1 /*****************************************************************************
2 * macroblock.c: macroblock common functions
3 *****************************************************************************
4 * Copyright (C) 2003-2010 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
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 )
317 for( int i = 0; i <= 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 h->intra_border_backup[1][j] = h->intra_border_backup[i][j];
326 CHECKED_MALLOC( h->deblock_strength[i], sizeof(**h->deblock_strength) * h->mb.i_mb_width );
327 h->deblock_strength[1] = h->deblock_strength[i];
330 /* Allocate scratch buffer */
331 int scratch_size = 0;
334 int buf_hpel = (h->thread[0]->fdec->i_width[0]+48) * sizeof(int16_t);
335 int buf_ssim = h->param.analyse.b_ssim * 8 * (h->param.i_width/4+3) * sizeof(int);
336 int me_range = X264_MIN(h->param.analyse.i_me_range, h->param.analyse.i_mv_range);
337 int buf_tesa = (h->param.analyse.i_me_method >= X264_ME_ESA) *
338 ((me_range*2+24) * sizeof(int16_t) + (me_range+4) * (me_range+1) * 4 * sizeof(mvsad_t));
339 scratch_size = X264_MAX3( buf_hpel, buf_ssim, buf_tesa );
341 int buf_mbtree = h->param.rc.b_mb_tree * ((h->mb.i_mb_width+3)&~3) * sizeof(int);
342 scratch_size = X264_MAX( scratch_size, buf_mbtree );
343 CHECKED_MALLOC( h->scratch_buffer, scratch_size );
350 void x264_macroblock_thread_free( x264_t *h, int b_lookahead )
353 for( int i = 0; i <= h->param.b_interlaced; i++ )
355 x264_free( h->deblock_strength[i] );
356 for( int j = 0; j < 2; j++ )
357 x264_free( h->intra_border_backup[i][j] - 16 );
359 x264_free( h->scratch_buffer );
362 void x264_macroblock_slice_init( x264_t *h )
364 h->mb.mv[0] = h->fdec->mv[0];
365 h->mb.mv[1] = h->fdec->mv[1];
366 h->mb.mvr[0][0] = h->fdec->mv16x16;
367 h->mb.ref[0] = h->fdec->ref[0];
368 h->mb.ref[1] = h->fdec->ref[1];
369 h->mb.type = h->fdec->mb_type;
370 h->mb.partition = h->fdec->mb_partition;
372 h->fdec->i_ref[0] = h->i_ref0;
373 h->fdec->i_ref[1] = h->i_ref1;
374 for( int i = 0; i < h->i_ref0; i++ )
375 h->fdec->ref_poc[0][i] = h->fref0[i]->i_poc;
376 if( h->sh.i_type == SLICE_TYPE_B )
378 for( int i = 0; i < h->i_ref1; i++ )
379 h->fdec->ref_poc[1][i] = h->fref1[i]->i_poc;
381 map_col_to_list0(-1) = -1;
382 map_col_to_list0(-2) = -2;
383 for( int i = 0; i < h->fref1[0]->i_ref[0]; i++ )
385 int poc = h->fref1[0]->ref_poc[0][i];
386 map_col_to_list0(i) = -2;
387 for( int j = 0; j < h->i_ref0; j++ )
388 if( h->fref0[j]->i_poc == poc )
390 map_col_to_list0(i) = j;
395 else if( h->sh.i_type == SLICE_TYPE_P )
397 memset( h->mb.cache.skip, 0, sizeof( h->mb.cache.skip ) );
399 if( h->sh.i_disable_deblocking_filter_idc != 1 && h->param.analyse.i_weighted_pred == X264_WEIGHTP_SMART )
401 deblock_ref_table(-2) = -2;
402 deblock_ref_table(-1) = -1;
403 for( int i = 0; i < h->i_ref0 << h->sh.b_mbaff; i++ )
405 /* Mask off high bits to avoid frame num collisions with -1/-2.
406 * In current x264 frame num values don't cover a range of more
407 * than 32, so 6 bits is enough for uniqueness. */
408 if( !h->mb.b_interlaced )
409 deblock_ref_table(i) = h->fref0[i]->i_frame_num&63;
411 deblock_ref_table(i) = ((h->fref0[i>>1]->i_frame_num&63)<<1) + (i&1);
416 /* init with not available (for top right idx=7,15) */
417 memset( h->mb.cache.ref, -2, sizeof( h->mb.cache.ref ) );
420 for( int field = 0; field <= h->sh.b_mbaff; field++ )
422 int curpoc = h->fdec->i_poc + field*h->sh.i_delta_poc_bottom;
423 int refpoc = h->fref0[0]->i_poc;
424 if( h->sh.b_mbaff && field )
425 refpoc += h->sh.i_delta_poc_bottom;
426 int delta = curpoc - refpoc;
428 h->fdec->inv_ref_poc[field] = (256 + delta/2) / delta;
431 h->mb.i_neighbour4[6] =
432 h->mb.i_neighbour4[9] =
433 h->mb.i_neighbour4[12] =
434 h->mb.i_neighbour4[14] = MB_LEFT|MB_TOP|MB_TOPLEFT|MB_TOPRIGHT;
435 h->mb.i_neighbour4[3] =
436 h->mb.i_neighbour4[7] =
437 h->mb.i_neighbour4[11] =
438 h->mb.i_neighbour4[13] =
439 h->mb.i_neighbour4[15] =
440 h->mb.i_neighbour8[3] = MB_LEFT|MB_TOP|MB_TOPLEFT;
443 void x264_macroblock_thread_init( x264_t *h )
445 h->mb.i_me_method = h->param.analyse.i_me_method;
446 h->mb.i_subpel_refine = h->param.analyse.i_subpel_refine;
447 if( h->sh.i_type == SLICE_TYPE_B && (h->mb.i_subpel_refine == 6 || h->mb.i_subpel_refine == 8) )
448 h->mb.i_subpel_refine--;
449 h->mb.b_chroma_me = h->param.analyse.b_chroma_me &&
450 ((h->sh.i_type == SLICE_TYPE_P && h->mb.i_subpel_refine >= 5) ||
451 (h->sh.i_type == SLICE_TYPE_B && h->mb.i_subpel_refine >= 9));
452 h->mb.b_dct_decimate = h->sh.i_type == SLICE_TYPE_B ||
453 (h->param.analyse.b_dct_decimate && h->sh.i_type != SLICE_TYPE_I);
466 h->mb.pic.p_fenc[0] = h->mb.pic.fenc_buf;
467 h->mb.pic.p_fenc[1] = h->mb.pic.fenc_buf + 16*FENC_STRIDE;
468 h->mb.pic.p_fenc[2] = h->mb.pic.fenc_buf + 16*FENC_STRIDE + 8;
469 h->mb.pic.p_fdec[0] = h->mb.pic.fdec_buf + 2*FDEC_STRIDE;
470 h->mb.pic.p_fdec[1] = h->mb.pic.fdec_buf + 19*FDEC_STRIDE;
471 h->mb.pic.p_fdec[2] = h->mb.pic.fdec_buf + 19*FDEC_STRIDE + 16;
474 void x264_prefetch_fenc( x264_t *h, x264_frame_t *fenc, int i_mb_x, int i_mb_y )
476 int stride_y = fenc->i_stride[0];
477 int stride_uv = fenc->i_stride[1];
478 int off_y = 16 * i_mb_x + 16 * i_mb_y * stride_y;
479 int off_uv = 16 * i_mb_x + 8 * i_mb_y * stride_uv;
480 h->mc.prefetch_fenc( fenc->plane[0]+off_y, stride_y,
481 fenc->plane[1]+off_uv, stride_uv, i_mb_x );
484 NOINLINE void x264_copy_column8( pixel *dst, pixel *src )
486 // input pointers are offset by 4 rows because that's faster (smaller instruction size on x86)
487 for( int i = -4; i < 4; i++ )
488 dst[i*FDEC_STRIDE] = src[i*FDEC_STRIDE];
491 static void ALWAYS_INLINE x264_macroblock_load_pic_pointers( x264_t *h, int mb_x, int mb_y, int i, int b_interlaced )
493 int w = (i ? 8 : 16);
494 int i_stride = h->fdec->i_stride[i];
495 int i_stride2 = i_stride << b_interlaced;
496 int i_pix_offset = b_interlaced
497 ? 16 * mb_x + w * (mb_y&~1) * i_stride + (mb_y&1) * i_stride
498 : 16 * mb_x + w * mb_y * i_stride;
499 pixel *plane_fdec = &h->fdec->plane[i][i_pix_offset];
500 pixel *intra_fdec = &h->intra_border_backup[mb_y&1][i][mb_x*16];
501 int ref_pix_offset[2] = { i_pix_offset, i_pix_offset };
502 x264_frame_t **fref[2] = { h->fref0, h->fref1 };
504 ref_pix_offset[1] += (1-2*(mb_y&1)) * i_stride;
505 h->mb.pic.i_stride[i] = i_stride2;
506 h->mb.pic.p_fenc_plane[i] = &h->fenc->plane[i][i_pix_offset];
509 h->mc.load_deinterleave_8x8x2_fenc( h->mb.pic.p_fenc[1], h->mb.pic.p_fenc_plane[1], i_stride2 );
510 memcpy( h->mb.pic.p_fdec[1]-FDEC_STRIDE, intra_fdec, 8*sizeof(pixel) );
511 memcpy( h->mb.pic.p_fdec[2]-FDEC_STRIDE, intra_fdec+8, 8*sizeof(pixel) );
515 h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fenc[0], FENC_STRIDE, h->mb.pic.p_fenc_plane[0], i_stride2, 16 );
516 memcpy( h->mb.pic.p_fdec[0]-FDEC_STRIDE, intra_fdec, 24*sizeof(pixel) );
520 for( int j = 0; j < w; j++ )
523 h->mb.pic.p_fdec[1][-1+j*FDEC_STRIDE] = plane_fdec[-2+j*i_stride2];
524 h->mb.pic.p_fdec[2][-1+j*FDEC_STRIDE] = plane_fdec[-1+j*i_stride2];
527 h->mb.pic.p_fdec[0][-1+j*FDEC_STRIDE] = plane_fdec[-1+j*i_stride2];
529 for( int j = 0; j < h->mb.pic.i_fref[0]; j++ )
531 h->mb.pic.p_fref[0][j][i?4:0] = &fref[0][j >> b_interlaced]->plane[i][ref_pix_offset[j&1]];
534 for( int k = 1; k < 4; k++ )
535 h->mb.pic.p_fref[0][j][k] = &fref[0][j >> b_interlaced]->filtered[k][ref_pix_offset[j&1]];
536 if( h->sh.weight[j][0].weightfn )
537 h->mb.pic.p_fref_w[j] = &h->fenc->weighted[j >> b_interlaced][ref_pix_offset[j&1]];
539 h->mb.pic.p_fref_w[j] = h->mb.pic.p_fref[0][j][0];
542 if( h->sh.i_type == SLICE_TYPE_B )
543 for( int j = 0; j < h->mb.pic.i_fref[1]; j++ )
545 h->mb.pic.p_fref[1][j][i?4:0] = &fref[1][j >> b_interlaced]->plane[i][ref_pix_offset[j&1]];
547 for( int k = 1; k < 4; k++ )
548 h->mb.pic.p_fref[1][j][k] = &fref[1][j >> b_interlaced]->filtered[k][ref_pix_offset[j&1]];
552 static void inline x264_macroblock_cache_load_neighbours( x264_t *h, int mb_x, int mb_y )
554 int top = (mb_y - (1 << h->mb.b_interlaced)) * h->mb.i_mb_stride + mb_x;
558 h->mb.i_mb_xy = mb_y * h->mb.i_mb_stride + mb_x;
559 h->mb.i_b8_xy = 2*(mb_y * h->mb.i_b8_stride + mb_x);
560 h->mb.i_b4_xy = 4*(mb_y * h->mb.i_b4_stride + mb_x);
561 h->mb.i_neighbour = 0;
562 h->mb.i_neighbour_intra = 0;
563 h->mb.i_neighbour_frame = 0;
564 h->mb.i_mb_top_xy = -1;
565 h->mb.i_mb_left_xy = -1;
566 h->mb.i_mb_topleft_xy = -1;
567 h->mb.i_mb_topright_xy = -1;
568 h->mb.i_mb_type_top = -1;
569 h->mb.i_mb_type_left = -1;
570 h->mb.i_mb_type_topleft = -1;
571 h->mb.i_mb_type_topright = -1;
575 h->mb.i_neighbour_frame |= MB_LEFT;
576 h->mb.i_mb_left_xy = h->mb.i_mb_xy - 1;
577 h->mb.i_mb_type_left = h->mb.type[h->mb.i_mb_left_xy];
578 if( h->mb.i_mb_xy > h->sh.i_first_mb )
580 h->mb.i_neighbour |= MB_LEFT;
582 if( !h->param.b_constrained_intra || IS_INTRA( h->mb.i_mb_type_left ) )
583 h->mb.i_neighbour_intra |= MB_LEFT;
587 /* We can't predict from the previous threadslice since it hasn't been encoded yet. */
588 if( (h->i_threadslice_start >> h->mb.b_interlaced) != (mb_y >> h->mb.b_interlaced) )
592 h->mb.i_neighbour_frame |= MB_TOP;
593 h->mb.i_mb_top_xy = top;
594 h->mb.i_mb_type_top = h->mb.type[h->mb.i_mb_top_xy];
595 if( top >= h->sh.i_first_mb )
597 h->mb.i_neighbour |= MB_TOP;
599 if( !h->param.b_constrained_intra || IS_INTRA( h->mb.i_mb_type_top ) )
600 h->mb.i_neighbour_intra |= MB_TOP;
602 /* We only need to prefetch the top blocks because the left was just written
603 * to as part of the previous cache_save. Since most target CPUs use write-allocate
604 * caches, left blocks are near-guaranteed to be in L1 cache. Top--not so much. */
605 x264_prefetch( &h->mb.cbp[top] );
606 x264_prefetch( h->mb.intra4x4_pred_mode[top] );
607 x264_prefetch( &h->mb.non_zero_count[top][12] );
608 /* These aren't always allocated, but prefetching an invalid address can't hurt. */
609 x264_prefetch( &h->mb.mb_transform_size[top] );
610 x264_prefetch( &h->mb.skipbp[top] );
614 if( mb_x > 0 && top - 1 >= 0 )
616 h->mb.i_neighbour_frame |= MB_TOPLEFT;
617 h->mb.i_mb_topleft_xy = top - 1;
618 h->mb.i_mb_type_topleft = h->mb.type[h->mb.i_mb_topleft_xy];
619 if( top - 1 >= h->sh.i_first_mb )
621 h->mb.i_neighbour |= MB_TOPLEFT;
623 if( !h->param.b_constrained_intra || IS_INTRA( h->mb.i_mb_type_topleft ) )
624 h->mb.i_neighbour_intra |= MB_TOPLEFT;
628 if( mb_x < h->mb.i_mb_width - 1 && top + 1 >= 0 )
630 h->mb.i_neighbour_frame |= MB_TOPRIGHT;
631 h->mb.i_mb_topright_xy = top + 1;
632 h->mb.i_mb_type_topright = h->mb.type[h->mb.i_mb_topright_xy];
633 if( top + 1 >= h->sh.i_first_mb )
635 h->mb.i_neighbour |= MB_TOPRIGHT;
637 if( !h->param.b_constrained_intra || IS_INTRA( h->mb.i_mb_type_topright ) )
638 h->mb.i_neighbour_intra |= MB_TOPRIGHT;
644 void x264_macroblock_cache_load( x264_t *h, int mb_x, int mb_y )
646 x264_macroblock_cache_load_neighbours( h, mb_x, mb_y );
648 int left = h->mb.i_mb_left_xy;
649 int top = h->mb.i_mb_top_xy;
650 int top_y = mb_y - (1 << h->mb.b_interlaced);
651 int s8x8 = h->mb.i_b8_stride;
652 int s4x4 = h->mb.i_b4_stride;
653 int top_8x8 = (2*top_y+1) * s8x8 + 2*mb_x;
654 int top_4x4 = (4*top_y+3) * s4x4 + 4*mb_x;
655 int lists = (1 << h->sh.i_type) & 3;
657 /* GCC pessimizes direct loads from heap-allocated arrays due to aliasing. */
658 /* By only dereferencing them once, we avoid this issue. */
659 int8_t (*i4x4)[8] = h->mb.intra4x4_pred_mode;
660 uint8_t (*nnz)[24] = h->mb.non_zero_count;
661 int16_t *cbp = h->mb.cbp;
664 if( h->mb.i_neighbour & MB_TOP )
666 h->mb.cache.i_cbp_top = cbp[top];
668 CP32( &h->mb.cache.intra4x4_pred_mode[x264_scan8[0] - 8], &i4x4[top][0] );
670 /* load non_zero_count */
671 CP32( &h->mb.cache.non_zero_count[x264_scan8[0] - 8], &nnz[top][12] );
672 /* shift because x264_scan8[16] is misaligned */
673 M32( &h->mb.cache.non_zero_count[x264_scan8[16+0] - 9] ) = M16( &nnz[top][18] ) << 8;
674 M32( &h->mb.cache.non_zero_count[x264_scan8[16+4] - 9] ) = M16( &nnz[top][22] ) << 8;
676 /* Finish the prefetching */
677 for( int l = 0; l < lists; l++ )
679 x264_prefetch( &h->mb.mv[l][top_4x4-1] );
680 /* Top right being not in the same cacheline as top left will happen
681 * once every 4 MBs, so one extra prefetch is worthwhile */
682 x264_prefetch( &h->mb.mv[l][top_4x4+4] );
683 x264_prefetch( &h->mb.ref[l][top_8x8-1] );
684 x264_prefetch( &h->mb.mvd[l][top] );
689 h->mb.cache.i_cbp_top = -1;
692 M32( &h->mb.cache.intra4x4_pred_mode[x264_scan8[0] - 8] ) = 0xFFFFFFFFU;
694 /* load non_zero_count */
695 M32( &h->mb.cache.non_zero_count[x264_scan8[ 0] - 8] ) = 0x80808080U;
696 M32( &h->mb.cache.non_zero_count[x264_scan8[16+0] - 9] ) = 0x80808080U;
697 M32( &h->mb.cache.non_zero_count[x264_scan8[16+4] - 9] ) = 0x80808080U;
700 if( h->mb.i_neighbour & MB_LEFT )
702 h->mb.cache.i_cbp_left = cbp[left];
705 h->mb.cache.intra4x4_pred_mode[x264_scan8[0 ] - 1] = i4x4[left][4];
706 h->mb.cache.intra4x4_pred_mode[x264_scan8[2 ] - 1] = i4x4[left][5];
707 h->mb.cache.intra4x4_pred_mode[x264_scan8[8 ] - 1] = i4x4[left][6];
708 h->mb.cache.intra4x4_pred_mode[x264_scan8[10] - 1] = i4x4[left][3];
710 /* load non_zero_count */
711 h->mb.cache.non_zero_count[x264_scan8[0 ] - 1] = nnz[left][3];
712 h->mb.cache.non_zero_count[x264_scan8[2 ] - 1] = nnz[left][7];
713 h->mb.cache.non_zero_count[x264_scan8[8 ] - 1] = nnz[left][11];
714 h->mb.cache.non_zero_count[x264_scan8[10] - 1] = nnz[left][15];
716 h->mb.cache.non_zero_count[x264_scan8[16+0] - 1] = nnz[left][16+1];
717 h->mb.cache.non_zero_count[x264_scan8[16+2] - 1] = nnz[left][16+3];
719 h->mb.cache.non_zero_count[x264_scan8[16+4+0] - 1] = nnz[left][16+4+1];
720 h->mb.cache.non_zero_count[x264_scan8[16+4+2] - 1] = nnz[left][16+4+3];
724 h->mb.cache.i_cbp_left = -1;
726 h->mb.cache.intra4x4_pred_mode[x264_scan8[0 ] - 1] =
727 h->mb.cache.intra4x4_pred_mode[x264_scan8[2 ] - 1] =
728 h->mb.cache.intra4x4_pred_mode[x264_scan8[8 ] - 1] =
729 h->mb.cache.intra4x4_pred_mode[x264_scan8[10] - 1] = -1;
731 /* load non_zero_count */
732 h->mb.cache.non_zero_count[x264_scan8[0 ] - 1] =
733 h->mb.cache.non_zero_count[x264_scan8[2 ] - 1] =
734 h->mb.cache.non_zero_count[x264_scan8[8 ] - 1] =
735 h->mb.cache.non_zero_count[x264_scan8[10] - 1] =
736 h->mb.cache.non_zero_count[x264_scan8[16+0] - 1] =
737 h->mb.cache.non_zero_count[x264_scan8[16+2] - 1] =
738 h->mb.cache.non_zero_count[x264_scan8[16+4+0] - 1] =
739 h->mb.cache.non_zero_count[x264_scan8[16+4+2] - 1] = 0x80;
742 if( h->pps->b_transform_8x8_mode )
744 h->mb.cache.i_neighbour_transform_size =
745 ( (h->mb.i_neighbour & MB_LEFT) && h->mb.mb_transform_size[left] )
746 + ( (h->mb.i_neighbour & MB_TOP) && h->mb.mb_transform_size[top] );
751 h->mb.pic.i_fref[0] = h->i_ref0 << h->mb.b_interlaced;
752 h->mb.pic.i_fref[1] = h->i_ref1 << h->mb.b_interlaced;
753 h->mb.cache.i_neighbour_interlaced =
754 !!(h->mb.i_neighbour & MB_LEFT)
755 + !!(h->mb.i_neighbour & MB_TOP);
758 if( !h->mb.b_interlaced )
760 x264_copy_column8( h->mb.pic.p_fdec[0]-1+ 4*FDEC_STRIDE, h->mb.pic.p_fdec[0]+15+ 4*FDEC_STRIDE );
761 x264_copy_column8( h->mb.pic.p_fdec[0]-1+12*FDEC_STRIDE, h->mb.pic.p_fdec[0]+15+12*FDEC_STRIDE );
762 x264_copy_column8( h->mb.pic.p_fdec[1]-1+ 4*FDEC_STRIDE, h->mb.pic.p_fdec[1]+ 7+ 4*FDEC_STRIDE );
763 x264_copy_column8( h->mb.pic.p_fdec[2]-1+ 4*FDEC_STRIDE, h->mb.pic.p_fdec[2]+ 7+ 4*FDEC_STRIDE );
764 x264_macroblock_load_pic_pointers( h, mb_x, mb_y, 0, 0 );
765 x264_macroblock_load_pic_pointers( h, mb_x, mb_y, 1, 0 );
769 x264_macroblock_load_pic_pointers( h, mb_x, mb_y, 0, 1 );
770 x264_macroblock_load_pic_pointers( h, mb_x, mb_y, 1, 1 );
773 if( h->fdec->integral )
775 int offset = 16 * (mb_x + mb_y * h->fdec->i_stride[0]);
776 for( int i = 0; i < h->mb.pic.i_fref[0]; i++ )
777 h->mb.pic.p_integral[0][i] = &h->fref0[i]->integral[offset];
778 for( int i = 0; i < h->mb.pic.i_fref[1]; i++ )
779 h->mb.pic.p_integral[1][i] = &h->fref1[i]->integral[offset];
782 x264_prefetch_fenc( h, h->fenc, mb_x, mb_y );
784 /* load ref/mv/mvd */
785 for( int l = 0; l < lists; l++ )
787 int16_t (*mv)[2] = h->mb.mv[l];
788 int8_t *ref = h->mb.ref[l];
790 int i8 = x264_scan8[0] - 1 - 1*8;
791 if( h->mb.i_neighbour & MB_TOPLEFT )
793 h->mb.cache.ref[l][i8] = ref[top_8x8 - 1];
794 CP32( h->mb.cache.mv[l][i8], mv[top_4x4 - 1] );
798 h->mb.cache.ref[l][i8] = -2;
799 M32( h->mb.cache.mv[l][i8] ) = 0;
802 i8 = x264_scan8[0] - 8;
803 if( h->mb.i_neighbour & MB_TOP )
805 h->mb.cache.ref[l][i8+0] =
806 h->mb.cache.ref[l][i8+1] = ref[top_8x8 + 0];
807 h->mb.cache.ref[l][i8+2] =
808 h->mb.cache.ref[l][i8+3] = ref[top_8x8 + 1];
809 CP128( h->mb.cache.mv[l][i8], mv[top_4x4] );
813 M128( h->mb.cache.mv[l][i8] ) = M128_ZERO;
814 M32( &h->mb.cache.ref[l][i8] ) = (uint8_t)(-2) * 0x01010101U;
817 i8 = x264_scan8[0] + 4 - 1*8;
818 if( h->mb.i_neighbour & MB_TOPRIGHT )
820 h->mb.cache.ref[l][i8] = ref[top_8x8 + 2];
821 CP32( h->mb.cache.mv[l][i8], mv[top_4x4 + 4] );
824 h->mb.cache.ref[l][i8] = -2;
826 i8 = x264_scan8[0] - 1;
827 if( h->mb.i_neighbour & MB_LEFT )
829 const int ir = h->mb.i_b8_xy - 1;
830 const int iv = h->mb.i_b4_xy - 1;
831 h->mb.cache.ref[l][i8+0*8] =
832 h->mb.cache.ref[l][i8+1*8] = ref[ir + 0*s8x8];
833 h->mb.cache.ref[l][i8+2*8] =
834 h->mb.cache.ref[l][i8+3*8] = ref[ir + 1*s8x8];
836 CP32( h->mb.cache.mv[l][i8+0*8], mv[iv + 0*s4x4] );
837 CP32( h->mb.cache.mv[l][i8+1*8], mv[iv + 1*s4x4] );
838 CP32( h->mb.cache.mv[l][i8+2*8], mv[iv + 2*s4x4] );
839 CP32( h->mb.cache.mv[l][i8+3*8], mv[iv + 3*s4x4] );
843 for( int i = 0; i < 4; i++ )
845 h->mb.cache.ref[l][i8+i*8] = -2;
846 M32( h->mb.cache.mv[l][i8+i*8] ) = 0;
850 if( h->param.b_cabac )
852 uint8_t (*mvd)[8][2] = h->mb.mvd[l];
853 if( h->mb.i_neighbour & MB_TOP )
854 CP64( h->mb.cache.mvd[l][x264_scan8[0] - 8], mvd[top][0] );
856 M64( h->mb.cache.mvd[l][x264_scan8[0] - 8] ) = 0;
858 if( h->mb.i_neighbour & MB_LEFT )
860 CP16( h->mb.cache.mvd[l][x264_scan8[0 ] - 1], mvd[left][4] );
861 CP16( h->mb.cache.mvd[l][x264_scan8[2 ] - 1], mvd[left][5] );
862 CP16( h->mb.cache.mvd[l][x264_scan8[8 ] - 1], mvd[left][6] );
863 CP16( h->mb.cache.mvd[l][x264_scan8[10] - 1], mvd[left][3] );
866 for( int i = 0; i < 4; i++ )
867 M16( h->mb.cache.mvd[l][x264_scan8[0]-1+i*8] ) = 0;
872 if( h->sh.i_type == SLICE_TYPE_B )
874 h->mb.bipred_weight = h->mb.bipred_weight_buf[h->mb.b_interlaced&(mb_y&1)];
875 h->mb.dist_scale_factor = h->mb.dist_scale_factor_buf[h->mb.b_interlaced&(mb_y&1)];
876 if( h->param.b_cabac )
879 x264_macroblock_cache_skip( h, 0, 0, 4, 4, 0 );
880 skipbp = (h->mb.i_neighbour & MB_LEFT) ? h->mb.skipbp[left] : 0;
881 h->mb.cache.skip[x264_scan8[0] - 1] = skipbp & 0x2;
882 h->mb.cache.skip[x264_scan8[8] - 1] = skipbp & 0x8;
883 skipbp = (h->mb.i_neighbour & MB_TOP) ? h->mb.skipbp[top] : 0;
884 h->mb.cache.skip[x264_scan8[0] - 8] = skipbp & 0x4;
885 h->mb.cache.skip[x264_scan8[4] - 8] = skipbp & 0x8;
889 if( h->sh.i_type == SLICE_TYPE_P )
890 x264_mb_predict_mv_pskip( h, h->mb.cache.pskip_mv );
892 h->mb.i_neighbour4[0] =
893 h->mb.i_neighbour8[0] = (h->mb.i_neighbour_intra & (MB_TOP|MB_LEFT|MB_TOPLEFT))
894 | ((h->mb.i_neighbour_intra & MB_TOP) ? MB_TOPRIGHT : 0);
895 h->mb.i_neighbour4[4] =
896 h->mb.i_neighbour4[1] = MB_LEFT | ((h->mb.i_neighbour_intra & MB_TOP) ? (MB_TOP|MB_TOPLEFT|MB_TOPRIGHT) : 0);
897 h->mb.i_neighbour4[2] =
898 h->mb.i_neighbour4[8] =
899 h->mb.i_neighbour4[10] =
900 h->mb.i_neighbour8[2] = MB_TOP|MB_TOPRIGHT | ((h->mb.i_neighbour_intra & MB_LEFT) ? (MB_LEFT|MB_TOPLEFT) : 0);
901 h->mb.i_neighbour4[5] =
902 h->mb.i_neighbour8[1] = MB_LEFT | (h->mb.i_neighbour_intra & MB_TOPRIGHT)
903 | ((h->mb.i_neighbour_intra & MB_TOP) ? MB_TOP|MB_TOPLEFT : 0);
906 void x264_macroblock_cache_load_neighbours_deblock( x264_t *h, int mb_x, int mb_y )
908 int deblock_on_slice_edges = h->sh.i_disable_deblocking_filter_idc != 2;
910 h->mb.i_neighbour = 0;
911 h->mb.i_mb_xy = mb_y * h->mb.i_mb_stride + mb_x;
915 h->mb.i_mb_left_xy = h->mb.i_mb_xy - 1;
916 if( deblock_on_slice_edges || h->mb.slice_table[h->mb.i_mb_left_xy] == h->mb.slice_table[h->mb.i_mb_xy] )
917 h->mb.i_neighbour |= MB_LEFT;
920 if( mb_y > h->mb.b_interlaced )
922 h->mb.i_mb_top_xy = h->mb.i_mb_xy - (h->mb.i_mb_stride << h->mb.b_interlaced);
923 if( deblock_on_slice_edges || h->mb.slice_table[h->mb.i_mb_top_xy] == h->mb.slice_table[h->mb.i_mb_xy] )
924 h->mb.i_neighbour |= MB_TOP;
928 void x264_macroblock_cache_load_deblock( x264_t *h )
930 if( IS_INTRA( h->mb.type[h->mb.i_mb_xy] ) )
933 /* If we have multiple slices and we're deblocking on slice edges, we
934 * have to reload neighbour data. */
935 if( h->sh.i_first_mb && h->sh.i_disable_deblocking_filter_idc != 2 )
937 int old_neighbour = h->mb.i_neighbour;
938 int mb_x = h->mb.i_mb_x;
939 int mb_y = h->mb.i_mb_y;
940 x264_macroblock_cache_load_neighbours_deblock( h, mb_x, mb_y );
941 int new_neighbour = h->mb.i_neighbour;
942 h->mb.i_neighbour &= ~old_neighbour;
943 if( h->mb.i_neighbour )
945 int top_y = mb_y - (1 << h->mb.b_interlaced);
946 int top_8x8 = (2*top_y+1) * h->mb.i_b8_stride + 2*mb_x;
947 int top_4x4 = (4*top_y+3) * h->mb.i_b4_stride + 4*mb_x;
948 int s8x8 = h->mb.i_b8_stride;
949 int s4x4 = h->mb.i_b4_stride;
951 uint8_t (*nnz)[24] = h->mb.non_zero_count;
953 if( h->mb.i_neighbour & MB_TOP )
954 CP32( &h->mb.cache.non_zero_count[x264_scan8[0] - 8], &nnz[h->mb.i_mb_top_xy][12] );
956 if( h->mb.i_neighbour & MB_LEFT )
958 int left = h->mb.i_mb_left_xy;
959 h->mb.cache.non_zero_count[x264_scan8[0 ] - 1] = nnz[left][3];
960 h->mb.cache.non_zero_count[x264_scan8[2 ] - 1] = nnz[left][7];
961 h->mb.cache.non_zero_count[x264_scan8[8 ] - 1] = nnz[left][11];
962 h->mb.cache.non_zero_count[x264_scan8[10] - 1] = nnz[left][15];
965 for( int l = 0; l <= (h->sh.i_type == SLICE_TYPE_B); l++ )
967 int16_t (*mv)[2] = h->mb.mv[l];
968 int8_t *ref = h->mb.ref[l];
970 int i8 = x264_scan8[0] - 8;
971 if( h->mb.i_neighbour & MB_TOP )
973 h->mb.cache.ref[l][i8+0] =
974 h->mb.cache.ref[l][i8+1] = ref[top_8x8 + 0];
975 h->mb.cache.ref[l][i8+2] =
976 h->mb.cache.ref[l][i8+3] = ref[top_8x8 + 1];
977 CP128( h->mb.cache.mv[l][i8], mv[top_4x4] );
980 i8 = x264_scan8[0] - 1;
981 if( h->mb.i_neighbour & MB_LEFT )
983 int ir = h->mb.i_b8_xy - 1;
984 int iv = h->mb.i_b4_xy - 1;
985 h->mb.cache.ref[l][i8+0*8] =
986 h->mb.cache.ref[l][i8+1*8] = ref[ir + 0*s8x8];
987 h->mb.cache.ref[l][i8+2*8] =
988 h->mb.cache.ref[l][i8+3*8] = ref[ir + 1*s8x8];
990 CP32( h->mb.cache.mv[l][i8+0*8], mv[iv + 0*s4x4] );
991 CP32( h->mb.cache.mv[l][i8+1*8], mv[iv + 1*s4x4] );
992 CP32( h->mb.cache.mv[l][i8+2*8], mv[iv + 2*s4x4] );
993 CP32( h->mb.cache.mv[l][i8+3*8], mv[iv + 3*s4x4] );
997 h->mb.i_neighbour = new_neighbour;
1000 if( h->param.analyse.i_weighted_pred == X264_WEIGHTP_SMART && h->sh.i_type == SLICE_TYPE_P )
1002 /* Handle reference frame duplicates */
1003 int i8 = x264_scan8[0] - 8;
1004 h->mb.cache.ref[0][i8+0] =
1005 h->mb.cache.ref[0][i8+1] = deblock_ref_table(h->mb.cache.ref[0][i8+0]);
1006 h->mb.cache.ref[0][i8+2] =
1007 h->mb.cache.ref[0][i8+3] = deblock_ref_table(h->mb.cache.ref[0][i8+2]);
1009 i8 = x264_scan8[0] - 1;
1010 h->mb.cache.ref[0][i8+0*8] =
1011 h->mb.cache.ref[0][i8+1*8] = deblock_ref_table(h->mb.cache.ref[0][i8+0*8]);
1012 h->mb.cache.ref[0][i8+2*8] =
1013 h->mb.cache.ref[0][i8+3*8] = deblock_ref_table(h->mb.cache.ref[0][i8+2*8]);
1015 int ref0 = deblock_ref_table(h->mb.cache.ref[0][x264_scan8[ 0]]);
1016 int ref1 = deblock_ref_table(h->mb.cache.ref[0][x264_scan8[ 4]]);
1017 int ref2 = deblock_ref_table(h->mb.cache.ref[0][x264_scan8[ 8]]);
1018 int ref3 = deblock_ref_table(h->mb.cache.ref[0][x264_scan8[12]]);
1019 uint32_t reftop = pack16to32( (uint8_t)ref0, (uint8_t)ref1 ) * 0x0101;
1020 uint32_t refbot = pack16to32( (uint8_t)ref2, (uint8_t)ref3 ) * 0x0101;
1022 M32( &h->mb.cache.ref[0][x264_scan8[0]+8*0] ) = reftop;
1023 M32( &h->mb.cache.ref[0][x264_scan8[0]+8*1] ) = reftop;
1024 M32( &h->mb.cache.ref[0][x264_scan8[0]+8*2] ) = refbot;
1025 M32( &h->mb.cache.ref[0][x264_scan8[0]+8*3] ) = refbot;
1028 /* Munge NNZ for cavlc + 8x8dct */
1029 if( !h->param.b_cabac && h->pps->b_transform_8x8_mode )
1031 uint8_t (*nnz)[24] = h->mb.non_zero_count;
1032 int top = h->mb.i_mb_top_xy;
1033 int left = h->mb.i_mb_left_xy;
1035 if( (h->mb.i_neighbour & MB_TOP) && h->mb.mb_transform_size[top] )
1037 int i8 = x264_scan8[0] - 8;
1038 int nnz_top0 = M16( &nnz[top][8] ) | M16( &nnz[top][12] );
1039 int nnz_top1 = M16( &nnz[top][10] ) | M16( &nnz[top][14] );
1040 M16( &h->mb.cache.non_zero_count[i8+0] ) = nnz_top0 ? 0x0101 : 0;
1041 M16( &h->mb.cache.non_zero_count[i8+2] ) = nnz_top1 ? 0x0101 : 0;
1044 if( (h->mb.i_neighbour & MB_LEFT) && h->mb.mb_transform_size[left] )
1046 int i8 = x264_scan8[0] - 1;
1047 int nnz_left0 = M16( &nnz[left][2] ) | M16( &nnz[left][6] );
1048 int nnz_left1 = M16( &nnz[left][10] ) | M16( &nnz[left][14] );
1049 h->mb.cache.non_zero_count[i8+8*0] = !!nnz_left0;
1050 h->mb.cache.non_zero_count[i8+8*1] = !!nnz_left0;
1051 h->mb.cache.non_zero_count[i8+8*2] = !!nnz_left1;
1052 h->mb.cache.non_zero_count[i8+8*3] = !!nnz_left1;
1055 if( h->mb.mb_transform_size[h->mb.i_mb_xy] )
1057 int nnz0 = M16( &h->mb.cache.non_zero_count[x264_scan8[ 0]] ) | M16( &h->mb.cache.non_zero_count[x264_scan8[ 2]] );
1058 int nnz1 = M16( &h->mb.cache.non_zero_count[x264_scan8[ 4]] ) | M16( &h->mb.cache.non_zero_count[x264_scan8[ 6]] );
1059 int nnz2 = M16( &h->mb.cache.non_zero_count[x264_scan8[ 8]] ) | M16( &h->mb.cache.non_zero_count[x264_scan8[10]] );
1060 int nnz3 = M16( &h->mb.cache.non_zero_count[x264_scan8[12]] ) | M16( &h->mb.cache.non_zero_count[x264_scan8[14]] );
1061 uint32_t nnztop = pack16to32( !!nnz0, !!nnz1 ) * 0x0101;
1062 uint32_t nnzbot = pack16to32( !!nnz2, !!nnz3 ) * 0x0101;
1064 M32( &h->mb.cache.non_zero_count[x264_scan8[0]+8*0] ) = nnztop;
1065 M32( &h->mb.cache.non_zero_count[x264_scan8[0]+8*1] ) = nnztop;
1066 M32( &h->mb.cache.non_zero_count[x264_scan8[0]+8*2] ) = nnzbot;
1067 M32( &h->mb.cache.non_zero_count[x264_scan8[0]+8*3] ) = nnzbot;
1072 static void ALWAYS_INLINE twiddle_topleft_pixel( pixel *dst, pixel *src, int b_interlaced )
1074 // We update intra_border_backup in-place, so the topleft neighbor will no longer
1075 // exist there when load_pic_pointers wants it. Move it within p_fdec instead.
1085 static void ALWAYS_INLINE x264_macroblock_store_pic( x264_t *h, int mb_x, int mb_y, int i, int b_interlaced )
1088 int i_stride = h->fdec->i_stride[i];
1089 int i_stride2 = i_stride << b_interlaced;
1090 int i_pix_offset = b_interlaced
1091 ? 16 * mb_x + w * (mb_y&~1) * i_stride + (mb_y&1) * i_stride
1092 : 16 * mb_x + w * mb_y * i_stride;
1093 pixel *intra_fdec = &h->intra_border_backup[mb_y&1][i][mb_x*16];
1096 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] );
1097 memcpy( intra_fdec, h->mb.pic.p_fdec[1]+FDEC_STRIDE*7, 8*sizeof(pixel) );
1098 memcpy( intra_fdec+8, h->mb.pic.p_fdec[2]+FDEC_STRIDE*7, 8*sizeof(pixel) );
1099 twiddle_topleft_pixel( h->mb.pic.p_fdec[1]-FDEC_STRIDE-1, h->mb.pic.p_fdec[1]-FDEC_STRIDE+7, b_interlaced );
1100 twiddle_topleft_pixel( h->mb.pic.p_fdec[2]-FDEC_STRIDE-1, h->mb.pic.p_fdec[2]-FDEC_STRIDE+7, b_interlaced );
1104 h->mc.copy[PIXEL_16x16]( &h->fdec->plane[0][i_pix_offset], i_stride2, h->mb.pic.p_fdec[0], FDEC_STRIDE, 16 );
1105 memcpy( intra_fdec, h->mb.pic.p_fdec[0]+FDEC_STRIDE*15, 16*sizeof(pixel) );
1106 twiddle_topleft_pixel( h->mb.pic.p_fdec[0]-FDEC_STRIDE-1, h->mb.pic.p_fdec[0]-FDEC_STRIDE+15, b_interlaced );
1110 void x264_macroblock_cache_save( x264_t *h )
1112 const int i_mb_xy = h->mb.i_mb_xy;
1113 const int i_mb_type = x264_mb_type_fix[h->mb.i_type];
1114 const int s8x8 = h->mb.i_b8_stride;
1115 const int s4x4 = h->mb.i_b4_stride;
1116 const int i_mb_4x4 = h->mb.i_b4_xy;
1117 const int i_mb_8x8 = h->mb.i_b8_xy;
1119 /* GCC pessimizes direct stores to heap-allocated arrays due to aliasing. */
1120 /* By only dereferencing them once, we avoid this issue. */
1121 int8_t *i4x4 = h->mb.intra4x4_pred_mode[i_mb_xy];
1122 uint8_t *nnz = h->mb.non_zero_count[i_mb_xy];
1124 if( h->mb.b_interlaced )
1126 x264_macroblock_store_pic( h, h->mb.i_mb_x, h->mb.i_mb_y, 0, 1 );
1127 x264_macroblock_store_pic( h, h->mb.i_mb_x, h->mb.i_mb_y, 1, 1 );
1131 x264_macroblock_store_pic( h, h->mb.i_mb_x, h->mb.i_mb_y, 0, 0 );
1132 x264_macroblock_store_pic( h, h->mb.i_mb_x, h->mb.i_mb_y, 1, 0 );
1135 x264_prefetch_fenc( h, h->fdec, h->mb.i_mb_x, h->mb.i_mb_y );
1137 h->mb.type[i_mb_xy] = i_mb_type;
1138 h->mb.slice_table[i_mb_xy] = h->sh.i_first_mb;
1139 h->mb.partition[i_mb_xy] = IS_INTRA( i_mb_type ) ? D_16x16 : h->mb.i_partition;
1140 h->mb.i_mb_prev_xy = i_mb_xy;
1143 if( i_mb_type == I_4x4 )
1145 CP32( &i4x4[0], &h->mb.cache.intra4x4_pred_mode[x264_scan8[10]] );
1146 M32( &i4x4[4] ) = pack8to32( h->mb.cache.intra4x4_pred_mode[x264_scan8[5] ],
1147 h->mb.cache.intra4x4_pred_mode[x264_scan8[7] ],
1148 h->mb.cache.intra4x4_pred_mode[x264_scan8[13] ], 0);
1150 else if( !h->param.b_constrained_intra || IS_INTRA(i_mb_type) )
1151 M64( i4x4 ) = I_PRED_4x4_DC * 0x0101010101010101ULL;
1153 M64( i4x4 ) = (uint8_t)(-1) * 0x0101010101010101ULL;
1156 if( i_mb_type == I_PCM )
1158 h->mb.qp[i_mb_xy] = 0;
1159 h->mb.i_last_dqp = 0;
1160 h->mb.i_cbp_chroma = 2;
1161 h->mb.i_cbp_luma = 0xf;
1162 h->mb.cbp[i_mb_xy] = 0x72f; /* all set */
1163 h->mb.b_transform_8x8 = 0;
1164 for( int i = 0; i < 24; i++ )
1165 h->mb.cache.non_zero_count[x264_scan8[i]] = h->param.b_cabac ? 1 : 16;
1169 if( h->mb.i_type != I_16x16 && h->mb.i_cbp_luma == 0 && h->mb.i_cbp_chroma == 0 )
1170 h->mb.i_qp = h->mb.i_last_qp;
1171 h->mb.qp[i_mb_xy] = h->mb.i_qp;
1172 h->mb.i_last_dqp = h->mb.i_qp - h->mb.i_last_qp;
1173 h->mb.i_last_qp = h->mb.i_qp;
1176 /* save non zero count */
1177 CP32( &nnz[0*4], &h->mb.cache.non_zero_count[x264_scan8[0]+0*8] );
1178 CP32( &nnz[1*4], &h->mb.cache.non_zero_count[x264_scan8[0]+1*8] );
1179 CP32( &nnz[2*4], &h->mb.cache.non_zero_count[x264_scan8[0]+2*8] );
1180 CP32( &nnz[3*4], &h->mb.cache.non_zero_count[x264_scan8[0]+3*8] );
1181 M16( &nnz[16+0*2] ) = M32( &h->mb.cache.non_zero_count[x264_scan8[16+0*2]-1] ) >> 8;
1182 M16( &nnz[16+1*2] ) = M32( &h->mb.cache.non_zero_count[x264_scan8[16+1*2]-1] ) >> 8;
1183 M16( &nnz[16+2*2] ) = M32( &h->mb.cache.non_zero_count[x264_scan8[16+2*2]-1] ) >> 8;
1184 M16( &nnz[16+3*2] ) = M32( &h->mb.cache.non_zero_count[x264_scan8[16+3*2]-1] ) >> 8;
1186 if( h->mb.i_cbp_luma == 0 && h->mb.i_type != I_8x8 )
1187 h->mb.b_transform_8x8 = 0;
1188 h->mb.mb_transform_size[i_mb_xy] = h->mb.b_transform_8x8;
1190 if( h->sh.i_type != SLICE_TYPE_I )
1192 int16_t (*mv0)[2] = &h->mb.mv[0][i_mb_4x4];
1193 int16_t (*mv1)[2] = &h->mb.mv[1][i_mb_4x4];
1194 int8_t *ref0 = &h->mb.ref[0][i_mb_8x8];
1195 int8_t *ref1 = &h->mb.ref[1][i_mb_8x8];
1196 if( !IS_INTRA( i_mb_type ) )
1198 ref0[0+0*s8x8] = h->mb.cache.ref[0][x264_scan8[0]];
1199 ref0[1+0*s8x8] = h->mb.cache.ref[0][x264_scan8[4]];
1200 ref0[0+1*s8x8] = h->mb.cache.ref[0][x264_scan8[8]];
1201 ref0[1+1*s8x8] = h->mb.cache.ref[0][x264_scan8[12]];
1202 CP128( &mv0[0*s4x4], h->mb.cache.mv[0][x264_scan8[0]+8*0] );
1203 CP128( &mv0[1*s4x4], h->mb.cache.mv[0][x264_scan8[0]+8*1] );
1204 CP128( &mv0[2*s4x4], h->mb.cache.mv[0][x264_scan8[0]+8*2] );
1205 CP128( &mv0[3*s4x4], h->mb.cache.mv[0][x264_scan8[0]+8*3] );
1206 if( h->sh.i_type == SLICE_TYPE_B )
1208 ref1[0+0*s8x8] = h->mb.cache.ref[1][x264_scan8[0]];
1209 ref1[1+0*s8x8] = h->mb.cache.ref[1][x264_scan8[4]];
1210 ref1[0+1*s8x8] = h->mb.cache.ref[1][x264_scan8[8]];
1211 ref1[1+1*s8x8] = h->mb.cache.ref[1][x264_scan8[12]];
1212 CP128( &mv1[0*s4x4], h->mb.cache.mv[1][x264_scan8[0]+8*0] );
1213 CP128( &mv1[1*s4x4], h->mb.cache.mv[1][x264_scan8[0]+8*1] );
1214 CP128( &mv1[2*s4x4], h->mb.cache.mv[1][x264_scan8[0]+8*2] );
1215 CP128( &mv1[3*s4x4], h->mb.cache.mv[1][x264_scan8[0]+8*3] );
1220 M16( &ref0[0*s8x8] ) = (uint8_t)(-1) * 0x0101;
1221 M16( &ref0[1*s8x8] ) = (uint8_t)(-1) * 0x0101;
1222 M128( &mv0[0*s4x4] ) = M128_ZERO;
1223 M128( &mv0[1*s4x4] ) = M128_ZERO;
1224 M128( &mv0[2*s4x4] ) = M128_ZERO;
1225 M128( &mv0[3*s4x4] ) = M128_ZERO;
1226 if( h->sh.i_type == SLICE_TYPE_B )
1228 M16( &ref1[0*s8x8] ) = (uint8_t)(-1) * 0x0101;
1229 M16( &ref1[1*s8x8] ) = (uint8_t)(-1) * 0x0101;
1230 M128( &mv1[0*s4x4] ) = M128_ZERO;
1231 M128( &mv1[1*s4x4] ) = M128_ZERO;
1232 M128( &mv1[2*s4x4] ) = M128_ZERO;
1233 M128( &mv1[3*s4x4] ) = M128_ZERO;
1238 if( h->param.b_cabac )
1240 uint8_t (*mvd0)[2] = h->mb.mvd[0][i_mb_xy];
1241 uint8_t (*mvd1)[2] = h->mb.mvd[1][i_mb_xy];
1242 if( IS_INTRA(i_mb_type) && i_mb_type != I_PCM )
1243 h->mb.chroma_pred_mode[i_mb_xy] = x264_mb_pred_mode8x8c_fix[ h->mb.i_chroma_pred_mode ];
1245 h->mb.chroma_pred_mode[i_mb_xy] = I_PRED_CHROMA_DC;
1247 if( (0x3FF30 >> i_mb_type) & 1 ) /* !INTRA && !SKIP && !DIRECT */
1249 CP64( mvd0[0], h->mb.cache.mvd[0][x264_scan8[10]] );
1250 CP16( mvd0[4], h->mb.cache.mvd[0][x264_scan8[5 ]] );
1251 CP16( mvd0[5], h->mb.cache.mvd[0][x264_scan8[7 ]] );
1252 CP16( mvd0[6], h->mb.cache.mvd[0][x264_scan8[13]] );
1253 if( h->sh.i_type == SLICE_TYPE_B )
1255 CP64( mvd1[0], h->mb.cache.mvd[1][x264_scan8[10]] );
1256 CP16( mvd1[4], h->mb.cache.mvd[1][x264_scan8[5 ]] );
1257 CP16( mvd1[5], h->mb.cache.mvd[1][x264_scan8[7 ]] );
1258 CP16( mvd1[6], h->mb.cache.mvd[1][x264_scan8[13]] );
1263 M128( mvd0[0] ) = M128_ZERO;
1264 if( h->sh.i_type == SLICE_TYPE_B )
1265 M128( mvd1[0] ) = M128_ZERO;
1268 if( h->sh.i_type == SLICE_TYPE_B )
1270 if( i_mb_type == B_SKIP || i_mb_type == B_DIRECT )
1271 h->mb.skipbp[i_mb_xy] = 0xf;
1272 else if( i_mb_type == B_8x8 )
1274 int skipbp = ( h->mb.i_sub_partition[0] == D_DIRECT_8x8 ) << 0;
1275 skipbp |= ( h->mb.i_sub_partition[1] == D_DIRECT_8x8 ) << 1;
1276 skipbp |= ( h->mb.i_sub_partition[2] == D_DIRECT_8x8 ) << 2;
1277 skipbp |= ( h->mb.i_sub_partition[3] == D_DIRECT_8x8 ) << 3;
1278 h->mb.skipbp[i_mb_xy] = skipbp;
1281 h->mb.skipbp[i_mb_xy] = 0;
1287 void x264_macroblock_bipred_init( x264_t *h )
1289 for( int field = 0; field <= h->sh.b_mbaff; field++ )
1290 for( int i_ref0 = 0; i_ref0 < (h->i_ref0<<h->sh.b_mbaff); i_ref0++ )
1292 int poc0 = h->fref0[i_ref0>>h->sh.b_mbaff]->i_poc;
1293 if( h->sh.b_mbaff && field^(i_ref0&1) )
1294 poc0 += h->sh.i_delta_poc_bottom;
1295 for( int i_ref1 = 0; i_ref1 < (h->i_ref1<<h->sh.b_mbaff); i_ref1++ )
1297 int dist_scale_factor;
1298 int poc1 = h->fref1[i_ref1>>h->sh.b_mbaff]->i_poc;
1299 if( h->sh.b_mbaff && field^(i_ref1&1) )
1300 poc1 += h->sh.i_delta_poc_bottom;
1301 int cur_poc = h->fdec->i_poc + field*h->sh.i_delta_poc_bottom;
1302 int td = x264_clip3( poc1 - poc0, -128, 127 );
1303 if( td == 0 /* || pic0 is a long-term ref */ )
1304 dist_scale_factor = 256;
1307 int tb = x264_clip3( cur_poc - poc0, -128, 127 );
1308 int tx = (16384 + (abs(td) >> 1)) / td;
1309 dist_scale_factor = x264_clip3( (tb * tx + 32) >> 6, -1024, 1023 );
1312 h->mb.dist_scale_factor_buf[field][i_ref0][i_ref1] = dist_scale_factor;
1314 dist_scale_factor >>= 2;
1315 if( h->param.analyse.b_weighted_bipred
1316 && dist_scale_factor >= -64
1317 && dist_scale_factor <= 128 )
1319 h->mb.bipred_weight_buf[field][i_ref0][i_ref1] = 64 - dist_scale_factor;
1320 // ssse3 implementation of biweight doesn't support the extrema.
1321 // if we ever generate them, we'll have to drop that optimization.
1322 assert( dist_scale_factor >= -63 && dist_scale_factor <= 127 );
1325 h->mb.bipred_weight_buf[field][i_ref0][i_ref1] = 32;