv0)[2] = &h->mb.mv[0][i_mb_4x4];
1149 int16_t (*mv1)[2] = &h->mb.mv[1][i_mb_4x4];
1150 int8_t *ref0 = &h->mb.ref[0][i_mb_8x8];
1151 int8_t *ref1 = &h->mb.ref[1][i_mb_8x8];
1152 if( !IS_INTRA( i_mb_type ) )
1154 ref0[0+0*s8x8] = h->mb.cache.ref[0][x264_scan8[0]];
1155 ref0[1+0*s8x8] = h->mb.cache.ref[0][x264_scan8[4]];
1156 ref0[0+1*s8x8] = h->mb.cache.ref[0][x264_scan8[8]];
1157 ref0[1+1*s8x8] = h->mb.cache.ref[0][x264_scan8[12]];
1158 CP128( &mv0[0*s4x4], h->mb.cache.mv[0][x264_scan8[0]+8*0] );
1159 CP128( &mv0[1*s4x4], h->mb.cache.mv[0][x264_scan8[0]+8*1] );
1160 CP128( &mv0[2*s4x4], h->mb.cache.mv[0][x264_scan8[0]+8*2] );
1161 CP128( &mv0[3*s4x4], h->mb.cache.mv[0][x264_scan8[0]+8*3] );
1162 if( h->sh.i_type == SLICE_TYPE_B )
1164 ref1[0+0*s8x8] = h->mb.cache.ref[1][x264_scan8[0]];
1165 ref1[1+0*s8x8] = h->mb.cache.ref[1][x264_scan8[4]];
1166 ref1[0+1*s8x8] = h->mb.cache.ref[1][x264_scan8[8]];
1167 ref1[1+1*s8x8] = h->mb.cache.ref[1][x264_scan8[12]];
1168 CP128( &mv1[0*s4x4], h->mb.cache.mv[1][x264_scan8[0]+8*0] );
1169 CP128( &mv1[1*s4x4], h->mb.cache.mv[1][x264_scan8[0]+8*1] );
1170 CP128( &mv1[2*s4x4], h->mb.cache.mv[1][x264_scan8[0]+8*2] );
1171 CP128( &mv1[3*s4x4], h->mb.cache.mv[1][x264_scan8[0]+8*3] );
1176 M16( &ref0[0*s8x8] ) = (uint8_t)(-1) * 0x0101;
1177 M16( &ref0[1*s8x8] ) = (uint8_t)(-1) * 0x0101;
1178 M128( &mv0[0*s4x4] ) = M128_ZERO;
1179 M128( &mv0[1*s4x4] ) = M128_ZERO;
1180 M128( &mv0[2*s4x4] ) = M128_ZERO;
1181 M128( &mv0[3*s4x4] ) = M128_ZERO;
1182 if( h->sh.i_type == SLICE_TYPE_B )
1184 M16( &ref1[0*s8x8] ) = (uint8_t)(-1) * 0x0101;
1185 M16( &ref1[1*s8x8] ) = (uint8_t)(-1) * 0x0101;
1186 M128( &mv1[0*s4x4] ) = M128_ZERO;
1187 M128( &mv1[1*s4x4] ) = M128_ZERO;
1188 M128( &mv1[2*s4x4] ) = M128_ZERO;
1189 M128( &mv1[3*s4x4] ) = M128_ZERO;
1194 if( h->param.b_cabac )
1196 uint8_t (*mvd0)[2] = h->mb.mvd[0][i_mb_xy];
1197 uint8_t (*mvd1)[2] = h->mb.mvd[1][i_mb_xy];
1198 if( IS_INTRA(i_mb_type) && i_mb_type != I_PCM )
1199 h->mb.chroma_pred_mode[i_mb_xy] = x264_mb_pred_mode8x8c_fix[ h->mb.i_chroma_pred_mode ];
1201 h->mb.chroma_pred_mode[i_mb_xy] = I_PRED_CHROMA_DC;
1203 if( !IS_INTRA( i_mb_type ) && !IS_SKIP( i_mb_type ) && !IS_DIRECT( i_mb_type ) )
1205 CP64( mvd0[0], h->mb.cache.mvd[0][x264_scan8[10]] );
1206 CP16( mvd0[4], h->mb.cache.mvd[0][x264_scan8[5 ]] );
1207 CP16( mvd0[5], h->mb.cache.mvd[0][x264_scan8[7 ]] );
1208 CP16( mvd0[6], h->mb.cache.mvd[0][x264_scan8[13]] );
1209 if( h->sh.i_type == SLICE_TYPE_B )
1211 CP64( mvd1[0], h->mb.cache.mvd[1][x264_scan8[10]] );
1212 CP16( mvd1[4], h->mb.cache.mvd[1][x264_scan8[5 ]] );
1213 CP16( mvd1[5], h->mb.cache.mvd[1][x264_scan8[7 ]] );
1214 CP16( mvd1[6], h->mb.cache.mvd[1][x264_scan8[13]] );
1219 M128( mvd0[0] ) = M128_ZERO;
1220 if( h->sh.i_type == SLICE_TYPE_B )
1221 M128( mvd1[0] ) = M128_ZERO;
1224 if( h->sh.i_type == SLICE_TYPE_B )
1226 if( i_mb_type == B_SKIP || i_mb_type == B_DIRECT )
1227 h->mb.skipbp[i_mb_xy] = 0xf;
1228 else if( i_mb_type == B_8x8 )
1230 int skipbp = ( h->mb.i_sub_partition[0] == D_DIRECT_8x8 ) << 0;
1231 skipbp |= ( h->mb.i_sub_partition[1] == D_DIRECT_8x8 ) << 1;
1232 skipbp |= ( h->mb.i_sub_partition[2] == D_DIRECT_8x8 ) << 2;
1233 skipbp |= ( h->mb.i_sub_partition[3] == D_DIRECT_8x8 ) << 3;
1234 h->mb.skipbp[i_mb_xy] = skipbp;
1237 h->mb.skipbp[i_mb_xy] = 0;
1243 void x264_macroblock_bipred_init( x264_t *h )
1245 for( int field = 0; field <= h->sh.b_mbaff; field++ )
1246 for( int i_ref0 = 0; i_ref0 < (h->i_ref0<<h->sh.b_mbaff); i_ref0++ )
1248 int poc0 = h->fref0[i_ref0>>h->sh.b_mbaff]->i_poc;
1249 if( h->sh.b_mbaff && field^(i_ref0&1) )
1250 poc0 += h->sh.i_delta_poc_bottom;
1251 for( int i_ref1 = 0; i_ref1 < (h->i_ref1<<h->sh.b_mbaff); i_ref1++ )
1253 int dist_scale_factor;
1254 int poc1 = h->fref1[i_ref1>>h->sh.b_mbaff]->i_poc;
1255 if( h->sh.b_mbaff && field^(i_ref1&1) )
1256 poc1 += h->sh.i_delta_poc_bottom;
1257 int cur_poc = h->fdec->i_poc + field*h->sh.i_delta_poc_bottom;
1258 int td = x264_clip3( poc1 - poc0, -128, 127 );
1259 if( td == 0 /* || pic0 is a long-term ref */ )
1260 dist_scale_factor = 256;
1263 int tb = x264_clip3( cur_poc - poc0, -128, 127 );
1264 int tx = (16384 + (abs(td) >> 1)) / td;
1265 dist_scale_factor = x264_clip3( (tb * tx + 32) >> 6, -1024, 1023 );
1268 h->mb.dist_scale_factor_buf[field][i_ref0][i_ref1] = dist_scale_factor;
1270 dist_scale_factor >>= 2;
1271 if( h->param.analyse.b_weighted_bipred
1272 && dist_scale_factor >= -64
1273 && dist_scale_factor <= 128 )
1275 h->mb.bipred_weight_buf[field][i_ref0][i_ref1] = 64 - dist_scale_factor;
1276 // ssse3 implementation of biweight doesn't support the extrema.
1277 // if we ever generate them, we'll have to drop that optimization.
1278 assert( dist_scale_factor >= -63 && dist_scale_factor <= 127 );
1281 h->mb.bipred_weight_buf[field][i_ref0][i_ref1] = 32;