1 /*****************************************************************************
2 * macroblock.c: h264 encoder library
3 *****************************************************************************
4 * Copyright (C) 2003-2008 x264 project
6 * Authors: Laurent Aimar <fenrir@via.ecp.fr>
7 * Loren Merritt <lorenm@u.washington.edu>
8 * Fiona Glaser <fiona@x264.com>
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 *****************************************************************************/
25 #include "common/common.h"
26 #include "macroblock.h"
28 /* These chroma DC functions don't have assembly versions and are only used here. */
30 #define ZIG(i,y,x) level[i] = dct[x*2+y];
31 static inline void zigzag_scan_2x2_dc( int16_t level[4], int16_t dct[4] )
40 #define IDCT_DEQUANT_START \
41 int d0 = dct[0] + dct[1]; \
42 int d1 = dct[2] + dct[3]; \
43 int d2 = dct[0] - dct[1]; \
44 int d3 = dct[2] - dct[3]; \
45 int dmf = dequant_mf[i_qp%6][0] << i_qp/6;
47 static inline void idct_dequant_2x2_dc( int16_t dct[4], int16_t dct4x4[4][16], int dequant_mf[6][16], int i_qp )
50 dct4x4[0][0] = (d0 + d1) * dmf >> 5;
51 dct4x4[1][0] = (d0 - d1) * dmf >> 5;
52 dct4x4[2][0] = (d2 + d3) * dmf >> 5;
53 dct4x4[3][0] = (d2 - d3) * dmf >> 5;
56 static inline void idct_dequant_2x2_dconly( int16_t out[4], int16_t dct[4], int dequant_mf[6][16], int i_qp )
59 out[0] = (d0 + d1) * dmf >> 5;
60 out[1] = (d0 - d1) * dmf >> 5;
61 out[2] = (d2 + d3) * dmf >> 5;
62 out[3] = (d2 - d3) * dmf >> 5;
65 static inline void dct2x2dc( int16_t d[4], int16_t dct4x4[4][16] )
67 int d0 = dct4x4[0][0] + dct4x4[1][0];
68 int d1 = dct4x4[2][0] + dct4x4[3][0];
69 int d2 = dct4x4[0][0] - dct4x4[1][0];
70 int d3 = dct4x4[2][0] - dct4x4[3][0];
81 static ALWAYS_INLINE int x264_quant_4x4( x264_t *h, int16_t dct[16], int i_qp, int i_ctxBlockCat, int b_intra, int idx )
83 int i_quant_cat = b_intra ? CQM_4IY : CQM_4PY;
85 return x264_quant_4x4_trellis( h, dct, i_quant_cat, i_qp, i_ctxBlockCat, b_intra, 0, idx );
87 return h->quantf.quant_4x4( dct, h->quant4_mf[i_quant_cat][i_qp], h->quant4_bias[i_quant_cat][i_qp] );
90 static ALWAYS_INLINE int x264_quant_8x8( x264_t *h, int16_t dct[64], int i_qp, int b_intra, int idx )
92 int i_quant_cat = b_intra ? CQM_8IY : CQM_8PY;
94 return x264_quant_8x8_trellis( h, dct, i_quant_cat, i_qp, b_intra, idx );
96 return h->quantf.quant_8x8( dct, h->quant8_mf[i_quant_cat][i_qp], h->quant8_bias[i_quant_cat][i_qp] );
99 /* All encoding functions must output the correct CBP and NNZ values.
100 * The entropy coding functions will check CBP first, then NNZ, before
101 * actually reading the DCT coefficients. NNZ still must be correct even
102 * if CBP is zero because of the use of NNZ values for context selection.
103 * "NNZ" need only be 0 or 1 rather than the exact coefficient count because
104 * that is only needed in CAVLC, and will be calculated by CAVLC's residual
105 * coding and stored as necessary. */
107 /* This means that decimation can be done merely by adjusting the CBP and NNZ
108 * rather than memsetting the coefficients. */
110 void x264_mb_encode_i4x4( x264_t *h, int idx, int i_qp )
113 uint8_t *p_src = &h->mb.pic.p_fenc[0][block_idx_xy_fenc[idx]];
114 uint8_t *p_dst = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[idx]];
115 ALIGNED_ARRAY_16( int16_t, dct4x4,[16] );
117 if( h->mb.b_lossless )
119 nz = h->zigzagf.sub_4x4( h->dct.luma4x4[idx], p_src, p_dst );
120 h->mb.cache.non_zero_count[x264_scan8[idx]] = nz;
121 h->mb.i_cbp_luma |= nz<<(idx>>2);
125 h->dctf.sub4x4_dct( dct4x4, p_src, p_dst );
127 nz = x264_quant_4x4( h, dct4x4, i_qp, DCT_LUMA_4x4, 1, idx );
128 h->mb.cache.non_zero_count[x264_scan8[idx]] = nz;
131 h->mb.i_cbp_luma |= 1<<(idx>>2);
132 h->zigzagf.scan_4x4( h->dct.luma4x4[idx], dct4x4 );
133 h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4IY], i_qp );
134 h->dctf.add4x4_idct( p_dst, dct4x4 );
138 #define STORE_8x8_NNZ(idx,nz)\
140 M16( &h->mb.cache.non_zero_count[x264_scan8[idx*4+0]] ) = nz * 0x0101;\
141 M16( &h->mb.cache.non_zero_count[x264_scan8[idx*4+2]] ) = nz * 0x0101;\
144 #define CLEAR_16x16_NNZ \
146 M32( &h->mb.cache.non_zero_count[x264_scan8[ 0]] ) = 0;\
147 M32( &h->mb.cache.non_zero_count[x264_scan8[ 2]] ) = 0;\
148 M32( &h->mb.cache.non_zero_count[x264_scan8[ 8]] ) = 0;\
149 M32( &h->mb.cache.non_zero_count[x264_scan8[10]] ) = 0;\
152 void x264_mb_encode_i8x8( x264_t *h, int idx, int i_qp )
155 int y = 8 * (idx>>1);
157 uint8_t *p_src = &h->mb.pic.p_fenc[0][x+y*FENC_STRIDE];
158 uint8_t *p_dst = &h->mb.pic.p_fdec[0][x+y*FDEC_STRIDE];
159 ALIGNED_ARRAY_16( int16_t, dct8x8,[64] );
161 if( h->mb.b_lossless )
163 nz = h->zigzagf.sub_8x8( h->dct.luma8x8[idx], p_src, p_dst );
164 STORE_8x8_NNZ(idx,nz);
165 h->mb.i_cbp_luma |= nz<<idx;
169 h->dctf.sub8x8_dct8( dct8x8, p_src, p_dst );
171 nz = x264_quant_8x8( h, dct8x8, i_qp, 1, idx );
174 h->mb.i_cbp_luma |= 1<<idx;
175 h->zigzagf.scan_8x8( h->dct.luma8x8[idx], dct8x8 );
176 h->quantf.dequant_8x8( dct8x8, h->dequant8_mf[CQM_8IY], i_qp );
177 h->dctf.add8x8_idct8( p_dst, dct8x8 );
178 STORE_8x8_NNZ(idx,1);
181 STORE_8x8_NNZ(idx,0);
184 static void x264_mb_encode_i16x16( x264_t *h, int i_qp )
186 uint8_t *p_src = h->mb.pic.p_fenc[0];
187 uint8_t *p_dst = h->mb.pic.p_fdec[0];
189 ALIGNED_ARRAY_16( int16_t, dct4x4,[16],[16] );
190 ALIGNED_ARRAY_16( int16_t, dct_dc4x4,[16] );
193 int decimate_score = h->mb.b_dct_decimate ? 0 : 9;
195 if( h->mb.b_lossless )
197 for( int i = 0; i < 16; i++ )
199 int oe = block_idx_xy_fenc[i];
200 int od = block_idx_xy_fdec[i];
201 nz = h->zigzagf.sub_4x4ac( h->dct.luma4x4[i], p_src+oe, p_dst+od, &dct_dc4x4[block_idx_yx_1d[i]] );
202 h->mb.cache.non_zero_count[x264_scan8[i]] = nz;
203 h->mb.i_cbp_luma |= nz;
205 h->mb.i_cbp_luma *= 0xf;
206 h->mb.cache.non_zero_count[x264_scan8[24]] = array_non_zero( dct_dc4x4 );
207 h->zigzagf.scan_4x4( h->dct.luma16x16_dc, dct_dc4x4 );
211 h->dctf.sub16x16_dct( dct4x4, p_src, p_dst );
213 for( int i = 0; i < 16; i++ )
216 dct_dc4x4[block_idx_xy_1d[i]] = dct4x4[i][0];
219 /* quant/scan/dequant */
220 nz = x264_quant_4x4( h, dct4x4[i], i_qp, DCT_LUMA_AC, 1, i );
221 h->mb.cache.non_zero_count[x264_scan8[i]] = nz;
224 h->zigzagf.scan_4x4( h->dct.luma4x4[i], dct4x4[i] );
225 h->quantf.dequant_4x4( dct4x4[i], h->dequant4_mf[CQM_4IY], i_qp );
226 if( decimate_score < 6 ) decimate_score += h->quantf.decimate_score15( h->dct.luma4x4[i] );
227 h->mb.i_cbp_luma = 0xf;
231 /* Writing the 16 CBFs in an i16x16 block is quite costly, so decimation can save many bits. */
232 /* More useful with CAVLC, but still useful with CABAC. */
233 if( decimate_score < 6 )
235 h->mb.i_cbp_luma = 0;
239 h->dctf.dct4x4dc( dct_dc4x4 );
240 if( h->mb.b_trellis )
241 nz = x264_quant_dc_trellis( h, dct_dc4x4, CQM_4IY, i_qp, DCT_LUMA_DC, 1, 0 );
243 nz = 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 );
245 h->mb.cache.non_zero_count[x264_scan8[24]] = nz;
248 h->zigzagf.scan_4x4( h->dct.luma16x16_dc, dct_dc4x4 );
250 /* output samples to fdec */
251 h->dctf.idct4x4dc( dct_dc4x4 );
252 h->quantf.dequant_4x4_dc( dct_dc4x4, h->dequant4_mf[CQM_4IY], i_qp ); /* XXX not inversed */
253 if( h->mb.i_cbp_luma )
254 for( int i = 0; i < 16; i++ )
255 dct4x4[i][0] = dct_dc4x4[block_idx_xy_1d[i]];
258 /* put pixels to fdec */
259 if( h->mb.i_cbp_luma )
260 h->dctf.add16x16_idct( p_dst, dct4x4 );
262 h->dctf.add16x16_idct_dc( p_dst, dct_dc4x4 );
265 static inline int idct_dequant_round_2x2_dc( int16_t ref[4], int16_t dct[4], int dequant_mf[6][16], int i_qp )
268 idct_dequant_2x2_dconly( out, dct, dequant_mf, i_qp );
269 return ((ref[0] ^ (out[0]+32))
270 | (ref[1] ^ (out[1]+32))
271 | (ref[2] ^ (out[2]+32))
272 | (ref[3] ^ (out[3]+32))) >> 6;
275 /* Round down coefficients losslessly in DC-only chroma blocks.
276 * Unlike luma blocks, this can't be done with a lookup table or
277 * other shortcut technique because of the interdependencies
278 * between the coefficients due to the chroma DC transform. */
279 static inline int x264_mb_optimize_chroma_dc( x264_t *h, int b_inter, int i_qp, int16_t dct2x2[4] )
281 int16_t dct2x2_orig[4];
284 /* If the QP is too high, there's no benefit to rounding optimization. */
285 if( h->dequant4_mf[CQM_4IC + b_inter][i_qp%6][0] << (i_qp/6) > 32*64 )
288 idct_dequant_2x2_dconly( dct2x2_orig, dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
289 dct2x2_orig[0] += 32;
290 dct2x2_orig[1] += 32;
291 dct2x2_orig[2] += 32;
292 dct2x2_orig[3] += 32;
294 /* If the DC coefficients already round to zero, terminate early. */
295 if( !((dct2x2_orig[0]|dct2x2_orig[1]|dct2x2_orig[2]|dct2x2_orig[3])>>6) )
298 /* Start with the highest frequency coefficient... is this the best option? */
299 for( nz = 0, coeff = h->quantf.coeff_last[DCT_CHROMA_DC]( dct2x2 ); coeff >= 0; coeff-- )
301 int level = dct2x2[coeff];
302 int sign = level>>31 | 1; /* dct2x2[coeff] < 0 ? -1 : 1 */
306 dct2x2[coeff] = level - sign;
307 if( idct_dequant_round_2x2_dc( dct2x2_orig, dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qp ) )
310 dct2x2[coeff] = level;
320 void x264_mb_encode_8x8_chroma( x264_t *h, int b_inter, int i_qp )
323 int b_decimate = b_inter && h->mb.b_dct_decimate;
324 ALIGNED_ARRAY_16( int16_t, dct2x2,[4] );
325 h->mb.i_cbp_chroma = 0;
327 /* Early termination: check variance of chroma residual before encoding.
328 * Don't bother trying early termination at low QPs.
329 * Values are experimentally derived. */
330 if( b_decimate && i_qp >= (h->mb.b_trellis ? 12 : 18) )
332 int thresh = (x264_lambda2_tab[i_qp] + 32) >> 6;
334 int score = h->pixf.var2_8x8( h->mb.pic.p_fenc[1], FENC_STRIDE, h->mb.pic.p_fdec[1], FDEC_STRIDE, &ssd[0] );
335 if( score < thresh*4 )
336 score += h->pixf.var2_8x8( h->mb.pic.p_fenc[2], FENC_STRIDE, h->mb.pic.p_fdec[2], FDEC_STRIDE, &ssd[1] );
337 if( score < thresh*4 )
339 h->mb.cache.non_zero_count[x264_scan8[16]] = 0;
340 h->mb.cache.non_zero_count[x264_scan8[17]] = 0;
341 h->mb.cache.non_zero_count[x264_scan8[18]] = 0;
342 h->mb.cache.non_zero_count[x264_scan8[19]] = 0;
343 h->mb.cache.non_zero_count[x264_scan8[20]] = 0;
344 h->mb.cache.non_zero_count[x264_scan8[21]] = 0;
345 h->mb.cache.non_zero_count[x264_scan8[22]] = 0;
346 h->mb.cache.non_zero_count[x264_scan8[23]] = 0;
347 M16( &h->mb.cache.non_zero_count[x264_scan8[25]] ) = 0;
349 for( int ch = 0; ch < 2; ch++ )
351 if( ssd[ch] > thresh )
353 h->dctf.sub8x8_dct_dc( dct2x2, h->mb.pic.p_fenc[1+ch], h->mb.pic.p_fdec[1+ch] );
354 if( h->mb.b_trellis )
355 nz_dc = x264_quant_dc_trellis( h, dct2x2, CQM_4IC+b_inter, i_qp, DCT_CHROMA_DC, !b_inter, 1 );
357 nz_dc = 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 );
361 if( !x264_mb_optimize_chroma_dc( h, b_inter, i_qp, dct2x2 ) )
363 h->mb.cache.non_zero_count[x264_scan8[25]+ch] = 1;
364 zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct2x2 );
365 idct_dequant_2x2_dconly( dct2x2, dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
366 h->dctf.add8x8_idct_dc( h->mb.pic.p_fdec[1+ch], dct2x2 );
367 h->mb.i_cbp_chroma = 1;
375 for( int ch = 0; ch < 2; ch++ )
377 uint8_t *p_src = h->mb.pic.p_fenc[1+ch];
378 uint8_t *p_dst = h->mb.pic.p_fdec[1+ch];
379 int i_decimate_score = 0;
382 ALIGNED_ARRAY_16( int16_t, dct4x4,[4],[16] );
384 if( h->mb.b_lossless )
386 for( int i = 0; i < 4; i++ )
388 int oe = block_idx_x[i]*4 + block_idx_y[i]*4*FENC_STRIDE;
389 int od = block_idx_x[i]*4 + block_idx_y[i]*4*FDEC_STRIDE;
390 nz = h->zigzagf.sub_4x4ac( h->dct.luma4x4[16+i+ch*4], p_src+oe, p_dst+od, &h->dct.chroma_dc[ch][i] );
391 h->mb.cache.non_zero_count[x264_scan8[16+i+ch*4]] = nz;
392 h->mb.i_cbp_chroma |= nz;
394 h->mb.cache.non_zero_count[x264_scan8[25]+ch] = array_non_zero( h->dct.chroma_dc[ch] );
398 h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
399 dct2x2dc( dct2x2, dct4x4 );
400 /* calculate dct coeffs */
401 for( int i = 0; i < 4; i++ )
403 if( h->mb.b_trellis )
404 nz = x264_quant_4x4_trellis( h, dct4x4[i], CQM_4IC+b_inter, i_qp, DCT_CHROMA_AC, !b_inter, 1, 0 );
406 nz = h->quantf.quant_4x4( dct4x4[i], h->quant4_mf[CQM_4IC+b_inter][i_qp], h->quant4_bias[CQM_4IC+b_inter][i_qp] );
407 h->mb.cache.non_zero_count[x264_scan8[16+i+ch*4]] = nz;
411 h->zigzagf.scan_4x4( h->dct.luma4x4[16+i+ch*4], dct4x4[i] );
412 h->quantf.dequant_4x4( dct4x4[i], h->dequant4_mf[CQM_4IC + b_inter], i_qp );
414 i_decimate_score += h->quantf.decimate_score15( h->dct.luma4x4[16+i+ch*4] );
418 if( h->mb.b_trellis )
419 nz_dc = x264_quant_dc_trellis( h, dct2x2, CQM_4IC+b_inter, i_qp, DCT_CHROMA_DC, !b_inter, 1 );
421 nz_dc = 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 );
423 h->mb.cache.non_zero_count[x264_scan8[25]+ch] = nz_dc;
425 if( (b_decimate && i_decimate_score < 7) || !nz_ac )
427 /* Decimate the block */
428 h->mb.cache.non_zero_count[x264_scan8[16+0]+24*ch] = 0;
429 h->mb.cache.non_zero_count[x264_scan8[16+1]+24*ch] = 0;
430 h->mb.cache.non_zero_count[x264_scan8[16+2]+24*ch] = 0;
431 h->mb.cache.non_zero_count[x264_scan8[16+3]+24*ch] = 0;
432 if( !nz_dc ) /* Whole block is empty */
434 if( !x264_mb_optimize_chroma_dc( h, b_inter, i_qp, dct2x2 ) )
436 h->mb.cache.non_zero_count[x264_scan8[25]+ch] = 0;
440 zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct2x2 );
441 idct_dequant_2x2_dconly( dct2x2, dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
442 h->dctf.add8x8_idct_dc( p_dst, dct2x2 );
446 h->mb.i_cbp_chroma = 1;
449 zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct2x2 );
450 idct_dequant_2x2_dc( dct2x2, dct4x4, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
452 h->dctf.add8x8_idct( p_dst, dct4x4 );
456 /* 0 = none, 1 = DC only, 2 = DC+AC */
457 h->mb.i_cbp_chroma = ((!!M16( &h->mb.cache.non_zero_count[x264_scan8[25]] )) | h->mb.i_cbp_chroma) + h->mb.i_cbp_chroma;
460 static void x264_macroblock_encode_skip( x264_t *h )
462 M32( &h->mb.cache.non_zero_count[x264_scan8[0]+0*8] ) = 0;
463 M32( &h->mb.cache.non_zero_count[x264_scan8[0]+1*8] ) = 0;
464 M32( &h->mb.cache.non_zero_count[x264_scan8[0]+2*8] ) = 0;
465 M32( &h->mb.cache.non_zero_count[x264_scan8[0]+3*8] ) = 0;
466 for( int i = 16; i < 24; i++ )
467 h->mb.cache.non_zero_count[x264_scan8[i]] = 0;
468 h->mb.i_cbp_luma = 0;
469 h->mb.i_cbp_chroma = 0;
470 h->mb.cbp[h->mb.i_mb_xy] = 0;
473 /*****************************************************************************
474 * x264_macroblock_encode_pskip:
475 * Encode an already marked skip block
476 *****************************************************************************/
477 static void x264_macroblock_encode_pskip( x264_t *h )
479 /* don't do pskip motion compensation if it was already done in macroblock_analyse */
480 if( !h->mb.b_skip_mc )
482 int mvx = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][0],
483 h->mb.mv_min[0], h->mb.mv_max[0] );
484 int mvy = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][1],
485 h->mb.mv_min[1], h->mb.mv_max[1] );
487 h->mc.mc_luma( h->mb.pic.p_fdec[0], FDEC_STRIDE,
488 h->mb.pic.p_fref[0][0], h->mb.pic.i_stride[0],
489 mvx, mvy, 16, 16, &h->sh.weight[0][0] );
491 /* Special case for mv0, which is (of course) very common in P-skip mode. */
494 h->mc.mc_chroma( h->mb.pic.p_fdec[1], FDEC_STRIDE,
495 h->mb.pic.p_fref[0][0][4], h->mb.pic.i_stride[1],
497 h->mc.mc_chroma( h->mb.pic.p_fdec[2], FDEC_STRIDE,
498 h->mb.pic.p_fref[0][0][5], h->mb.pic.i_stride[2],
503 h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fref[0][0][4], h->mb.pic.i_stride[1], 8 );
504 h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fref[0][0][5], h->mb.pic.i_stride[2], 8 );
507 if( h->sh.weight[0][1].weightfn )
508 h->sh.weight[0][1].weightfn[8>>2]( h->mb.pic.p_fdec[1], FDEC_STRIDE,
509 h->mb.pic.p_fdec[1], FDEC_STRIDE,
510 &h->sh.weight[0][1], 8 );
512 if( h->sh.weight[0][2].weightfn )
513 h->sh.weight[0][2].weightfn[8>>2]( h->mb.pic.p_fdec[2], FDEC_STRIDE,
514 h->mb.pic.p_fdec[2], FDEC_STRIDE,
515 &h->sh.weight[0][2], 8 );
518 x264_macroblock_encode_skip( h );
521 /*****************************************************************************
522 * Intra prediction for predictive lossless mode.
523 *****************************************************************************/
525 /* Note that these functions take a shortcut (mc.copy instead of actual pixel prediction) which assumes
526 * that the edge pixels of the reconstructed frame are the same as that of the source frame. This means
527 * they will only work correctly if the neighboring blocks are losslessly coded. In practice, this means
528 * lossless mode cannot be mixed with lossy mode within a frame. */
529 /* This can be resolved by explicitly copying the edge pixels after doing the mc.copy, but this doesn't
530 * need to be done unless we decide to allow mixing lossless and lossy compression. */
532 void x264_predict_lossless_8x8_chroma( x264_t *h, int i_mode )
534 int stride = h->fenc->i_stride[1] << h->mb.b_interlaced;
535 if( i_mode == I_PRED_CHROMA_V )
537 h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fenc_plane[1]-stride, stride, 8 );
538 h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fenc_plane[2]-stride, stride, 8 );
540 else if( i_mode == I_PRED_CHROMA_H )
542 h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fenc_plane[1]-1, stride, 8 );
543 h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fenc_plane[2]-1, stride, 8 );
547 h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[1] );
548 h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[2] );
552 void x264_predict_lossless_4x4( x264_t *h, uint8_t *p_dst, int idx, int i_mode )
554 int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
555 uint8_t *p_src = h->mb.pic.p_fenc_plane[0] + block_idx_x[idx]*4 + block_idx_y[idx]*4 * stride;
557 if( i_mode == I_PRED_4x4_V )
558 h->mc.copy[PIXEL_4x4]( p_dst, FDEC_STRIDE, p_src-stride, stride, 4 );
559 else if( i_mode == I_PRED_4x4_H )
560 h->mc.copy[PIXEL_4x4]( p_dst, FDEC_STRIDE, p_src-1, stride, 4 );
562 h->predict_4x4[i_mode]( p_dst );
565 void x264_predict_lossless_8x8( x264_t *h, uint8_t *p_dst, int idx, int i_mode, uint8_t edge[33] )
567 int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
568 uint8_t *p_src = h->mb.pic.p_fenc_plane[0] + (idx&1)*8 + (idx>>1)*8*stride;
570 if( i_mode == I_PRED_8x8_V )
571 h->mc.copy[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src-stride, stride, 8 );
572 else if( i_mode == I_PRED_8x8_H )
573 h->mc.copy[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src-1, stride, 8 );
575 h->predict_8x8[i_mode]( p_dst, edge );
578 void x264_predict_lossless_16x16( x264_t *h, int i_mode )
580 int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
581 if( i_mode == I_PRED_16x16_V )
582 h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.p_fenc_plane[0]-stride, stride, 16 );
583 else if( i_mode == I_PRED_16x16_H )
584 h->mc.copy_16x16_unaligned( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.p_fenc_plane[0]-1, stride, 16 );
586 h->predict_16x16[i_mode]( h->mb.pic.p_fdec[0] );
589 /*****************************************************************************
590 * x264_macroblock_encode:
591 *****************************************************************************/
592 void x264_macroblock_encode( x264_t *h )
594 int i_qp = h->mb.i_qp;
595 int b_decimate = h->mb.b_dct_decimate;
596 int b_force_no_skip = 0;
598 h->mb.i_cbp_luma = 0;
599 h->mb.cache.non_zero_count[x264_scan8[24]] = 0;
601 if( h->mb.i_type == I_PCM )
603 /* if PCM is chosen, we need to store reconstructed frame data */
604 h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.p_fenc[0], FENC_STRIDE, 16 );
605 h->mc.copy[PIXEL_8x8] ( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fenc[1], FENC_STRIDE, 8 );
606 h->mc.copy[PIXEL_8x8] ( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fenc[2], FENC_STRIDE, 8 );
611 && h->mb.i_mb_xy == h->sh.i_first_mb + h->mb.i_mb_stride
612 && IS_SKIP(h->mb.type[h->sh.i_first_mb]) )
614 /* The first skip is predicted to be a frame mb pair.
615 * We don't yet support the aff part of mbaff, so force it to non-skip
616 * so that we can pick the aff flag. */
618 if( IS_SKIP(h->mb.i_type) )
620 if( h->mb.i_type == P_SKIP )
622 else if( h->mb.i_type == B_SKIP )
623 h->mb.i_type = B_DIRECT;
627 if( h->mb.i_type == P_SKIP )
630 x264_macroblock_encode_pskip( h );
633 if( h->mb.i_type == B_SKIP )
635 /* don't do bskip motion compensation if it was already done in macroblock_analyse */
636 if( !h->mb.b_skip_mc )
638 x264_macroblock_encode_skip( h );
642 if( h->mb.i_type == I_16x16 )
644 const int i_mode = h->mb.i_intra16x16_pred_mode;
645 h->mb.b_transform_8x8 = 0;
647 if( h->mb.b_lossless )
648 x264_predict_lossless_16x16( h, i_mode );
650 h->predict_16x16[i_mode]( h->mb.pic.p_fdec[0] );
652 /* encode the 16x16 macroblock */
653 x264_mb_encode_i16x16( h, i_qp );
655 else if( h->mb.i_type == I_8x8 )
657 ALIGNED_ARRAY_16( uint8_t, edge,[33] );
658 h->mb.b_transform_8x8 = 1;
659 /* If we already encoded 3 of the 4 i8x8 blocks, we don't have to do them again. */
660 if( h->mb.i_skip_intra )
662 h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i8x8_fdec_buf, 16, 16 );
663 M32( &h->mb.cache.non_zero_count[x264_scan8[ 0]] ) = h->mb.pic.i8x8_nnz_buf[0];
664 M32( &h->mb.cache.non_zero_count[x264_scan8[ 2]] ) = h->mb.pic.i8x8_nnz_buf[1];
665 M32( &h->mb.cache.non_zero_count[x264_scan8[ 8]] ) = h->mb.pic.i8x8_nnz_buf[2];
666 M32( &h->mb.cache.non_zero_count[x264_scan8[10]] ) = h->mb.pic.i8x8_nnz_buf[3];
667 h->mb.i_cbp_luma = h->mb.pic.i8x8_cbp;
668 /* In RD mode, restore the now-overwritten DCT data. */
669 if( h->mb.i_skip_intra == 2 )
670 h->mc.memcpy_aligned( h->dct.luma8x8, h->mb.pic.i8x8_dct_buf, sizeof(h->mb.pic.i8x8_dct_buf) );
672 for( int i = h->mb.i_skip_intra ? 3 : 0 ; i < 4; i++ )
674 uint8_t *p_dst = &h->mb.pic.p_fdec[0][8 * (i&1) + 8 * (i>>1) * FDEC_STRIDE];
675 int i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[4*i]];
676 h->predict_8x8_filter( p_dst, edge, h->mb.i_neighbour8[i], x264_pred_i4x4_neighbors[i_mode] );
678 if( h->mb.b_lossless )
679 x264_predict_lossless_8x8( h, p_dst, i, i_mode, edge );
681 h->predict_8x8[i_mode]( p_dst, edge );
683 x264_mb_encode_i8x8( h, i, i_qp );
686 else if( h->mb.i_type == I_4x4 )
688 h->mb.b_transform_8x8 = 0;
689 /* If we already encoded 15 of the 16 i4x4 blocks, we don't have to do them again. */
690 if( h->mb.i_skip_intra )
692 h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i4x4_fdec_buf, 16, 16 );
693 M32( &h->mb.cache.non_zero_count[x264_scan8[ 0]] ) = h->mb.pic.i4x4_nnz_buf[0];
694 M32( &h->mb.cache.non_zero_count[x264_scan8[ 2]] ) = h->mb.pic.i4x4_nnz_buf[1];
695 M32( &h->mb.cache.non_zero_count[x264_scan8[ 8]] ) = h->mb.pic.i4x4_nnz_buf[2];
696 M32( &h->mb.cache.non_zero_count[x264_scan8[10]] ) = h->mb.pic.i4x4_nnz_buf[3];
697 h->mb.i_cbp_luma = h->mb.pic.i4x4_cbp;
698 /* In RD mode, restore the now-overwritten DCT data. */
699 if( h->mb.i_skip_intra == 2 )
700 h->mc.memcpy_aligned( h->dct.luma4x4, h->mb.pic.i4x4_dct_buf, sizeof(h->mb.pic.i4x4_dct_buf) );
702 for( int i = h->mb.i_skip_intra ? 15 : 0 ; i < 16; i++ )
704 uint8_t *p_dst = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[i]];
705 int i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[i]];
707 if( (h->mb.i_neighbour4[i] & (MB_TOPRIGHT|MB_TOP)) == MB_TOP )
708 /* emulate missing topright samples */
709 M32( &p_dst[4-FDEC_STRIDE] ) = p_dst[3-FDEC_STRIDE] * 0x01010101U;
711 if( h->mb.b_lossless )
712 x264_predict_lossless_4x4( h, p_dst, i, i_mode );
714 h->predict_4x4[i_mode]( p_dst );
715 x264_mb_encode_i4x4( h, i, i_qp );
720 int i_decimate_mb = 0;
722 /* Don't repeat motion compensation if it was already done in non-RD transform analysis */
723 if( !h->mb.b_skip_mc )
726 if( h->mb.b_lossless )
728 if( h->mb.b_transform_8x8 )
729 for( int i8x8 = 0; i8x8 < 4; i8x8++ )
733 nz = h->zigzagf.sub_8x8( h->dct.luma8x8[i8x8],
734 h->mb.pic.p_fenc[0]+x+y*FENC_STRIDE,
735 h->mb.pic.p_fdec[0]+x+y*FDEC_STRIDE );
736 STORE_8x8_NNZ(i8x8,nz);
737 h->mb.i_cbp_luma |= nz << i8x8;
740 for( int i4x4 = 0; i4x4 < 16; i4x4++ )
742 nz = h->zigzagf.sub_4x4( h->dct.luma4x4[i4x4],
743 h->mb.pic.p_fenc[0]+block_idx_xy_fenc[i4x4],
744 h->mb.pic.p_fdec[0]+block_idx_xy_fdec[i4x4] );
745 h->mb.cache.non_zero_count[x264_scan8[i4x4]] = nz;
746 h->mb.i_cbp_luma |= nz << (i4x4>>2);
749 else if( h->mb.b_transform_8x8 )
751 ALIGNED_ARRAY_16( int16_t, dct8x8,[4],[64] );
752 b_decimate &= !h->mb.b_trellis; // 8x8 trellis is inherently optimal decimation
753 h->dctf.sub16x16_dct8( dct8x8, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
754 h->nr_count[1] += h->mb.b_noise_reduction * 4;
756 for( int idx = 0; idx < 4; idx++ )
758 if( h->mb.b_noise_reduction )
759 h->quantf.denoise_dct( dct8x8[idx], h->nr_residual_sum[1], h->nr_offset[1], 64 );
760 nz = x264_quant_8x8( h, dct8x8[idx], i_qp, 0, idx );
764 h->zigzagf.scan_8x8( h->dct.luma8x8[idx], dct8x8[idx] );
767 int i_decimate_8x8 = h->quantf.decimate_score64( h->dct.luma8x8[idx] );
768 i_decimate_mb += i_decimate_8x8;
769 if( i_decimate_8x8 >= 4 )
770 h->mb.i_cbp_luma |= 1<<idx;
773 h->mb.i_cbp_luma |= 1<<idx;
777 if( i_decimate_mb < 6 && b_decimate )
779 h->mb.i_cbp_luma = 0;
784 for( int idx = 0; idx < 4; idx++ )
786 if( h->mb.i_cbp_luma&(1<<idx) )
788 h->quantf.dequant_8x8( dct8x8[idx], h->dequant8_mf[CQM_8PY], i_qp );
789 h->dctf.add8x8_idct8( &h->mb.pic.p_fdec[0][(idx&1)*8 + (idx>>1)*8*FDEC_STRIDE], dct8x8[idx] );
790 STORE_8x8_NNZ(idx,1);
793 STORE_8x8_NNZ(idx,0);
799 ALIGNED_ARRAY_16( int16_t, dct4x4,[16],[16] );
800 h->dctf.sub16x16_dct( dct4x4, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
801 h->nr_count[0] += h->mb.b_noise_reduction * 16;
803 for( int i8x8 = 0; i8x8 < 4; i8x8++ )
805 int i_decimate_8x8 = 0;
808 /* encode one 4x4 block */
809 for( int i4x4 = 0; i4x4 < 4; i4x4++ )
811 int idx = i8x8 * 4 + i4x4;
813 if( h->mb.b_noise_reduction )
814 h->quantf.denoise_dct( dct4x4[idx], h->nr_residual_sum[0], h->nr_offset[0], 16 );
815 nz = x264_quant_4x4( h, dct4x4[idx], i_qp, DCT_LUMA_4x4, 0, idx );
816 h->mb.cache.non_zero_count[x264_scan8[idx]] = nz;
820 h->zigzagf.scan_4x4( h->dct.luma4x4[idx], dct4x4[idx] );
821 h->quantf.dequant_4x4( dct4x4[idx], h->dequant4_mf[CQM_4PY], i_qp );
822 if( b_decimate && i_decimate_8x8 < 6 )
823 i_decimate_8x8 += h->quantf.decimate_score16( h->dct.luma4x4[idx] );
828 /* decimate this 8x8 block */
829 i_decimate_mb += i_decimate_8x8;
832 if( i_decimate_8x8 < 4 )
833 STORE_8x8_NNZ(i8x8,0)
835 h->mb.i_cbp_luma |= 1<<i8x8;
839 h->dctf.add8x8_idct( &h->mb.pic.p_fdec[0][(i8x8&1)*8 + (i8x8>>1)*8*FDEC_STRIDE], &dct4x4[i8x8*4] );
840 h->mb.i_cbp_luma |= 1<<i8x8;
846 if( i_decimate_mb < 6 )
848 h->mb.i_cbp_luma = 0;
853 for( int i8x8 = 0; i8x8 < 4; i8x8++ )
854 if( h->mb.i_cbp_luma&(1<<i8x8) )
855 h->dctf.add8x8_idct( &h->mb.pic.p_fdec[0][(i8x8&1)*8 + (i8x8>>1)*8*FDEC_STRIDE], &dct4x4[i8x8*4] );
862 if( IS_INTRA( h->mb.i_type ) )
864 const int i_mode = h->mb.i_chroma_pred_mode;
865 if( h->mb.b_lossless )
866 x264_predict_lossless_8x8_chroma( h, i_mode );
869 h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[1] );
870 h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[2] );
874 /* encode the 8x8 blocks */
875 x264_mb_encode_8x8_chroma( h, !IS_INTRA( h->mb.i_type ), h->mb.i_chroma_qp );
878 int cbp = h->mb.i_cbp_chroma << 4 | h->mb.i_cbp_luma;
879 if( h->param.b_cabac )
880 cbp |= h->mb.cache.non_zero_count[x264_scan8[24]] << 8
881 | h->mb.cache.non_zero_count[x264_scan8[25]] << 9
882 | h->mb.cache.non_zero_count[x264_scan8[26]] << 10;
883 h->mb.cbp[h->mb.i_mb_xy] = cbp;
886 * XXX: in the me perhaps we should take x264_mb_predict_mv_pskip into account
887 * (if multiple mv give same result)*/
888 if( !b_force_no_skip )
890 if( h->mb.i_type == P_L0 && h->mb.i_partition == D_16x16 &&
891 !(h->mb.i_cbp_luma | h->mb.i_cbp_chroma) &&
892 M32( h->mb.cache.mv[0][x264_scan8[0]] ) == M32( h->mb.cache.pskip_mv )
893 && h->mb.cache.ref[0][x264_scan8[0]] == 0 )
895 h->mb.i_type = P_SKIP;
898 /* Check for B_SKIP */
899 if( h->mb.i_type == B_DIRECT && !(h->mb.i_cbp_luma | h->mb.i_cbp_chroma) )
901 h->mb.i_type = B_SKIP;
906 /*****************************************************************************
907 * x264_macroblock_probe_skip:
908 * Check if the current MB could be encoded as a [PB]_SKIP
909 *****************************************************************************/
910 int x264_macroblock_probe_skip( x264_t *h, int b_bidir )
912 ALIGNED_ARRAY_16( int16_t, dct4x4,[4],[16] );
913 ALIGNED_ARRAY_16( int16_t, dct2x2,[4] );
914 ALIGNED_ARRAY_16( int16_t, dctscan,[16] );
915 ALIGNED_4( int16_t mvp[2] );
917 int i_qp = h->mb.i_qp;
923 mvp[0] = x264_clip3( h->mb.cache.pskip_mv[0], h->mb.mv_min[0], h->mb.mv_max[0] );
924 mvp[1] = x264_clip3( h->mb.cache.pskip_mv[1], h->mb.mv_min[1], h->mb.mv_max[1] );
926 /* Motion compensation */
927 h->mc.mc_luma( h->mb.pic.p_fdec[0], FDEC_STRIDE,
928 h->mb.pic.p_fref[0][0], h->mb.pic.i_stride[0],
929 mvp[0], mvp[1], 16, 16, &h->sh.weight[0][0] );
932 for( int i8x8 = 0, i_decimate_mb = 0; i8x8 < 4; i8x8++ )
934 int fenc_offset = (i8x8&1) * 8 + (i8x8>>1) * FENC_STRIDE * 8;
935 int fdec_offset = (i8x8&1) * 8 + (i8x8>>1) * FDEC_STRIDE * 8;
937 h->dctf.sub8x8_dct( dct4x4, h->mb.pic.p_fenc[0] + fenc_offset,
938 h->mb.pic.p_fdec[0] + fdec_offset );
939 /* encode one 4x4 block */
940 for( int i4x4 = 0; i4x4 < 4; i4x4++ )
942 if( !h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] ) )
944 h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
945 i_decimate_mb += h->quantf.decimate_score16( dctscan );
946 if( i_decimate_mb >= 6 )
952 i_qp = h->mb.i_chroma_qp;
953 thresh = (x264_lambda2_tab[i_qp] + 32) >> 6;
955 for( int ch = 0; ch < 2; ch++ )
957 uint8_t *p_src = h->mb.pic.p_fenc[1+ch];
958 uint8_t *p_dst = h->mb.pic.p_fdec[1+ch];
962 /* Special case for mv0, which is (of course) very common in P-skip mode. */
965 h->mc.mc_chroma( h->mb.pic.p_fdec[1+ch], FDEC_STRIDE,
966 h->mb.pic.p_fref[0][0][4+ch], h->mb.pic.i_stride[1+ch],
967 mvp[0], mvp[1], 8, 8 );
970 h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[1+ch], FDEC_STRIDE, h->mb.pic.p_fref[0][0][4+ch], h->mb.pic.i_stride[1+ch], 8 );
972 if( h->sh.weight[0][1+ch].weightfn )
973 h->sh.weight[0][1+ch].weightfn[8>>2]( h->mb.pic.p_fdec[1+ch], FDEC_STRIDE,
974 h->mb.pic.p_fdec[1+ch], FDEC_STRIDE,
975 &h->sh.weight[0][1+ch], 8 );
978 /* there is almost never a termination during chroma, but we can't avoid the check entirely */
979 /* so instead we check SSD and skip the actual check if the score is low enough. */
980 ssd = h->pixf.ssd[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src, FENC_STRIDE );
984 /* The vast majority of chroma checks will terminate during the DC check or the higher
985 * threshold check, so we can save time by doing a DC-only DCT. */
986 h->dctf.sub8x8_dct_dc( dct2x2, p_src, p_dst );
988 if( h->quantf.quant_2x2_dc( dct2x2, h->quant4_mf[CQM_4PC][i_qp][0]>>1, h->quant4_bias[CQM_4PC][i_qp][0]<<1 ) )
991 /* If there wasn't a termination in DC, we can check against a much higher threshold. */
995 h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
997 /* calculate dct coeffs */
998 for( int i4x4 = 0, i_decimate_mb = 0; i4x4 < 4; i4x4++ )
1000 dct4x4[i4x4][0] = 0;
1001 if( !h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] ) )
1003 h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
1004 i_decimate_mb += h->quantf.decimate_score15( dctscan );
1005 if( i_decimate_mb >= 7 )
1010 h->mb.b_skip_mc = 1;
1014 /****************************************************************************
1015 * DCT-domain noise reduction / adaptive deadzone
1017 ****************************************************************************/
1019 void x264_noise_reduction_update( x264_t *h )
1021 for( int cat = 0; cat < 2; cat++ )
1023 int size = cat ? 64 : 16;
1024 const uint16_t *weight = cat ? x264_dct8_weight2_tab : x264_dct4_weight2_tab;
1026 if( h->nr_count[cat] > (cat ? (1<<16) : (1<<18)) )
1028 for( int i = 0; i < size; i++ )
1029 h->nr_residual_sum[cat][i] >>= 1;
1030 h->nr_count[cat] >>= 1;
1033 for( int i = 0; i < size; i++ )
1034 h->nr_offset[cat][i] =
1035 ((uint64_t)h->param.analyse.i_noise_reduction * h->nr_count[cat]
1036 + h->nr_residual_sum[cat][i]/2)
1037 / ((uint64_t)h->nr_residual_sum[cat][i] * weight[i]/256 + 1);
1041 /*****************************************************************************
1042 * RD only; 4 calls to this do not make up for one macroblock_encode.
1043 * doesn't transform chroma dc.
1044 *****************************************************************************/
1045 void x264_macroblock_encode_p8x8( x264_t *h, int i8 )
1047 int i_qp = h->mb.i_qp;
1048 uint8_t *p_fenc = h->mb.pic.p_fenc[0] + (i8&1)*8 + (i8>>1)*8*FENC_STRIDE;
1049 uint8_t *p_fdec = h->mb.pic.p_fdec[0] + (i8&1)*8 + (i8>>1)*8*FDEC_STRIDE;
1050 int b_decimate = h->mb.b_dct_decimate;
1054 if( !h->mb.b_skip_mc )
1055 x264_mb_mc_8x8( h, i8 );
1057 if( h->mb.b_lossless )
1059 if( h->mb.b_transform_8x8 )
1061 nnz8x8 = h->zigzagf.sub_8x8( h->dct.luma8x8[i8], p_fenc, p_fdec );
1062 STORE_8x8_NNZ(i8,nnz8x8);
1066 for( int i4 = i8*4; i4 < i8*4+4; i4++ )
1068 nz = h->zigzagf.sub_4x4( h->dct.luma4x4[i4],
1069 h->mb.pic.p_fenc[0]+block_idx_xy_fenc[i4],
1070 h->mb.pic.p_fdec[0]+block_idx_xy_fdec[i4] );
1071 h->mb.cache.non_zero_count[x264_scan8[i4]] = nz;
1075 for( int ch = 0; ch < 2; ch++ )
1078 p_fenc = h->mb.pic.p_fenc[1+ch] + (i8&1)*4 + (i8>>1)*4*FENC_STRIDE;
1079 p_fdec = h->mb.pic.p_fdec[1+ch] + (i8&1)*4 + (i8>>1)*4*FDEC_STRIDE;
1080 nz = h->zigzagf.sub_4x4ac( h->dct.luma4x4[16+i8+ch*4], p_fenc, p_fdec, &dc );
1081 h->mb.cache.non_zero_count[x264_scan8[16+i8+ch*4]] = nz;
1086 if( h->mb.b_transform_8x8 )
1088 ALIGNED_ARRAY_16( int16_t, dct8x8,[64] );
1089 h->dctf.sub8x8_dct8( dct8x8, p_fenc, p_fdec );
1090 nnz8x8 = x264_quant_8x8( h, dct8x8, i_qp, 0, i8 );
1093 h->zigzagf.scan_8x8( h->dct.luma8x8[i8], dct8x8 );
1095 if( b_decimate && !h->mb.b_trellis )
1096 nnz8x8 = 4 <= h->quantf.decimate_score64( h->dct.luma8x8[i8] );
1100 h->quantf.dequant_8x8( dct8x8, h->dequant8_mf[CQM_8PY], i_qp );
1101 h->dctf.add8x8_idct8( p_fdec, dct8x8 );
1102 STORE_8x8_NNZ(i8,1);
1105 STORE_8x8_NNZ(i8,0);
1108 STORE_8x8_NNZ(i8,0);
1112 int i_decimate_8x8 = 0;
1113 ALIGNED_ARRAY_16( int16_t, dct4x4,[4],[16] );
1114 h->dctf.sub8x8_dct( dct4x4, p_fenc, p_fdec );
1115 for( int i4 = 0; i4 < 4; i4++ )
1117 nz = x264_quant_4x4( h, dct4x4[i4], i_qp, DCT_LUMA_4x4, 0, i8*4+i4 );
1118 h->mb.cache.non_zero_count[x264_scan8[i8*4+i4]] = nz;
1121 h->zigzagf.scan_4x4( h->dct.luma4x4[i8*4+i4], dct4x4[i4] );
1122 h->quantf.dequant_4x4( dct4x4[i4], h->dequant4_mf[CQM_4PY], i_qp );
1124 i_decimate_8x8 += h->quantf.decimate_score16( h->dct.luma4x4[i8*4+i4] );
1129 if( b_decimate && i_decimate_8x8 < 4 )
1133 h->dctf.add8x8_idct( p_fdec, dct4x4 );
1135 STORE_8x8_NNZ(i8,0);
1138 i_qp = h->mb.i_chroma_qp;
1140 for( int ch = 0; ch < 2; ch++ )
1142 ALIGNED_ARRAY_16( int16_t, dct4x4,[16] );
1143 p_fenc = h->mb.pic.p_fenc[1+ch] + (i8&1)*4 + (i8>>1)*4*FENC_STRIDE;
1144 p_fdec = h->mb.pic.p_fdec[1+ch] + (i8&1)*4 + (i8>>1)*4*FDEC_STRIDE;
1146 h->dctf.sub4x4_dct( dct4x4, p_fenc, p_fdec );
1149 if( h->mb.b_trellis )
1150 nz = x264_quant_4x4_trellis( h, dct4x4, CQM_4PC, i_qp, DCT_CHROMA_AC, 0, 1, 0 );
1152 nz = h->quantf.quant_4x4( dct4x4, h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] );
1154 h->mb.cache.non_zero_count[x264_scan8[16+i8+ch*4]] = nz;
1157 h->zigzagf.scan_4x4( h->dct.luma4x4[16+i8+ch*4], dct4x4 );
1158 h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4PC], i_qp );
1159 h->dctf.add4x4_idct( p_fdec, dct4x4 );
1163 h->mb.i_cbp_luma &= ~(1 << i8);
1164 h->mb.i_cbp_luma |= nnz8x8 << i8;
1165 h->mb.i_cbp_chroma = 0x02;
1168 /*****************************************************************************
1169 * RD only, luma only
1170 *****************************************************************************/
1171 void x264_macroblock_encode_p4x4( x264_t *h, int i4 )
1173 int i_qp = h->mb.i_qp;
1174 uint8_t *p_fenc = &h->mb.pic.p_fenc[0][block_idx_xy_fenc[i4]];
1175 uint8_t *p_fdec = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[i4]];
1178 /* Don't need motion compensation as this function is only used in qpel-RD, which caches pixel data. */
1180 if( h->mb.b_lossless )
1182 nz = h->zigzagf.sub_4x4( h->dct.luma4x4[i4], p_fenc, p_fdec );
1183 h->mb.cache.non_zero_count[x264_scan8[i4]] = nz;
1187 ALIGNED_ARRAY_16( int16_t, dct4x4,[16] );
1188 h->dctf.sub4x4_dct( dct4x4, p_fenc, p_fdec );
1189 nz = x264_quant_4x4( h, dct4x4, i_qp, DCT_LUMA_4x4, 0, i4 );
1190 h->mb.cache.non_zero_count[x264_scan8[i4]] = nz;
1193 h->zigzagf.scan_4x4( h->dct.luma4x4[i4], dct4x4 );
1194 h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4PY], i_qp );
1195 h->dctf.add4x4_idct( p_fdec, dct4x4 );