Replace High 4:4:4 profile lossless with High 4:4:4 Predictive.
[x262.git] / encoder / macroblock.c
1 /*****************************************************************************
2  * macroblock.c: h264 encoder library
3  *****************************************************************************
4  * Copyright (C) 2003-2008 x264 project
5  *
6  * Authors: Laurent Aimar <fenrir@via.ecp.fr>
7  *          Loren Merritt <lorenm@u.washington.edu>
8  *          Jason Garrett-Glaser <darkshikari@gmail.com>
9  *
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.
14  *
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.
19  *
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  *****************************************************************************/
24
25 #include "common/common.h"
26 #include "macroblock.h"
27
28 #define ZIG(i,y,x) level[i] = dct[x][y];
29 static inline void zigzag_scan_2x2_dc( int16_t level[4], int16_t dct[2][2] )
30 {
31     ZIG(0,0,0)
32     ZIG(1,0,1)
33     ZIG(2,1,0)
34     ZIG(3,1,1)
35 }
36 #undef ZIG
37
38 /* (ref: JVT-B118)
39  * x264_mb_decimate_score: given dct coeffs it returns a score to see if we could empty this dct coeffs
40  * to 0 (low score means set it to null)
41  * Used in inter macroblock (luma and chroma)
42  *  luma: for a 8x8 block: if score < 4 -> null
43  *        for the complete mb: if score < 6 -> null
44  *  chroma: for the complete mb: if score < 7 -> null
45  */
46 static int x264_mb_decimate_score( int16_t *dct, int i_max )
47 {
48     static const int i_ds_table4[16] = {
49         3,2,2,1,1,1,0,0,0,0,0,0,0,0,0,0 };
50     static const int i_ds_table8[64] = {
51         3,3,3,3,2,2,2,2,2,2,2,2,1,1,1,1,
52         1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,0,
53         0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
54         0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0 };
55
56     const int *ds_table = (i_max == 64) ? i_ds_table8 : i_ds_table4;
57     int i_score = 0;
58     int idx = i_max - 1;
59
60     while( idx >= 0 && dct[idx] == 0 )
61         idx--;
62
63     while( idx >= 0 )
64     {
65         int i_run;
66
67         if( (unsigned)(dct[idx--] + 1) > 2 )
68             return 9;
69
70         i_run = 0;
71         while( idx >= 0 && dct[idx] == 0 )
72         {
73             idx--;
74             i_run++;
75         }
76         i_score += ds_table[i_run];
77     }
78
79     return i_score;
80 }
81
82 static ALWAYS_INLINE void x264_quant_4x4( x264_t *h, int16_t dct[4][4], int i_qp, int i_ctxBlockCat, int b_intra, int idx )
83 {
84     int i_quant_cat = b_intra ? CQM_4IY : CQM_4PY;
85     if( h->mb.b_trellis )
86         x264_quant_4x4_trellis( h, dct, i_quant_cat, i_qp, i_ctxBlockCat, b_intra, idx );
87     else
88         h->quantf.quant_4x4( dct, h->quant4_mf[i_quant_cat][i_qp], h->quant4_bias[i_quant_cat][i_qp] );
89 }
90
91 static ALWAYS_INLINE void x264_quant_8x8( x264_t *h, int16_t dct[8][8], int i_qp, int b_intra, int idx )
92 {
93     int i_quant_cat = b_intra ? CQM_8IY : CQM_8PY;
94     if( h->mb.b_trellis )
95         x264_quant_8x8_trellis( h, dct, i_quant_cat, i_qp, b_intra, idx );
96     else
97         h->quantf.quant_8x8( dct, h->quant8_mf[i_quant_cat][i_qp], h->quant8_bias[i_quant_cat][i_qp] );
98 }
99
100 void x264_mb_encode_i4x4( x264_t *h, int idx, int i_qp )
101 {
102     uint8_t *p_src = &h->mb.pic.p_fenc[0][block_idx_xy_fenc[idx]];
103     uint8_t *p_dst = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[idx]];
104     DECLARE_ALIGNED_16( int16_t dct4x4[4][4] );
105
106     if( h->mb.b_lossless )
107     {
108         h->zigzagf.sub_4x4( h->dct.luma4x4[idx], p_src, p_dst );
109         return;
110     }
111
112     h->dctf.sub4x4_dct( dct4x4, p_src, p_dst );
113
114     x264_quant_4x4( h, dct4x4, i_qp, DCT_LUMA_4x4, 1, idx );
115
116     if( array_non_zero( dct4x4 ) )
117     {
118         h->zigzagf.scan_4x4( h->dct.luma4x4[idx], dct4x4 );
119         h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4IY], i_qp );
120
121         /* output samples to fdec */
122         h->dctf.add4x4_idct( p_dst, dct4x4 );
123     }
124     else
125         memset( h->dct.luma4x4[idx], 0, sizeof(h->dct.luma4x4[idx]));
126 }
127
128 void x264_mb_encode_i8x8( x264_t *h, int idx, int i_qp )
129 {
130     int x = 8 * (idx&1);
131     int y = 8 * (idx>>1);
132     uint8_t *p_src = &h->mb.pic.p_fenc[0][x+y*FENC_STRIDE];
133     uint8_t *p_dst = &h->mb.pic.p_fdec[0][x+y*FDEC_STRIDE];
134     DECLARE_ALIGNED_16( int16_t dct8x8[8][8] );
135
136     if( h->mb.b_lossless )
137     {
138         h->zigzagf.sub_8x8( h->dct.luma8x8[idx], p_src, p_dst );
139         return;
140     }
141
142     h->dctf.sub8x8_dct8( dct8x8, p_src, p_dst );
143
144     x264_quant_8x8( h, dct8x8, i_qp, 1, idx );
145
146     h->zigzagf.scan_8x8( h->dct.luma8x8[idx], dct8x8 );
147     h->quantf.dequant_8x8( dct8x8, h->dequant8_mf[CQM_8IY], i_qp );
148     h->dctf.add8x8_idct8( p_dst, dct8x8 );
149 }
150
151 static void x264_mb_encode_i16x16( x264_t *h, int i_qp )
152 {
153     uint8_t  *p_src = h->mb.pic.p_fenc[0];
154     uint8_t  *p_dst = h->mb.pic.p_fdec[0];
155
156     DECLARE_ALIGNED_16( int16_t dct4x4[16][4][4] );
157     DECLARE_ALIGNED_16( int16_t dct_dc4x4[4][4] );
158
159     int i;
160
161     if( h->mb.b_lossless )
162     {
163         for( i = 0; i < 16; i++ )
164         {
165             int oe = block_idx_xy_fenc[i];
166             int od = block_idx_xy_fdec[i];
167             h->zigzagf.sub_4x4( h->dct.luma4x4[i], p_src+oe, p_dst+od );
168             dct_dc4x4[0][block_idx_yx_1d[i]] = h->dct.luma4x4[i][0];
169             h->dct.luma4x4[i][0] = 0;
170         }
171         h->zigzagf.scan_4x4( h->dct.luma16x16_dc, dct_dc4x4 );
172         return;
173     }
174
175     h->dctf.sub16x16_dct( dct4x4, p_src, p_dst );
176     for( i = 0; i < 16; i++ )
177     {
178         /* copy dc coeff */
179         dct_dc4x4[0][block_idx_xy_1d[i]] = dct4x4[i][0][0];
180         dct4x4[i][0][0] = 0;
181
182         /* quant/scan/dequant */
183         x264_quant_4x4( h, dct4x4[i], i_qp, DCT_LUMA_AC, 1, i );
184
185         h->zigzagf.scan_4x4( h->dct.luma4x4[i], dct4x4[i] );
186         h->quantf.dequant_4x4( dct4x4[i], h->dequant4_mf[CQM_4IY], i_qp );
187     }
188
189     h->dctf.dct4x4dc( dct_dc4x4 );
190     h->quantf.quant_4x4_dc( dct_dc4x4, h->quant4_mf[CQM_4IY][i_qp][0]>>1, h->quant4_bias[CQM_4IY][i_qp][0]<<1 );
191     h->zigzagf.scan_4x4( h->dct.luma16x16_dc, dct_dc4x4 );
192
193     /* output samples to fdec */
194     h->dctf.idct4x4dc( dct_dc4x4 );
195     x264_mb_dequant_4x4_dc( dct_dc4x4, h->dequant4_mf[CQM_4IY], i_qp );  /* XXX not inversed */
196
197     /* calculate dct coeffs */
198     for( i = 0; i < 16; i++ )
199     {
200         /* copy dc coeff */
201         dct4x4[i][0][0] = dct_dc4x4[0][block_idx_xy_1d[i]];
202     }
203     /* put pixels to fdec */
204     h->dctf.add16x16_idct( p_dst, dct4x4 );
205 }
206
207 void x264_mb_encode_8x8_chroma( x264_t *h, int b_inter, int i_qp )
208 {
209     int i, ch;
210     int b_decimate = b_inter && (h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate);
211
212     for( ch = 0; ch < 2; ch++ )
213     {
214         uint8_t  *p_src = h->mb.pic.p_fenc[1+ch];
215         uint8_t  *p_dst = h->mb.pic.p_fdec[1+ch];
216         int i_decimate_score = 0;
217
218         DECLARE_ALIGNED_16( int16_t dct2x2[2][2]  );
219         DECLARE_ALIGNED_16( int16_t dct4x4[4][4][4] );
220
221         if( h->mb.b_lossless )
222         {
223             for( i = 0; i < 4; i++ )
224             {
225                 int oe = block_idx_x[i]*4 + block_idx_y[i]*4*FENC_STRIDE;
226                 int od = block_idx_x[i]*4 + block_idx_y[i]*4*FDEC_STRIDE;
227                 h->zigzagf.sub_4x4( h->dct.luma4x4[16+i+ch*4], p_src+oe, p_dst+od );
228                 h->dct.chroma_dc[ch][i] = h->dct.luma4x4[16+i+ch*4][0];
229                 h->dct.luma4x4[16+i+ch*4][0] = 0;
230             }
231             continue;
232         }
233
234         h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
235         /* calculate dct coeffs */
236         for( i = 0; i < 4; i++ )
237         {
238             /* copy dc coeff */
239             dct2x2[i>>1][i&1] = dct4x4[i][0][0];
240             dct4x4[i][0][0] = 0;
241
242             /* no trellis; it doesn't seem to help chroma noticeably */
243             h->quantf.quant_4x4( dct4x4[i], h->quant4_mf[CQM_4IC+b_inter][i_qp], h->quant4_bias[CQM_4IC+b_inter][i_qp] );
244             h->zigzagf.scan_4x4( h->dct.luma4x4[16+i+ch*4], dct4x4[i] );
245
246             if( b_decimate )
247                 i_decimate_score += x264_mb_decimate_score( h->dct.luma4x4[16+i+ch*4]+1, 15 );
248         }
249
250         h->dctf.dct2x2dc( dct2x2 );
251         h->quantf.quant_2x2_dc( dct2x2, h->quant4_mf[CQM_4IC+b_inter][i_qp][0]>>1, h->quant4_bias[CQM_4IC+b_inter][i_qp][0]<<1 );
252         zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct2x2 );
253
254         /* output samples to fdec */
255         h->dctf.idct2x2dc( dct2x2 );
256         x264_mb_dequant_2x2_dc( dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qp );  /* XXX not inversed */
257
258         if( b_decimate && i_decimate_score < 7 )
259         {
260             /* Near null chroma 8x8 block so make it null (bits saving) */
261             memset( &h->dct.luma4x4[16+ch*4], 0, 4 * sizeof( *h->dct.luma4x4 ) );
262             if( !array_non_zero( dct2x2 ) )
263                 continue;
264             memset( dct4x4, 0, sizeof( dct4x4 ) );
265         }
266         else
267         {
268             for( i = 0; i < 4; i++ )
269                 h->quantf.dequant_4x4( dct4x4[i], h->dequant4_mf[CQM_4IC + b_inter], i_qp );
270         }
271         dct4x4[0][0][0] = dct2x2[0][0];
272         dct4x4[1][0][0] = dct2x2[0][1];
273         dct4x4[2][0][0] = dct2x2[1][0];
274         dct4x4[3][0][0] = dct2x2[1][1];
275         h->dctf.add8x8_idct( p_dst, dct4x4 );
276     }
277
278     /* coded block pattern */
279     h->mb.i_cbp_chroma = 0;
280     for( i = 0; i < 8; i++ )
281     {
282         int nz = array_non_zero( h->dct.luma4x4[16+i] );
283         h->mb.cache.non_zero_count[x264_scan8[16+i]] = nz;
284         h->mb.i_cbp_chroma |= nz;
285     }
286     if( h->mb.i_cbp_chroma )
287         h->mb.i_cbp_chroma = 2;    /* dc+ac (we can't do only ac) */
288     else if( array_non_zero( h->dct.chroma_dc ) )
289         h->mb.i_cbp_chroma = 1;    /* dc only */
290 }
291
292 static void x264_macroblock_encode_skip( x264_t *h )
293 {
294     h->mb.i_cbp_luma = 0x00;
295     h->mb.i_cbp_chroma = 0x00;
296     memset( h->mb.cache.non_zero_count, 0, X264_SCAN8_SIZE );
297     /* store cbp */
298     h->mb.cbp[h->mb.i_mb_xy] = 0;
299 }
300
301 /*****************************************************************************
302  * x264_macroblock_encode_pskip:
303  *  Encode an already marked skip block
304  *****************************************************************************/
305 static void x264_macroblock_encode_pskip( x264_t *h )
306 {
307     const int mvx = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][0],
308                                 h->mb.mv_min[0], h->mb.mv_max[0] );
309     const int mvy = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][1],
310                                 h->mb.mv_min[1], h->mb.mv_max[1] );
311
312     /* don't do pskip motion compensation if it was already done in macroblock_analyse */
313     if( !h->mb.b_skip_mc )
314     {
315         h->mc.mc_luma( h->mb.pic.p_fdec[0],    FDEC_STRIDE,
316                        h->mb.pic.p_fref[0][0], h->mb.pic.i_stride[0],
317                        mvx, mvy, 16, 16 );
318
319         h->mc.mc_chroma( h->mb.pic.p_fdec[1],       FDEC_STRIDE,
320                          h->mb.pic.p_fref[0][0][4], h->mb.pic.i_stride[1],
321                          mvx, mvy, 8, 8 );
322
323         h->mc.mc_chroma( h->mb.pic.p_fdec[2],       FDEC_STRIDE,
324                          h->mb.pic.p_fref[0][0][5], h->mb.pic.i_stride[2],
325                          mvx, mvy, 8, 8 );
326     }
327
328     x264_macroblock_encode_skip( h );
329 }
330
331 /*****************************************************************************
332  * Intra prediction for predictive lossless mode.
333  *****************************************************************************/
334
335 /* Note that these functions take a shortcut (mc.copy instead of actual pixel prediction) which assumes
336  * that the edge pixels of the reconstructed frame are the same as that of the source frame.  This means
337  * they will only work correctly if the neighboring blocks are losslessly coded.  In practice, this means
338  * lossless mode cannot be mixed with lossy mode within a frame. */
339 /* This can be resolved by explicitly copying the edge pixels after doing the mc.copy, but this doesn't
340  * need to be done unless we decide to allow mixing lossless and lossy compression. */
341
342 void x264_predict_lossless_8x8_chroma( x264_t *h, int i_mode )
343 {
344     int stride = h->fenc->i_stride[1] << h->mb.b_interlaced;
345     if( i_mode == I_PRED_CHROMA_V )
346     {
347         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fenc_plane[1]-stride, stride, 8 );
348         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fenc_plane[2]-stride, stride, 8 );
349     }
350     else if( i_mode == I_PRED_CHROMA_H )
351     {
352         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fenc_plane[1]-1, stride, 8 );
353         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fenc_plane[2]-1, stride, 8 );
354     }
355     else
356     {
357         h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[1] );
358         h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[2] );
359     }
360 }
361
362 void x264_predict_lossless_4x4( x264_t *h, uint8_t *p_dst, int idx, int i_mode )
363 {
364     int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
365     uint8_t *p_src = h->mb.pic.p_fenc_plane[0] + block_idx_x[idx]*4 + block_idx_y[idx]*4 * stride;
366
367     if( i_mode == I_PRED_4x4_V )
368         h->mc.copy[PIXEL_4x4]( p_dst, FDEC_STRIDE, p_src-stride, stride, 4 );
369     else if( i_mode == I_PRED_4x4_H )
370         h->mc.copy[PIXEL_4x4]( p_dst, FDEC_STRIDE, p_src-1, stride, 4 );
371     else
372         h->predict_4x4[i_mode]( p_dst );
373 }
374
375 void x264_predict_lossless_8x8( x264_t *h, uint8_t *p_dst, int idx, int i_mode, uint8_t edge[33] )
376 {
377     int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
378     uint8_t *p_src = h->mb.pic.p_fenc_plane[0] + (idx&1)*8 + (idx>>1)*8*stride;
379
380     if( i_mode == I_PRED_8x8_V )
381         h->mc.copy[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src-stride, stride, 8 );
382     else if( i_mode == I_PRED_8x8_H )
383         h->mc.copy[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src-1, stride, 8 );
384     else
385         h->predict_8x8[i_mode]( p_dst, edge );
386 }
387
388 void x264_predict_lossless_16x16( x264_t *h, int i_mode )
389 {
390     int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
391     if( i_mode == I_PRED_16x16_V )
392         h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.p_fenc_plane[0]-stride, stride, 16 );
393     else if( i_mode == I_PRED_16x16_H )
394         h->mc.copy_16x16_unaligned( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.p_fenc_plane[0]-1, stride, 16 );
395     else
396         h->predict_16x16[i_mode]( h->mb.pic.p_fdec[0] );
397 }
398
399 /*****************************************************************************
400  * x264_macroblock_encode:
401  *****************************************************************************/
402 void x264_macroblock_encode( x264_t *h )
403 {
404     int i_cbp_dc = 0;
405     int i_qp = h->mb.i_qp;
406     int b_decimate = h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate;
407     int b_force_no_skip = 0;
408     int i,j,idx;
409     uint8_t nnz8x8[4] = {1,1,1,1};
410
411     if( h->sh.b_mbaff
412         && h->mb.i_mb_xy == h->sh.i_first_mb + h->mb.i_mb_stride
413         && IS_SKIP(h->mb.type[h->sh.i_first_mb]) )
414     {
415         /* The first skip is predicted to be a frame mb pair.
416          * We don't yet support the aff part of mbaff, so force it to non-skip
417          * so that we can pick the aff flag. */
418         b_force_no_skip = 1;
419         if( IS_SKIP(h->mb.i_type) )
420         {
421             if( h->mb.i_type == P_SKIP )
422                 h->mb.i_type = P_L0;
423             else if( h->mb.i_type == B_SKIP )
424                 h->mb.i_type = B_DIRECT;
425         }
426     }
427
428     if( h->mb.i_type == P_SKIP )
429     {
430         /* A bit special */
431         x264_macroblock_encode_pskip( h );
432         return;
433     }
434     if( h->mb.i_type == B_SKIP )
435     {
436         /* don't do bskip motion compensation if it was already done in macroblock_analyse */
437         if( !h->mb.b_skip_mc )
438             x264_mb_mc( h );
439         x264_macroblock_encode_skip( h );
440         return;
441     }
442
443     if( h->mb.i_type == I_16x16 )
444     {
445         const int i_mode = h->mb.i_intra16x16_pred_mode;
446         h->mb.b_transform_8x8 = 0;
447
448         if( h->mb.b_lossless )
449             x264_predict_lossless_16x16( h, i_mode );
450         else
451             h->predict_16x16[i_mode]( h->mb.pic.p_fdec[0] );
452
453         /* encode the 16x16 macroblock */
454         x264_mb_encode_i16x16( h, i_qp );
455     }
456     else if( h->mb.i_type == I_8x8 )
457     {
458         DECLARE_ALIGNED_16( uint8_t edge[33] );
459         h->mb.b_transform_8x8 = 1;
460         /* If we already encoded 3 of the 4 i8x8 blocks, we don't have to do them again. */
461         if( h->mb.i_skip_intra )
462         {
463             h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i8x8_fdec_buf, 16, 16 );
464             /* In RD mode, restore the now-overwritten DCT data. */
465             if( h->mb.i_skip_intra == 2 )
466                 h->mc.memcpy_aligned( h->dct.luma8x8, h->mb.pic.i8x8_dct_buf, sizeof(h->mb.pic.i8x8_dct_buf) );
467         }
468         for( i = h->mb.i_skip_intra ? 3 : 0 ; i < 4; i++ )
469         {
470             uint8_t  *p_dst = &h->mb.pic.p_fdec[0][8 * (i&1) + 8 * (i>>1) * FDEC_STRIDE];
471             int      i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[4*i]];
472             x264_predict_8x8_filter( p_dst, edge, h->mb.i_neighbour8[i], x264_pred_i4x4_neighbors[i_mode] );
473
474             if( h->mb.b_lossless )
475                 x264_predict_lossless_8x8( h, p_dst, i, i_mode, edge );
476             else
477                 h->predict_8x8[i_mode]( p_dst, edge );
478
479             x264_mb_encode_i8x8( h, i, i_qp );
480         }
481         for( i = 0; i < 4; i++ )
482             nnz8x8[i] = array_non_zero( h->dct.luma8x8[i] );
483     }
484     else if( h->mb.i_type == I_4x4 )
485     {
486         h->mb.b_transform_8x8 = 0;
487         /* If we already encoded 15 of the 16 i4x4 blocks, we don't have to do them again. */
488         if( h->mb.i_skip_intra )
489         {
490             h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i4x4_fdec_buf, 16, 16 );
491             /* In RD mode, restore the now-overwritten DCT data. */
492             if( h->mb.i_skip_intra == 2 )
493                 h->mc.memcpy_aligned( h->dct.luma4x4, h->mb.pic.i4x4_dct_buf, sizeof(h->mb.pic.i4x4_dct_buf) );
494         }
495         for( i = h->mb.i_skip_intra ? 15 : 0 ; i < 16; i++ )
496         {
497             uint8_t  *p_dst = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[i]];
498             int      i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[i]];
499
500             if( (h->mb.i_neighbour4[i] & (MB_TOPRIGHT|MB_TOP)) == MB_TOP )
501                 /* emulate missing topright samples */
502                 *(uint32_t*) &p_dst[4-FDEC_STRIDE] = p_dst[3-FDEC_STRIDE] * 0x01010101U;
503
504             if( h->mb.b_lossless )
505                 x264_predict_lossless_4x4( h, p_dst, i, i_mode );
506             else
507                 h->predict_4x4[i_mode]( p_dst );
508             x264_mb_encode_i4x4( h, i, i_qp );
509         }
510     }
511     else    /* Inter MB */
512     {
513         int i8x8, i4x4;
514         int i_decimate_mb = 0;
515
516         /* Don't repeat motion compensation if it was already done in non-RD transform analysis */
517         if( !h->mb.b_skip_mc )
518             x264_mb_mc( h );
519
520         if( h->mb.b_lossless )
521         {
522             if( h->mb.b_transform_8x8 )
523                 for( i8x8 = 0; i8x8 < 4; i8x8++ )
524                 {
525                     int x = 8*(i8x8&1);
526                     int y = 8*(i8x8>>1);
527                     h->zigzagf.sub_8x8( h->dct.luma8x8[i8x8],
528                                         h->mb.pic.p_fenc[0]+x+y*FENC_STRIDE,
529                                         h->mb.pic.p_fdec[0]+x+y*FDEC_STRIDE );
530                     nnz8x8[i8x8] = array_non_zero( h->dct.luma8x8[i8x8] );
531                 }
532             else
533                 for( i4x4 = 0; i4x4 < 16; i4x4++ )
534                 {
535                     h->zigzagf.sub_4x4( h->dct.luma4x4[i4x4],
536                                         h->mb.pic.p_fenc[0]+block_idx_xy_fenc[i4x4],
537                                         h->mb.pic.p_fdec[0]+block_idx_xy_fdec[i4x4] );
538                 }
539         }
540         else if( h->mb.b_transform_8x8 )
541         {
542             DECLARE_ALIGNED_16( int16_t dct8x8[4][8][8] );
543             b_decimate &= !h->mb.b_trellis; // 8x8 trellis is inherently optimal decimation
544             h->dctf.sub16x16_dct8( dct8x8, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
545             h->nr_count[1] += h->mb.b_noise_reduction * 4;
546
547             for( idx = 0; idx < 4; idx++ )
548             {
549                 if( h->mb.b_noise_reduction )
550                     h->quantf.denoise_dct( *dct8x8[idx], h->nr_residual_sum[1], h->nr_offset[1], 64 );
551                 x264_quant_8x8( h, dct8x8[idx], i_qp, 0, idx );
552
553                 h->zigzagf.scan_8x8( h->dct.luma8x8[idx], dct8x8[idx] );
554
555                 if( b_decimate )
556                 {
557                     int i_decimate_8x8 = x264_mb_decimate_score( h->dct.luma8x8[idx], 64 );
558                     i_decimate_mb += i_decimate_8x8;
559                     if( i_decimate_8x8 < 4 )
560                         nnz8x8[idx] = 0;
561                 }
562                 else
563                     nnz8x8[idx] = array_non_zero( dct8x8[idx] );
564             }
565
566             if( i_decimate_mb < 6 && b_decimate )
567                 *(uint32_t*)nnz8x8 = 0;
568             else
569             {
570                 for( idx = 0; idx < 4; idx++ )
571                     if( nnz8x8[idx] )
572                     {
573                         h->quantf.dequant_8x8( dct8x8[idx], h->dequant8_mf[CQM_8PY], i_qp );
574                         h->dctf.add8x8_idct8( &h->mb.pic.p_fdec[0][(idx&1)*8 + (idx>>1)*8*FDEC_STRIDE], dct8x8[idx] );
575                     }
576             }
577         }
578         else
579         {
580             DECLARE_ALIGNED_16( int16_t dct4x4[16][4][4] );
581             h->dctf.sub16x16_dct( dct4x4, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
582             h->nr_count[0] += h->mb.b_noise_reduction * 16;
583
584             for( i8x8 = 0; i8x8 < 4; i8x8++ )
585             {
586                 int i_decimate_8x8;
587
588                 /* encode one 4x4 block */
589                 i_decimate_8x8 = 0;
590                 for( i4x4 = 0; i4x4 < 4; i4x4++ )
591                 {
592                     idx = i8x8 * 4 + i4x4;
593
594                     if( h->mb.b_noise_reduction )
595                         h->quantf.denoise_dct( *dct4x4[idx], h->nr_residual_sum[0], h->nr_offset[0], 16 );
596                     x264_quant_4x4( h, dct4x4[idx], i_qp, DCT_LUMA_4x4, 0, idx );
597
598                     h->zigzagf.scan_4x4( h->dct.luma4x4[idx], dct4x4[idx] );
599
600                     if( b_decimate && i_decimate_8x8 <= 6 )
601                         i_decimate_8x8 += x264_mb_decimate_score( h->dct.luma4x4[idx], 16 );
602                 }
603
604                 /* decimate this 8x8 block */
605                 i_decimate_mb += i_decimate_8x8;
606                 if( i_decimate_8x8 < 4 && b_decimate )
607                     nnz8x8[i8x8] = 0;
608             }
609
610             if( i_decimate_mb < 6 && b_decimate )
611                 *(uint32_t*)nnz8x8 = 0;
612             else
613             {
614                 for( i8x8 = 0; i8x8 < 4; i8x8++ )
615                     if( nnz8x8[i8x8] )
616                     {
617                         for( i = 0; i < 4; i++ )
618                             h->quantf.dequant_4x4( dct4x4[i8x8*4+i], h->dequant4_mf[CQM_4PY], i_qp );
619                         h->dctf.add8x8_idct( &h->mb.pic.p_fdec[0][(i8x8&1)*8 + (i8x8>>1)*8*FDEC_STRIDE], &dct4x4[i8x8*4] );
620                     }
621             }
622         }
623     }
624
625     /* encode chroma */
626     if( IS_INTRA( h->mb.i_type ) )
627     {
628         const int i_mode = h->mb.i_chroma_pred_mode;
629         if( h->mb.b_lossless )
630             x264_predict_lossless_8x8_chroma( h, i_mode );
631         else
632         {
633             h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[1] );
634             h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[2] );
635         }
636     }
637
638     /* encode the 8x8 blocks */
639     x264_mb_encode_8x8_chroma( h, !IS_INTRA( h->mb.i_type ), h->mb.i_chroma_qp );
640
641     /* coded block pattern and non_zero_count */
642     h->mb.i_cbp_luma = 0x00;
643     if( h->mb.i_type == I_16x16 )
644     {
645         for( i = 0; i < 16; i++ )
646         {
647             int nz = array_non_zero( h->dct.luma4x4[i] );
648             h->mb.cache.non_zero_count[x264_scan8[i]] = nz;
649             h->mb.i_cbp_luma |= nz;
650         }
651         h->mb.i_cbp_luma *= 0xf;
652     }
653     else
654     {
655         for( i = 0; i < 4; i++)
656         {
657             if(!nnz8x8[i])
658             {
659                 *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[0+i*4]] = 0;
660                 *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[2+i*4]] = 0;
661             }
662             else if( h->mb.b_transform_8x8 )
663             {
664                 *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[0+4*i]] = nnz8x8[i] * 0x0101;
665                 *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[2+4*i]] = nnz8x8[i] * 0x0101;
666                 h->mb.i_cbp_luma |= nnz8x8[i] << i;
667             }
668             else
669             {
670                 int nz, cbp = 0;
671                 for( j = 0; j < 4; j++ )
672                 {
673                     nz = array_non_zero( h->dct.luma4x4[j+4*i] );
674                     h->mb.cache.non_zero_count[x264_scan8[j+4*i]] = nz;
675                     cbp |= nz;
676                 }
677                 h->mb.i_cbp_luma |= cbp << i;
678             }
679         }
680     }
681
682     if( h->param.b_cabac )
683     {
684         i_cbp_dc = ( h->mb.i_type == I_16x16 && array_non_zero( h->dct.luma16x16_dc ) )
685                  | array_non_zero( h->dct.chroma_dc[0] ) << 1
686                  | array_non_zero( h->dct.chroma_dc[1] ) << 2;
687     }
688
689     /* store cbp */
690     h->mb.cbp[h->mb.i_mb_xy] = (i_cbp_dc << 8) | (h->mb.i_cbp_chroma << 4) | h->mb.i_cbp_luma;
691
692     /* Check for P_SKIP
693      * XXX: in the me perhaps we should take x264_mb_predict_mv_pskip into account
694      *      (if multiple mv give same result)*/
695     if( !b_force_no_skip )
696     {
697         if( h->mb.i_type == P_L0 && h->mb.i_partition == D_16x16 &&
698             !(h->mb.i_cbp_luma | h->mb.i_cbp_chroma) &&
699             *(uint32_t*)h->mb.cache.mv[0][x264_scan8[0]] == *(uint32_t*)h->mb.cache.pskip_mv
700             && h->mb.cache.ref[0][x264_scan8[0]] == 0 )
701         {
702             h->mb.i_type = P_SKIP;
703         }
704
705         /* Check for B_SKIP */
706         if( h->mb.i_type == B_DIRECT && !(h->mb.i_cbp_luma | h->mb.i_cbp_chroma) )
707         {
708             h->mb.i_type = B_SKIP;
709         }
710     }
711 }
712
713 /*****************************************************************************
714  * x264_macroblock_probe_skip:
715  *  Check if the current MB could be encoded as a [PB]_SKIP (it supposes you use
716  *  the previous QP
717  *****************************************************************************/
718 int x264_macroblock_probe_skip( x264_t *h, const int b_bidir )
719 {
720     DECLARE_ALIGNED_16( int16_t dct4x4[4][4][4] );
721     DECLARE_ALIGNED_16( int16_t dct2x2[2][2] );
722     DECLARE_ALIGNED_16( int16_t dctscan[16] );
723
724     int i_qp = h->mb.i_qp;
725     int mvp[2];
726     int ch, thresh;
727
728     int i8x8, i4x4;
729     int i_decimate_mb;
730
731     if( !b_bidir )
732     {
733         /* Get the MV */
734         mvp[0] = x264_clip3( h->mb.cache.pskip_mv[0], h->mb.mv_min[0], h->mb.mv_max[0] );
735         mvp[1] = x264_clip3( h->mb.cache.pskip_mv[1], h->mb.mv_min[1], h->mb.mv_max[1] );
736
737         /* Motion compensation */
738         h->mc.mc_luma( h->mb.pic.p_fdec[0],    FDEC_STRIDE,
739                        h->mb.pic.p_fref[0][0], h->mb.pic.i_stride[0],
740                        mvp[0], mvp[1], 16, 16 );
741     }
742
743     for( i8x8 = 0, i_decimate_mb = 0; i8x8 < 4; i8x8++ )
744     {
745         int fenc_offset = (i8x8&1) * 8 + (i8x8>>1) * FENC_STRIDE * 8;
746         int fdec_offset = (i8x8&1) * 8 + (i8x8>>1) * FDEC_STRIDE * 8;
747         /* get luma diff */
748         h->dctf.sub8x8_dct( dct4x4, h->mb.pic.p_fenc[0] + fenc_offset,
749                                     h->mb.pic.p_fdec[0] + fdec_offset );
750         /* encode one 4x4 block */
751         for( i4x4 = 0; i4x4 < 4; i4x4++ )
752         {
753             h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] );
754             if( !array_non_zero(dct4x4[i4x4]) )
755                 continue;
756             h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
757             i_decimate_mb += x264_mb_decimate_score( dctscan, 16 );
758             if( i_decimate_mb >= 6 )
759                 return 0;
760         }
761     }
762
763     /* encode chroma */
764     i_qp = h->mb.i_chroma_qp;
765     thresh = (x264_lambda2_tab[i_qp] + 32) >> 6;
766
767     for( ch = 0; ch < 2; ch++ )
768     {
769         uint8_t  *p_src = h->mb.pic.p_fenc[1+ch];
770         uint8_t  *p_dst = h->mb.pic.p_fdec[1+ch];
771
772         if( !b_bidir )
773         {
774             h->mc.mc_chroma( h->mb.pic.p_fdec[1+ch],       FDEC_STRIDE,
775                              h->mb.pic.p_fref[0][0][4+ch], h->mb.pic.i_stride[1+ch],
776                              mvp[0], mvp[1], 8, 8 );
777         }
778
779         /* there is almost never a termination during chroma, but we can't avoid the check entirely */
780         /* so instead we check SSD and skip the actual check if the score is low enough. */
781         if( h->pixf.ssd[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src, FENC_STRIDE ) < thresh )
782             continue;
783
784         h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
785
786         /* calculate dct DC */
787         dct2x2[0][0] = dct4x4[0][0][0];
788         dct2x2[0][1] = dct4x4[1][0][0];
789         dct2x2[1][0] = dct4x4[2][0][0];
790         dct2x2[1][1] = dct4x4[3][0][0];
791         h->dctf.dct2x2dc( dct2x2 );
792         h->quantf.quant_2x2_dc( dct2x2, h->quant4_mf[CQM_4PC][i_qp][0]>>1, h->quant4_bias[CQM_4PC][i_qp][0]<<1 );
793         if( array_non_zero(dct2x2) )
794             return 0;
795
796         /* calculate dct coeffs */
797         for( i4x4 = 0, i_decimate_mb = 0; i4x4 < 4; i4x4++ )
798         {
799             h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] );
800             if( !array_non_zero(dct4x4[i4x4]) )
801                 continue;
802             h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
803             i_decimate_mb += x264_mb_decimate_score( dctscan+1, 15 );
804             if( i_decimate_mb >= 7 )
805                 return 0;
806         }
807     }
808
809     h->mb.b_skip_mc = 1;
810     return 1;
811 }
812
813 /****************************************************************************
814  * DCT-domain noise reduction / adaptive deadzone
815  * from libavcodec
816  ****************************************************************************/
817
818 void x264_noise_reduction_update( x264_t *h )
819 {
820     int cat, i;
821     for( cat = 0; cat < 2; cat++ )
822     {
823         int size = cat ? 64 : 16;
824         const uint16_t *weight = cat ? x264_dct8_weight2_tab : x264_dct4_weight2_tab;
825
826         if( h->nr_count[cat] > (cat ? (1<<16) : (1<<18)) )
827         {
828             for( i = 0; i < size; i++ )
829                 h->nr_residual_sum[cat][i] >>= 1;
830             h->nr_count[cat] >>= 1;
831         }
832
833         for( i = 0; i < size; i++ )
834             h->nr_offset[cat][i] =
835                 ((uint64_t)h->param.analyse.i_noise_reduction * h->nr_count[cat]
836                  + h->nr_residual_sum[cat][i]/2)
837               / ((uint64_t)h->nr_residual_sum[cat][i] * weight[i]/256 + 1);
838     }
839 }
840
841 /*****************************************************************************
842  * RD only; 4 calls to this do not make up for one macroblock_encode.
843  * doesn't transform chroma dc.
844  *****************************************************************************/
845 void x264_macroblock_encode_p8x8( x264_t *h, int i8 )
846 {
847     int i_qp = h->mb.i_qp;
848     uint8_t *p_fenc = h->mb.pic.p_fenc[0] + (i8&1)*8 + (i8>>1)*8*FENC_STRIDE;
849     uint8_t *p_fdec = h->mb.pic.p_fdec[0] + (i8&1)*8 + (i8>>1)*8*FDEC_STRIDE;
850     int b_decimate = h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate;
851     int nnz8x8 = 0;
852     int ch;
853
854     x264_mb_mc_8x8( h, i8 );
855
856     if( h->mb.b_lossless )
857     {
858         int i4;
859         if( h->mb.b_transform_8x8 )
860         {
861             h->zigzagf.sub_4x4( h->dct.luma4x4[i8], p_fenc, p_fdec );
862             nnz8x8 = array_non_zero( h->dct.luma8x8[i8] );
863         }
864         else
865         {
866             for( i4 = i8*4; i4 < i8*4+4; i4++ )
867             {
868                 h->zigzagf.sub_4x4( h->dct.luma4x4[i4],
869                                     h->mb.pic.p_fenc[0]+block_idx_xy_fenc[i4],
870                                     h->mb.pic.p_fdec[0]+block_idx_xy_fdec[i4] );
871                 nnz8x8 |= array_non_zero( h->dct.luma4x4[i4] );
872             }
873         }
874         for( ch = 0; ch < 2; ch++ )
875         {
876             p_fenc = h->mb.pic.p_fenc[1+ch] + (i8&1)*4 + (i8>>1)*4*FENC_STRIDE;
877             p_fdec = h->mb.pic.p_fdec[1+ch] + (i8&1)*4 + (i8>>1)*4*FDEC_STRIDE;
878             h->zigzagf.sub_4x4( h->dct.luma4x4[16+i8+ch*4], p_fenc, p_fdec );
879             h->dct.luma4x4[16+i8+ch*4][0] = 0;
880         }
881     }
882     else
883     {
884         if( h->mb.b_transform_8x8 )
885         {
886             DECLARE_ALIGNED_16( int16_t dct8x8[8][8] );
887             h->dctf.sub8x8_dct8( dct8x8, p_fenc, p_fdec );
888             x264_quant_8x8( h, dct8x8, i_qp, 0, i8 );
889             h->zigzagf.scan_8x8( h->dct.luma8x8[i8], dct8x8 );
890
891             if( b_decimate && !h->mb.b_trellis )
892                 nnz8x8 = 4 <= x264_mb_decimate_score( h->dct.luma8x8[i8], 64 );
893             else
894                 nnz8x8 = array_non_zero( dct8x8 );
895
896             if( nnz8x8 )
897             {
898                 h->quantf.dequant_8x8( dct8x8, h->dequant8_mf[CQM_8PY], i_qp );
899                 h->dctf.add8x8_idct8( p_fdec, dct8x8 );
900             }
901         }
902         else
903         {
904             int i4;
905             DECLARE_ALIGNED_16( int16_t dct4x4[4][4][4] );
906             h->dctf.sub8x8_dct( dct4x4, p_fenc, p_fdec );
907             for( i4 = 0; i4 < 4; i4++ )
908                 x264_quant_4x4( h, dct4x4[i4], i_qp, DCT_LUMA_4x4, 0, i8*4+i4 );
909
910             for( i4 = 0; i4 < 4; i4++ )
911                 h->zigzagf.scan_4x4( h->dct.luma4x4[i8*4+i4], dct4x4[i4] );
912
913             if( b_decimate )
914             {
915                 int i_decimate_8x8 = 0;
916                 for( i4 = 0; i4 < 4 && i_decimate_8x8 < 4; i4++ )
917                     i_decimate_8x8 += x264_mb_decimate_score( h->dct.luma4x4[i8*4+i4], 16 );
918                 nnz8x8 = 4 <= i_decimate_8x8;
919             }
920             else
921                 nnz8x8 = array_non_zero( dct4x4 );
922
923             if( nnz8x8 )
924             {
925                 for( i4 = 0; i4 < 4; i4++ )
926                     h->quantf.dequant_4x4( dct4x4[i4], h->dequant4_mf[CQM_4PY], i_qp );
927                 h->dctf.add8x8_idct( p_fdec, dct4x4 );
928             }
929         }
930
931         i_qp = h->mb.i_chroma_qp;
932
933         for( ch = 0; ch < 2; ch++ )
934         {
935             DECLARE_ALIGNED_16( int16_t dct4x4[4][4] );
936             p_fenc = h->mb.pic.p_fenc[1+ch] + (i8&1)*4 + (i8>>1)*4*FENC_STRIDE;
937             p_fdec = h->mb.pic.p_fdec[1+ch] + (i8&1)*4 + (i8>>1)*4*FDEC_STRIDE;
938
939             h->dctf.sub4x4_dct( dct4x4, p_fenc, p_fdec );
940             h->quantf.quant_4x4( dct4x4, h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] );
941             h->zigzagf.scan_4x4( h->dct.luma4x4[16+i8+ch*4], dct4x4 );
942             h->dct.luma4x4[16+i8+ch*4][0] = 0;
943             if( array_non_zero( dct4x4 ) )
944             {
945                 h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4PC], i_qp );
946                 h->dctf.add4x4_idct( p_fdec, dct4x4 );
947             }
948         }
949     }
950     h->mb.i_cbp_luma &= ~(1 << i8);
951     h->mb.i_cbp_luma |= nnz8x8 << i8;
952     h->mb.i_cbp_chroma = 0x02;
953 }