1 /*****************************************************************************
2 * macroblock.c: macroblock encoding
3 *****************************************************************************
4 * Copyright (C) 2003-2011 x264 project
6 * Authors: Laurent Aimar <fenrir@via.ecp.fr>
7 * Loren Merritt <lorenm@u.washington.edu>
8 * Fiona Glaser <fiona@x264.com>
9 * Henrik Gramner <hengar-6@student.ltu.se>
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2 of the License, or
14 * (at your option) any later version.
16 * This program is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
19 * GNU General Public License for more details.
21 * You should have received a copy of the GNU General Public License
22 * along with this program; if not, write to the Free Software
23 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02111, USA.
25 * This program is also available under a commercial proprietary license.
26 * For more information, contact us at licensing@x264.com.
27 *****************************************************************************/
29 #include "common/common.h"
30 #include "macroblock.h"
32 /* These chroma DC functions don't have assembly versions and are only used here. */
34 #define ZIG(i,y,x) level[i] = dct[x*2+y];
35 static inline void zigzag_scan_2x2_dc( dctcoef level[4], dctcoef dct[4] )
44 static inline void zigzag_scan_2x4_dc( dctcoef level[8], dctcoef dct[8] )
56 #define IDCT_DEQUANT_2X2_START \
57 int d0 = dct[0] + dct[1]; \
58 int d1 = dct[2] + dct[3]; \
59 int d2 = dct[0] - dct[1]; \
60 int d3 = dct[2] - dct[3]; \
61 int dmf = dequant_mf[i_qp%6][0] << i_qp/6;
63 static inline void idct_dequant_2x2_dc( dctcoef dct[4], dctcoef dct4x4[4][16], int dequant_mf[6][16], int i_qp )
65 IDCT_DEQUANT_2X2_START
66 dct4x4[0][0] = (d0 + d1) * dmf >> 5;
67 dct4x4[1][0] = (d0 - d1) * dmf >> 5;
68 dct4x4[2][0] = (d2 + d3) * dmf >> 5;
69 dct4x4[3][0] = (d2 - d3) * dmf >> 5;
72 static inline void idct_dequant_2x2_dconly( dctcoef dct[4], int dequant_mf[6][16], int i_qp )
74 IDCT_DEQUANT_2X2_START
75 dct[0] = (d0 + d1) * dmf >> 5;
76 dct[1] = (d0 - d1) * dmf >> 5;
77 dct[2] = (d2 + d3) * dmf >> 5;
78 dct[3] = (d2 - d3) * dmf >> 5;
80 #undef IDCT_2X2_DEQUANT_START
82 static inline void dct2x2dc( dctcoef d[4], dctcoef dct4x4[4][16] )
84 int d0 = dct4x4[0][0] + dct4x4[1][0];
85 int d1 = dct4x4[2][0] + dct4x4[3][0];
86 int d2 = dct4x4[0][0] - dct4x4[1][0];
87 int d3 = dct4x4[2][0] - dct4x4[3][0];
98 static ALWAYS_INLINE int array_non_zero( dctcoef *v, int i_count )
102 for( int i = 0; i < i_count; i += 8/sizeof(dctcoef) )
108 for( int i = 0; i < i_count; i += 4/sizeof(dctcoef) )
115 static ALWAYS_INLINE int x264_quant_4x4( x264_t *h, dctcoef dct[16], int i_qp, int ctx_block_cat, int b_intra, int p, int idx )
117 int i_quant_cat = b_intra ? (p?CQM_4IC:CQM_4IY) : (p?CQM_4PC:CQM_4PY);
118 if( h->mb.b_noise_reduction )
119 h->quantf.denoise_dct( dct, h->nr_residual_sum[0+!!p*2], h->nr_offset[0+!!p*2], 16 );
120 if( h->mb.b_trellis )
121 return x264_quant_4x4_trellis( h, dct, i_quant_cat, i_qp, ctx_block_cat, b_intra, !!p, idx+p*16 );
123 return h->quantf.quant_4x4( dct, h->quant4_mf[i_quant_cat][i_qp], h->quant4_bias[i_quant_cat][i_qp] );
126 static ALWAYS_INLINE int x264_quant_8x8( x264_t *h, dctcoef dct[64], int i_qp, int ctx_block_cat, int b_intra, int p, int idx )
128 int i_quant_cat = b_intra ? (p?CQM_8IC:CQM_8IY) : (p?CQM_8PC:CQM_8PY);
129 if( h->mb.b_noise_reduction )
130 h->quantf.denoise_dct( dct, h->nr_residual_sum[1+!!p*2], h->nr_offset[1+!!p*2], 64 );
131 if( h->mb.b_trellis )
132 return x264_quant_8x8_trellis( h, dct, i_quant_cat, i_qp, ctx_block_cat, b_intra, !!p, idx+p*4 );
134 return h->quantf.quant_8x8( dct, h->quant8_mf[i_quant_cat][i_qp], h->quant8_bias[i_quant_cat][i_qp] );
137 /* All encoding functions must output the correct CBP and NNZ values.
138 * The entropy coding functions will check CBP first, then NNZ, before
139 * actually reading the DCT coefficients. NNZ still must be correct even
140 * if CBP is zero because of the use of NNZ values for context selection.
141 * "NNZ" need only be 0 or 1 rather than the exact coefficient count because
142 * that is only needed in CAVLC, and will be calculated by CAVLC's residual
143 * coding and stored as necessary. */
145 /* This means that decimation can be done merely by adjusting the CBP and NNZ
146 * rather than memsetting the coefficients. */
148 void x264_mb_encode_i4x4( x264_t *h, int p, int idx, int i_qp, int i_mode )
151 pixel *p_src = &h->mb.pic.p_fenc[p][block_idx_xy_fenc[idx]];
152 pixel *p_dst = &h->mb.pic.p_fdec[p][block_idx_xy_fdec[idx]];
153 ALIGNED_ARRAY_16( dctcoef, dct4x4,[16] );
155 if( h->mb.b_lossless )
156 x264_predict_lossless_4x4( h, p_dst, p, idx, i_mode );
158 h->predict_4x4[i_mode]( p_dst );
160 if( h->mb.b_lossless )
162 nz = h->zigzagf.sub_4x4( h->dct.luma4x4[p*16+idx], p_src, p_dst );
163 h->mb.cache.non_zero_count[x264_scan8[p*16+idx]] = nz;
164 h->mb.i_cbp_luma |= nz<<(idx>>2);
168 h->dctf.sub4x4_dct( dct4x4, p_src, p_dst );
170 nz = x264_quant_4x4( h, dct4x4, i_qp, ctx_cat_plane[DCT_LUMA_4x4][p], 1, p, idx );
171 h->mb.cache.non_zero_count[x264_scan8[p*16+idx]] = nz;
174 h->mb.i_cbp_luma |= 1<<(idx>>2);
175 h->zigzagf.scan_4x4( h->dct.luma4x4[p*16+idx], dct4x4 );
176 h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[p?CQM_4IC:CQM_4IY], i_qp );
177 h->dctf.add4x4_idct( p_dst, dct4x4 );
181 #define STORE_8x8_NNZ( p, idx, nz )\
184 M16( &h->mb.cache.non_zero_count[x264_scan8[p*16+idx*4]+0] ) = (nz) * 0x0101;\
185 M16( &h->mb.cache.non_zero_count[x264_scan8[p*16+idx*4]+8] ) = (nz) * 0x0101;\
188 #define CLEAR_16x16_NNZ( p ) \
191 M32( &h->mb.cache.non_zero_count[x264_scan8[16*p+ 0]] ) = 0;\
192 M32( &h->mb.cache.non_zero_count[x264_scan8[16*p+ 2]] ) = 0;\
193 M32( &h->mb.cache.non_zero_count[x264_scan8[16*p+ 8]] ) = 0;\
194 M32( &h->mb.cache.non_zero_count[x264_scan8[16*p+10]] ) = 0;\
197 void x264_mb_encode_i8x8( x264_t *h, int p, int idx, int i_qp, int i_mode, pixel *edge )
202 pixel *p_src = &h->mb.pic.p_fenc[p][8*x + 8*y*FENC_STRIDE];
203 pixel *p_dst = &h->mb.pic.p_fdec[p][8*x + 8*y*FDEC_STRIDE];
204 ALIGNED_ARRAY_16( dctcoef, dct8x8,[64] );
205 ALIGNED_ARRAY_32( pixel, edge_buf,[36] );
209 h->predict_8x8_filter( p_dst, edge_buf, h->mb.i_neighbour8[idx], x264_pred_i4x4_neighbors[i_mode] );
213 if( h->mb.b_lossless )
214 x264_predict_lossless_8x8( h, p_dst, p, idx, i_mode, edge );
216 h->predict_8x8[i_mode]( p_dst, edge );
218 if( h->mb.b_lossless )
220 nz = h->zigzagf.sub_8x8( h->dct.luma8x8[p*4+idx], p_src, p_dst );
221 STORE_8x8_NNZ( p, idx, nz );
222 h->mb.i_cbp_luma |= nz<<idx;
226 h->dctf.sub8x8_dct8( dct8x8, p_src, p_dst );
228 nz = x264_quant_8x8( h, dct8x8, i_qp, ctx_cat_plane[DCT_LUMA_8x8][p], 1, p, idx );
231 h->mb.i_cbp_luma |= 1<<idx;
232 h->zigzagf.scan_8x8( h->dct.luma8x8[p*4+idx], dct8x8 );
233 h->quantf.dequant_8x8( dct8x8, h->dequant8_mf[p?CQM_8IC:CQM_8IY], i_qp );
234 h->dctf.add8x8_idct8( p_dst, dct8x8 );
235 STORE_8x8_NNZ( p, idx, 1 );
238 STORE_8x8_NNZ( p, idx, 0 );
241 static void x264_mb_encode_i16x16( x264_t *h, int p, int i_qp )
243 pixel *p_src = h->mb.pic.p_fenc[p];
244 pixel *p_dst = h->mb.pic.p_fdec[p];
246 ALIGNED_ARRAY_16( dctcoef, dct4x4,[16],[16] );
247 ALIGNED_ARRAY_16( dctcoef, dct_dc4x4,[16] );
249 int nz, block_cbp = 0;
250 int decimate_score = h->mb.b_dct_decimate ? 0 : 9;
251 int i_quant_cat = p ? CQM_4IC : CQM_4IY;
252 int i_mode = h->mb.i_intra16x16_pred_mode;
254 if( h->mb.b_lossless )
255 x264_predict_lossless_16x16( h, p, i_mode );
257 h->predict_16x16[i_mode]( h->mb.pic.p_fdec[p] );
259 if( h->mb.b_lossless )
261 for( int i = 0; i < 16; i++ )
263 int oe = block_idx_xy_fenc[i];
264 int od = block_idx_xy_fdec[i];
265 nz = h->zigzagf.sub_4x4ac( h->dct.luma4x4[16*p+i], p_src+oe, p_dst+od, &dct_dc4x4[block_idx_yx_1d[i]] );
266 h->mb.cache.non_zero_count[x264_scan8[16*p+i]] = nz;
269 h->mb.i_cbp_luma |= block_cbp * 0xf;
270 h->mb.cache.non_zero_count[x264_scan8[LUMA_DC+p]] = array_non_zero( dct_dc4x4, 16 );
271 h->zigzagf.scan_4x4( h->dct.luma16x16_dc[p], dct_dc4x4 );
275 h->dctf.sub16x16_dct( dct4x4, p_src, p_dst );
277 for( int i = 0; i < 16; i++ )
280 if( h->mb.b_noise_reduction )
281 h->quantf.denoise_dct( dct4x4[i], h->nr_residual_sum[0], h->nr_offset[0], 16 );
282 dct_dc4x4[block_idx_xy_1d[i]] = dct4x4[i][0];
285 /* quant/scan/dequant */
286 if( h->mb.b_trellis )
287 nz = x264_quant_4x4_trellis( h, dct4x4[i], i_quant_cat, i_qp, ctx_cat_plane[DCT_LUMA_AC][p], 1, !!p, i );
289 nz = h->quantf.quant_4x4( dct4x4[i], h->quant4_mf[i_quant_cat][i_qp], h->quant4_bias[i_quant_cat][i_qp] );
290 h->mb.cache.non_zero_count[x264_scan8[16*p+i]] = nz;
293 h->zigzagf.scan_4x4( h->dct.luma4x4[16*p+i], dct4x4[i] );
294 h->quantf.dequant_4x4( dct4x4[i], h->dequant4_mf[i_quant_cat], i_qp );
295 if( decimate_score < 6 ) decimate_score += h->quantf.decimate_score15( h->dct.luma4x4[16*p+i] );
300 /* Writing the 16 CBFs in an i16x16 block is quite costly, so decimation can save many bits. */
301 /* More useful with CAVLC, but still useful with CABAC. */
302 if( decimate_score < 6 )
304 CLEAR_16x16_NNZ( p );
308 h->mb.i_cbp_luma |= block_cbp;
310 h->dctf.dct4x4dc( dct_dc4x4 );
311 if( h->mb.b_trellis )
312 nz = x264_quant_luma_dc_trellis( h, dct_dc4x4, i_quant_cat, i_qp, ctx_cat_plane[DCT_LUMA_DC][p], 1, LUMA_DC+p );
314 nz = h->quantf.quant_4x4_dc( dct_dc4x4, h->quant4_mf[i_quant_cat][i_qp][0]>>1, h->quant4_bias[i_quant_cat][i_qp][0]<<1 );
316 h->mb.cache.non_zero_count[x264_scan8[LUMA_DC+p]] = nz;
319 h->zigzagf.scan_4x4( h->dct.luma16x16_dc[p], dct_dc4x4 );
321 /* output samples to fdec */
322 h->dctf.idct4x4dc( dct_dc4x4 );
323 h->quantf.dequant_4x4_dc( dct_dc4x4, h->dequant4_mf[i_quant_cat], i_qp ); /* XXX not inversed */
325 for( int i = 0; i < 16; i++ )
326 dct4x4[i][0] = dct_dc4x4[block_idx_xy_1d[i]];
329 /* put pixels to fdec */
331 h->dctf.add16x16_idct( p_dst, dct4x4 );
333 h->dctf.add16x16_idct_dc( p_dst, dct_dc4x4 );
336 /* Round down coefficients losslessly in DC-only chroma blocks.
337 * Unlike luma blocks, this can't be done with a lookup table or
338 * other shortcut technique because of the interdependencies
339 * between the coefficients due to the chroma DC transform. */
340 static ALWAYS_INLINE int x264_mb_optimize_chroma_dc( x264_t *h, dctcoef *dct_dc, int dequant_mf[6][16], int i_qp, int chroma422 )
342 int dmf = dequant_mf[i_qp%6][0] << i_qp/6;
344 /* If the QP is too high, there's no benefit to rounding optimization. */
349 return h->quantf.optimize_chroma_2x4_dc( dct_dc, dmf );
351 return h->quantf.optimize_chroma_2x2_dc( dct_dc, dmf );
354 static ALWAYS_INLINE void x264_mb_encode_chroma_internal( x264_t *h, int b_inter, int i_qp, int chroma422 )
357 int b_decimate = b_inter && h->mb.b_dct_decimate;
358 int (*dequant_mf)[16] = h->dequant4_mf[CQM_4IC + b_inter];
359 ALIGNED_ARRAY_16( dctcoef, dct_dc,[8] );
360 h->mb.i_cbp_chroma = 0;
361 h->nr_count[2] += h->mb.b_noise_reduction * 4;
363 /* Early termination: check variance of chroma residual before encoding.
364 * Don't bother trying early termination at low QPs.
365 * Values are experimentally derived. */
366 if( b_decimate && i_qp >= (h->mb.b_trellis ? 12 : 18) && !h->mb.b_noise_reduction )
368 int thresh = chroma422 ? (x264_lambda2_tab[i_qp] + 16) >> 5 : (x264_lambda2_tab[i_qp] + 32) >> 6;
370 int chromapix = chroma422 ? PIXEL_8x16 : PIXEL_8x8;
372 int score = h->pixf.var2[chromapix]( h->mb.pic.p_fenc[1], FENC_STRIDE, h->mb.pic.p_fdec[1], FDEC_STRIDE, &ssd[0] );
373 if( score < thresh*4 )
374 score += h->pixf.var2[chromapix]( h->mb.pic.p_fenc[2], FENC_STRIDE, h->mb.pic.p_fdec[2], FDEC_STRIDE, &ssd[1] );
375 if( score < thresh*4 )
377 M16( &h->mb.cache.non_zero_count[x264_scan8[16]] ) = 0;
378 M16( &h->mb.cache.non_zero_count[x264_scan8[18]] ) = 0;
379 M16( &h->mb.cache.non_zero_count[x264_scan8[32]] ) = 0;
380 M16( &h->mb.cache.non_zero_count[x264_scan8[34]] ) = 0;
383 M16( &h->mb.cache.non_zero_count[x264_scan8[24]] ) = 0;
384 M16( &h->mb.cache.non_zero_count[x264_scan8[26]] ) = 0;
385 M16( &h->mb.cache.non_zero_count[x264_scan8[40]] ) = 0;
386 M16( &h->mb.cache.non_zero_count[x264_scan8[42]] ) = 0;
388 h->mb.cache.non_zero_count[x264_scan8[CHROMA_DC+0]] = 0;
389 h->mb.cache.non_zero_count[x264_scan8[CHROMA_DC+1]] = 0;
391 for( int ch = 0; ch < 2; ch++ )
393 if( ssd[ch] > thresh )
395 pixel *p_src = h->mb.pic.p_fenc[1+ch];
396 pixel *p_dst = h->mb.pic.p_fdec[1+ch];
399 /* Cannot be replaced by two calls to sub8x8_dct_dc since the hadamard transform is different */
400 h->dctf.sub8x16_dct_dc( dct_dc, p_src, p_dst );
402 h->dctf.sub8x8_dct_dc( dct_dc, p_src, p_dst );
404 if( h->mb.b_trellis )
405 nz_dc = x264_quant_chroma_dc_trellis( h, dct_dc, i_qp+3*chroma422, !b_inter, CHROMA_DC+ch );
409 for( int i = 0; i <= chroma422; i++ )
410 nz_dc |= h->quantf.quant_2x2_dc( &dct_dc[4*i], h->quant4_mf[CQM_4IC+b_inter][i_qp+3*chroma422][0] >> 1,
411 h->quant4_bias[CQM_4IC+b_inter][i_qp+3*chroma422][0] << 1 );
416 if( !x264_mb_optimize_chroma_dc( h, dct_dc, dequant_mf, i_qp+3*chroma422, chroma422 ) )
418 h->mb.cache.non_zero_count[x264_scan8[CHROMA_DC+ch]] = 1;
421 zigzag_scan_2x4_dc( h->dct.chroma_dc[ch], dct_dc );
422 h->quantf.idct_dequant_2x4_dconly( dct_dc, dequant_mf, i_qp+3 );
426 zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct_dc );
427 idct_dequant_2x2_dconly( dct_dc, dequant_mf, i_qp );
430 for( int i = 0; i <= chroma422; i++ )
431 h->dctf.add8x8_idct_dc( p_dst + 8*i*FDEC_STRIDE, &dct_dc[4*i] );
432 h->mb.i_cbp_chroma = 1;
440 for( int ch = 0; ch < 2; ch++ )
442 pixel *p_src = h->mb.pic.p_fenc[1+ch];
443 pixel *p_dst = h->mb.pic.p_fdec[1+ch];
444 int i_decimate_score = 0;
447 ALIGNED_ARRAY_16( dctcoef, dct4x4,[8],[16] );
449 if( h->mb.b_lossless )
451 static const uint8_t chroma422_scan[8] = { 0, 2, 1, 5, 3, 6, 4, 7 };
453 for( int i = 0; i < (chroma422?8:4); i++ )
455 int oe = 4*(i&1) + 4*(i>>1)*FENC_STRIDE;
456 int od = 4*(i&1) + 4*(i>>1)*FDEC_STRIDE;
457 nz = h->zigzagf.sub_4x4ac( h->dct.luma4x4[16+i+(chroma422?i&4:0)+ch*16], p_src+oe, p_dst+od,
458 &h->dct.chroma_dc[ch][chroma422?chroma422_scan[i]:i] );
459 h->mb.cache.non_zero_count[x264_scan8[16+i+(chroma422?i&4:0)+ch*16]] = nz;
460 h->mb.i_cbp_chroma |= nz;
462 h->mb.cache.non_zero_count[x264_scan8[CHROMA_DC+ch]] = array_non_zero( h->dct.chroma_dc[ch], chroma422?8:4 );
466 for( int i = 0; i <= chroma422; i++ )
467 h->dctf.sub8x8_dct( &dct4x4[4*i], p_src + 8*i*FENC_STRIDE, p_dst + 8*i*FDEC_STRIDE );
469 if( h->mb.b_noise_reduction )
470 for( int i = 0; i < (chroma422?8:4); i++ )
471 h->quantf.denoise_dct( dct4x4[i], h->nr_residual_sum[2], h->nr_offset[2], 16 );
474 h->dctf.dct2x4dc( dct_dc, dct4x4 );
476 dct2x2dc( dct_dc, dct4x4 );
478 /* calculate dct coeffs */
479 for( int i = 0; i < (chroma422?8:4); i++ )
481 if( h->mb.b_trellis )
482 nz = x264_quant_4x4_trellis( h, dct4x4[i], CQM_4IC+b_inter, i_qp, DCT_CHROMA_AC, !b_inter, 1, 0 );
484 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] );
485 h->mb.cache.non_zero_count[x264_scan8[16+i+(chroma422?i&4:0)+ch*16]] = nz;
489 h->zigzagf.scan_4x4( h->dct.luma4x4[16+i+(chroma422?i&4:0)+ch*16], dct4x4[i] );
490 h->quantf.dequant_4x4( dct4x4[i], dequant_mf, i_qp );
492 i_decimate_score += h->quantf.decimate_score15( h->dct.luma4x4[16+i+(chroma422?i&4:0)+ch*16] );
496 if( h->mb.b_trellis )
497 nz_dc = x264_quant_chroma_dc_trellis( h, dct_dc, i_qp+3*chroma422, !b_inter, CHROMA_DC+ch );
501 for( int i = 0; i <= chroma422; i++ )
502 nz_dc |= h->quantf.quant_2x2_dc( &dct_dc[4*i], h->quant4_mf[CQM_4IC+b_inter][i_qp+3*chroma422][0] >> 1,
503 h->quant4_bias[CQM_4IC+b_inter][i_qp+3*chroma422][0] << 1 );
506 h->mb.cache.non_zero_count[x264_scan8[CHROMA_DC+ch]] = nz_dc;
508 if( (b_decimate && i_decimate_score < 7) || !nz_ac )
510 /* Decimate the block */
511 M16( &h->mb.cache.non_zero_count[x264_scan8[16+16*ch]] ) = 0;
512 M16( &h->mb.cache.non_zero_count[x264_scan8[18+16*ch]] ) = 0;
515 M16( &h->mb.cache.non_zero_count[x264_scan8[24+16*ch]] ) = 0;
516 M16( &h->mb.cache.non_zero_count[x264_scan8[26+16*ch]] ) = 0;
519 if( !nz_dc ) /* Whole block is empty */
521 if( !x264_mb_optimize_chroma_dc( h, dct_dc, dequant_mf, i_qp+3*chroma422, chroma422 ) )
523 h->mb.cache.non_zero_count[x264_scan8[CHROMA_DC+ch]] = 0;
529 zigzag_scan_2x4_dc( h->dct.chroma_dc[ch], dct_dc );
530 h->quantf.idct_dequant_2x4_dconly( dct_dc, dequant_mf, i_qp+3 );
534 zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct_dc );
535 idct_dequant_2x2_dconly( dct_dc, dequant_mf, i_qp );
538 for( int i = 0; i <= chroma422; i++ )
539 h->dctf.add8x8_idct_dc( p_dst + 8*i*FDEC_STRIDE, &dct_dc[4*i] );
543 h->mb.i_cbp_chroma = 1;
549 zigzag_scan_2x4_dc( h->dct.chroma_dc[ch], dct_dc );
550 h->quantf.idct_dequant_2x4_dc( dct_dc, dct4x4, dequant_mf, i_qp+3 );
554 zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct_dc );
555 idct_dequant_2x2_dc( dct_dc, dct4x4, dequant_mf, i_qp );
559 for( int i = 0; i <= chroma422; i++ )
560 h->dctf.add8x8_idct( p_dst + 8*i*FDEC_STRIDE, &dct4x4[4*i] );
564 /* 0 = none, 1 = DC only, 2 = DC+AC */
565 h->mb.i_cbp_chroma += (h->mb.cache.non_zero_count[x264_scan8[CHROMA_DC+0]] |
566 h->mb.cache.non_zero_count[x264_scan8[CHROMA_DC+1]] | h->mb.i_cbp_chroma);
569 void x264_mb_encode_chroma( x264_t *h, int b_inter, int i_qp )
571 if( CHROMA_FORMAT == CHROMA_420 )
572 x264_mb_encode_chroma_internal( h, b_inter, i_qp, 0 );
574 x264_mb_encode_chroma_internal( h, b_inter, i_qp, 1 );
577 static void x264_macroblock_encode_skip( x264_t *h )
579 M32( &h->mb.cache.non_zero_count[x264_scan8[ 0]] ) = 0;
580 M32( &h->mb.cache.non_zero_count[x264_scan8[ 2]] ) = 0;
581 M32( &h->mb.cache.non_zero_count[x264_scan8[ 8]] ) = 0;
582 M32( &h->mb.cache.non_zero_count[x264_scan8[10]] ) = 0;
583 M32( &h->mb.cache.non_zero_count[x264_scan8[16+ 0]] ) = 0;
584 M32( &h->mb.cache.non_zero_count[x264_scan8[16+ 2]] ) = 0;
585 M32( &h->mb.cache.non_zero_count[x264_scan8[32+ 0]] ) = 0;
586 M32( &h->mb.cache.non_zero_count[x264_scan8[32+ 2]] ) = 0;
587 if( CHROMA_FORMAT >= CHROMA_422 )
589 M32( &h->mb.cache.non_zero_count[x264_scan8[16+ 8]] ) = 0;
590 M32( &h->mb.cache.non_zero_count[x264_scan8[16+10]] ) = 0;
591 M32( &h->mb.cache.non_zero_count[x264_scan8[32+ 8]] ) = 0;
592 M32( &h->mb.cache.non_zero_count[x264_scan8[32+10]] ) = 0;
594 h->mb.i_cbp_luma = 0;
595 h->mb.i_cbp_chroma = 0;
596 h->mb.cbp[h->mb.i_mb_xy] = 0;
599 /*****************************************************************************
600 * Intra prediction for predictive lossless mode.
601 *****************************************************************************/
603 void x264_predict_lossless_chroma( x264_t *h, int i_mode )
605 int height = 16 >> h->mb.chroma_v_shift;
606 if( i_mode == I_PRED_CHROMA_V )
608 h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fenc[1]-FENC_STRIDE, FENC_STRIDE, height );
609 h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fenc[2]-FENC_STRIDE, FENC_STRIDE, height );
610 memcpy( h->mb.pic.p_fdec[1], h->mb.pic.p_fdec[1]-FDEC_STRIDE, 8*sizeof(pixel) );
611 memcpy( h->mb.pic.p_fdec[2], h->mb.pic.p_fdec[2]-FDEC_STRIDE, 8*sizeof(pixel) );
613 else if( i_mode == I_PRED_CHROMA_H )
615 h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fenc[1]-1, FENC_STRIDE, height );
616 h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fenc[2]-1, FENC_STRIDE, height );
617 x264_copy_column8( h->mb.pic.p_fdec[1]+4*FDEC_STRIDE, h->mb.pic.p_fdec[1]+4*FDEC_STRIDE-1 );
618 x264_copy_column8( h->mb.pic.p_fdec[2]+4*FDEC_STRIDE, h->mb.pic.p_fdec[2]+4*FDEC_STRIDE-1 );
619 if( CHROMA_FORMAT == CHROMA_422 )
621 x264_copy_column8( h->mb.pic.p_fdec[1]+12*FDEC_STRIDE, h->mb.pic.p_fdec[1]+12*FDEC_STRIDE-1 );
622 x264_copy_column8( h->mb.pic.p_fdec[2]+12*FDEC_STRIDE, h->mb.pic.p_fdec[2]+12*FDEC_STRIDE-1 );
627 h->predict_chroma[i_mode]( h->mb.pic.p_fdec[1] );
628 h->predict_chroma[i_mode]( h->mb.pic.p_fdec[2] );
632 void x264_predict_lossless_4x4( x264_t *h, pixel *p_dst, int p, int idx, int i_mode )
634 int stride = h->fenc->i_stride[p] << MB_INTERLACED;
635 pixel *p_src = h->mb.pic.p_fenc_plane[p] + block_idx_x[idx]*4 + block_idx_y[idx]*4 * stride;
637 if( i_mode == I_PRED_4x4_V )
638 h->mc.copy[PIXEL_4x4]( p_dst, FDEC_STRIDE, p_src-stride, stride, 4 );
639 else if( i_mode == I_PRED_4x4_H )
640 h->mc.copy[PIXEL_4x4]( p_dst, FDEC_STRIDE, p_src-1, stride, 4 );
642 h->predict_4x4[i_mode]( p_dst );
645 void x264_predict_lossless_8x8( x264_t *h, pixel *p_dst, int p, int idx, int i_mode, pixel edge[36] )
647 int stride = h->fenc->i_stride[p] << MB_INTERLACED;
648 pixel *p_src = h->mb.pic.p_fenc_plane[p] + (idx&1)*8 + (idx>>1)*8*stride;
650 if( i_mode == I_PRED_8x8_V )
651 h->mc.copy[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src-stride, stride, 8 );
652 else if( i_mode == I_PRED_8x8_H )
653 h->mc.copy[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src-1, stride, 8 );
655 h->predict_8x8[i_mode]( p_dst, edge );
658 void x264_predict_lossless_16x16( x264_t *h, int p, int i_mode )
660 int stride = h->fenc->i_stride[p] << MB_INTERLACED;
661 if( i_mode == I_PRED_16x16_V )
662 h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[p], FDEC_STRIDE, h->mb.pic.p_fenc_plane[p]-stride, stride, 16 );
663 else if( i_mode == I_PRED_16x16_H )
664 h->mc.copy_16x16_unaligned( h->mb.pic.p_fdec[p], FDEC_STRIDE, h->mb.pic.p_fenc_plane[p]-1, stride, 16 );
666 h->predict_16x16[i_mode]( h->mb.pic.p_fdec[p] );
669 /*****************************************************************************
670 * x264_macroblock_encode:
671 *****************************************************************************/
672 static ALWAYS_INLINE void x264_macroblock_encode_internal( x264_t *h, int plane_count, int chroma )
674 int i_qp = h->mb.i_qp;
675 int b_decimate = h->mb.b_dct_decimate;
676 int b_force_no_skip = 0;
678 h->mb.i_cbp_luma = 0;
679 for( int p = 0; p < plane_count; p++ )
680 h->mb.cache.non_zero_count[x264_scan8[LUMA_DC+p]] = 0;
682 if( h->mb.i_type == I_PCM )
684 /* if PCM is chosen, we need to store reconstructed frame data */
685 for( int p = 0; p < plane_count; p++ )
686 h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[p], FDEC_STRIDE, h->mb.pic.p_fenc[p], FENC_STRIDE, 16 );
689 int height = 16 >> h->mb.chroma_v_shift;
690 h->mc.copy[PIXEL_8x8] ( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fenc[1], FENC_STRIDE, height );
691 h->mc.copy[PIXEL_8x8] ( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fenc[2], FENC_STRIDE, height );
696 if( !h->mb.b_allow_skip )
699 if( IS_SKIP(h->mb.i_type) )
701 if( h->mb.i_type == P_SKIP )
703 else if( h->mb.i_type == B_SKIP )
704 h->mb.i_type = B_DIRECT;
708 if( h->mb.i_type == P_SKIP )
710 /* don't do pskip motion compensation if it was already done in macroblock_analyse */
711 if( !h->mb.b_skip_mc )
713 int mvx = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][0],
714 h->mb.mv_min[0], h->mb.mv_max[0] );
715 int mvy = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][1],
716 h->mb.mv_min[1], h->mb.mv_max[1] );
718 for( int p = 0; p < plane_count; p++ )
719 h->mc.mc_luma( h->mb.pic.p_fdec[p], FDEC_STRIDE,
720 &h->mb.pic.p_fref[0][0][p*4], h->mb.pic.i_stride[p],
721 mvx, mvy, 16, 16, &h->sh.weight[0][p] );
725 int v_shift = h->mb.chroma_v_shift;
726 int height = 16 >> v_shift;
728 /* Special case for mv0, which is (of course) very common in P-skip mode. */
730 h->mc.mc_chroma( h->mb.pic.p_fdec[1], h->mb.pic.p_fdec[2], FDEC_STRIDE,
731 h->mb.pic.p_fref[0][0][4], h->mb.pic.i_stride[1],
732 mvx, 2*mvy>>v_shift, 8, height );
734 h->mc.load_deinterleave_chroma_fdec( h->mb.pic.p_fdec[1], h->mb.pic.p_fref[0][0][4],
735 h->mb.pic.i_stride[1], height );
737 if( h->sh.weight[0][1].weightfn )
738 h->sh.weight[0][1].weightfn[8>>2]( h->mb.pic.p_fdec[1], FDEC_STRIDE,
739 h->mb.pic.p_fdec[1], FDEC_STRIDE,
740 &h->sh.weight[0][1], height );
741 if( h->sh.weight[0][2].weightfn )
742 h->sh.weight[0][2].weightfn[8>>2]( h->mb.pic.p_fdec[2], FDEC_STRIDE,
743 h->mb.pic.p_fdec[2], FDEC_STRIDE,
744 &h->sh.weight[0][2], height );
748 x264_macroblock_encode_skip( h );
751 if( h->mb.i_type == B_SKIP )
753 /* don't do bskip motion compensation if it was already done in macroblock_analyse */
754 if( !h->mb.b_skip_mc )
756 x264_macroblock_encode_skip( h );
760 if( h->mb.i_type == I_16x16 )
762 h->mb.b_transform_8x8 = 0;
764 for( int p = 0; p < plane_count; p++ )
766 x264_mb_encode_i16x16( h, p, i_qp );
767 i_qp = h->mb.i_chroma_qp;
770 else if( h->mb.i_type == I_8x8 )
772 h->mb.b_transform_8x8 = 1;
773 /* If we already encoded 3 of the 4 i8x8 blocks, we don't have to do them again. */
774 if( h->mb.i_skip_intra )
776 h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i8x8_fdec_buf, 16, 16 );
777 M32( &h->mb.cache.non_zero_count[x264_scan8[ 0]] ) = h->mb.pic.i8x8_nnz_buf[0];
778 M32( &h->mb.cache.non_zero_count[x264_scan8[ 2]] ) = h->mb.pic.i8x8_nnz_buf[1];
779 M32( &h->mb.cache.non_zero_count[x264_scan8[ 8]] ) = h->mb.pic.i8x8_nnz_buf[2];
780 M32( &h->mb.cache.non_zero_count[x264_scan8[10]] ) = h->mb.pic.i8x8_nnz_buf[3];
781 h->mb.i_cbp_luma = h->mb.pic.i8x8_cbp;
782 /* In RD mode, restore the now-overwritten DCT data. */
783 if( h->mb.i_skip_intra == 2 )
784 h->mc.memcpy_aligned( h->dct.luma8x8, h->mb.pic.i8x8_dct_buf, sizeof(h->mb.pic.i8x8_dct_buf) );
786 for( int p = 0; p < plane_count; p++ )
788 for( int i = (p == 0 && h->mb.i_skip_intra) ? 3 : 0 ; i < 4; i++ )
790 int i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[4*i]];
791 x264_mb_encode_i8x8( h, p, i, i_qp, i_mode, NULL );
793 i_qp = h->mb.i_chroma_qp;
796 else if( h->mb.i_type == I_4x4 )
798 h->mb.b_transform_8x8 = 0;
799 /* If we already encoded 15 of the 16 i4x4 blocks, we don't have to do them again. */
800 if( h->mb.i_skip_intra )
802 h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i4x4_fdec_buf, 16, 16 );
803 M32( &h->mb.cache.non_zero_count[x264_scan8[ 0]] ) = h->mb.pic.i4x4_nnz_buf[0];
804 M32( &h->mb.cache.non_zero_count[x264_scan8[ 2]] ) = h->mb.pic.i4x4_nnz_buf[1];
805 M32( &h->mb.cache.non_zero_count[x264_scan8[ 8]] ) = h->mb.pic.i4x4_nnz_buf[2];
806 M32( &h->mb.cache.non_zero_count[x264_scan8[10]] ) = h->mb.pic.i4x4_nnz_buf[3];
807 h->mb.i_cbp_luma = h->mb.pic.i4x4_cbp;
808 /* In RD mode, restore the now-overwritten DCT data. */
809 if( h->mb.i_skip_intra == 2 )
810 h->mc.memcpy_aligned( h->dct.luma4x4, h->mb.pic.i4x4_dct_buf, sizeof(h->mb.pic.i4x4_dct_buf) );
812 for( int p = 0; p < plane_count; p++ )
814 for( int i = (p == 0 && h->mb.i_skip_intra) ? 15 : 0 ; i < 16; i++ )
816 pixel *p_dst = &h->mb.pic.p_fdec[p][block_idx_xy_fdec[i]];
817 int i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[i]];
819 if( (h->mb.i_neighbour4[i] & (MB_TOPRIGHT|MB_TOP)) == MB_TOP )
820 /* emulate missing topright samples */
821 MPIXEL_X4( &p_dst[4-FDEC_STRIDE] ) = PIXEL_SPLAT_X4( p_dst[3-FDEC_STRIDE] );
823 x264_mb_encode_i4x4( h, p, i, i_qp, i_mode );
825 i_qp = h->mb.i_chroma_qp;
830 int i_decimate_mb = 0;
832 /* Don't repeat motion compensation if it was already done in non-RD transform analysis */
833 if( !h->mb.b_skip_mc )
836 if( h->mb.b_lossless )
838 if( h->mb.b_transform_8x8 )
839 for( int p = 0; p < plane_count; p++ )
840 for( int i8x8 = 0; i8x8 < 4; i8x8++ )
844 nz = h->zigzagf.sub_8x8( h->dct.luma8x8[p*4+i8x8], h->mb.pic.p_fenc[p] + 8*x + 8*y*FENC_STRIDE,
845 h->mb.pic.p_fdec[p] + 8*x + 8*y*FDEC_STRIDE );
846 STORE_8x8_NNZ( p, i8x8, nz );
847 h->mb.i_cbp_luma |= nz << i8x8;
850 for( int p = 0; p < plane_count; p++ )
851 for( int i4x4 = 0; i4x4 < 16; i4x4++ )
853 nz = h->zigzagf.sub_4x4( h->dct.luma4x4[p*16+i4x4],
854 h->mb.pic.p_fenc[p]+block_idx_xy_fenc[i4x4],
855 h->mb.pic.p_fdec[p]+block_idx_xy_fdec[i4x4] );
856 h->mb.cache.non_zero_count[x264_scan8[p*16+i4x4]] = nz;
857 h->mb.i_cbp_luma |= nz << (i4x4>>2);
860 else if( h->mb.b_transform_8x8 )
862 ALIGNED_ARRAY_16( dctcoef, dct8x8,[4],[64] );
863 b_decimate &= !h->mb.b_trellis || !h->param.b_cabac; // 8x8 trellis is inherently optimal decimation for CABAC
865 for( int p = 0; p < plane_count; p++ )
867 h->dctf.sub16x16_dct8( dct8x8, h->mb.pic.p_fenc[p], h->mb.pic.p_fdec[p] );
868 h->nr_count[1+!!p*2] += h->mb.b_noise_reduction * 4;
871 for( int idx = 0; idx < 4; idx++ )
873 nz = x264_quant_8x8( h, dct8x8[idx], i_qp, ctx_cat_plane[DCT_LUMA_8x8][p], 0, p, idx );
877 h->zigzagf.scan_8x8( h->dct.luma8x8[p*4+idx], dct8x8[idx] );
880 int i_decimate_8x8 = h->quantf.decimate_score64( h->dct.luma8x8[p*4+idx] );
881 i_decimate_mb += i_decimate_8x8;
882 if( i_decimate_8x8 >= 4 )
890 if( i_decimate_mb < 6 && b_decimate )
893 CLEAR_16x16_NNZ( p );
897 for( int idx = 0; idx < 4; idx++ )
902 if( plane_cbp&(1<<idx) )
904 h->quantf.dequant_8x8( dct8x8[idx], h->dequant8_mf[p?CQM_8PC:CQM_8PY], i_qp );
905 h->dctf.add8x8_idct8( &h->mb.pic.p_fdec[p][8*x + 8*y*FDEC_STRIDE], dct8x8[idx] );
906 STORE_8x8_NNZ( p, idx, 1 );
909 STORE_8x8_NNZ( p, idx, 0 );
912 h->mb.i_cbp_luma |= plane_cbp;
913 i_qp = h->mb.i_chroma_qp;
918 ALIGNED_ARRAY_16( dctcoef, dct4x4,[16],[16] );
919 for( int p = 0; p < plane_count; p++ )
921 h->dctf.sub16x16_dct( dct4x4, h->mb.pic.p_fenc[p], h->mb.pic.p_fdec[p] );
922 h->nr_count[0+!!p*2] += h->mb.b_noise_reduction * 16;
925 for( int i8x8 = 0; i8x8 < 4; i8x8++ )
927 int i_decimate_8x8 = 0;
930 /* encode one 4x4 block */
931 for( int i4x4 = 0; i4x4 < 4; i4x4++ )
933 int idx = i8x8 * 4 + i4x4;
935 nz = x264_quant_4x4( h, dct4x4[idx], i_qp, ctx_cat_plane[DCT_LUMA_4x4][p], 0, p, idx );
936 h->mb.cache.non_zero_count[x264_scan8[p*16+idx]] = nz;
940 h->zigzagf.scan_4x4( h->dct.luma4x4[p*16+idx], dct4x4[idx] );
941 h->quantf.dequant_4x4( dct4x4[idx], h->dequant4_mf[p?CQM_4PC:CQM_4PY], i_qp );
942 if( b_decimate && i_decimate_8x8 < 6 )
943 i_decimate_8x8 += h->quantf.decimate_score16( h->dct.luma4x4[p*16+idx] );
951 /* decimate this 8x8 block */
952 i_decimate_mb += i_decimate_8x8;
955 if( i_decimate_8x8 < 4 )
956 STORE_8x8_NNZ( p, i8x8, 0 );
958 plane_cbp |= 1<<i8x8;
962 h->dctf.add8x8_idct( &h->mb.pic.p_fdec[p][8*x + 8*y*FDEC_STRIDE], &dct4x4[i8x8*4] );
963 plane_cbp |= 1<<i8x8;
969 if( i_decimate_mb < 6 )
972 CLEAR_16x16_NNZ( p );
976 for( int i8x8 = 0; i8x8 < 4; i8x8++ )
977 if( plane_cbp&(1<<i8x8) )
978 h->dctf.add8x8_idct( &h->mb.pic.p_fdec[p][(i8x8&1)*8 + (i8x8>>1)*8*FDEC_STRIDE], &dct4x4[i8x8*4] );
981 h->mb.i_cbp_luma |= plane_cbp;
982 i_qp = h->mb.i_chroma_qp;
990 if( IS_INTRA( h->mb.i_type ) )
992 int i_mode = h->mb.i_chroma_pred_mode;
993 if( h->mb.b_lossless )
994 x264_predict_lossless_chroma( h, i_mode );
997 h->predict_chroma[i_mode]( h->mb.pic.p_fdec[1] );
998 h->predict_chroma[i_mode]( h->mb.pic.p_fdec[2] );
1002 /* encode the 8x8 blocks */
1003 x264_mb_encode_chroma( h, !IS_INTRA( h->mb.i_type ), h->mb.i_chroma_qp );
1006 h->mb.i_cbp_chroma = 0;
1009 int cbp = h->mb.i_cbp_chroma << 4 | h->mb.i_cbp_luma;
1010 if( h->param.b_cabac )
1011 cbp |= h->mb.cache.non_zero_count[x264_scan8[LUMA_DC ]] << 8
1012 | h->mb.cache.non_zero_count[x264_scan8[CHROMA_DC+0]] << 9
1013 | h->mb.cache.non_zero_count[x264_scan8[CHROMA_DC+1]] << 10;
1014 h->mb.cbp[h->mb.i_mb_xy] = cbp;
1017 * XXX: in the me perhaps we should take x264_mb_predict_mv_pskip into account
1018 * (if multiple mv give same result)*/
1019 if( !b_force_no_skip )
1021 if( h->mb.i_type == P_L0 && h->mb.i_partition == D_16x16 &&
1022 !(h->mb.i_cbp_luma | h->mb.i_cbp_chroma) &&
1023 M32( h->mb.cache.mv[0][x264_scan8[0]] ) == M32( h->mb.cache.pskip_mv )
1024 && h->mb.cache.ref[0][x264_scan8[0]] == 0 )
1026 h->mb.i_type = P_SKIP;
1029 /* Check for B_SKIP */
1030 if( h->mb.i_type == B_DIRECT && !(h->mb.i_cbp_luma | h->mb.i_cbp_chroma) )
1032 h->mb.i_type = B_SKIP;
1037 void x264_macroblock_encode( x264_t *h )
1040 x264_macroblock_encode_internal( h, 3, 0 );
1042 x264_macroblock_encode_internal( h, 1, 1 );
1045 /*****************************************************************************
1046 * x264_macroblock_probe_skip:
1047 * Check if the current MB could be encoded as a [PB]_SKIP
1048 *****************************************************************************/
1049 static ALWAYS_INLINE int x264_macroblock_probe_skip_internal( x264_t *h, int b_bidir, int plane_count, int chroma )
1051 ALIGNED_ARRAY_16( dctcoef, dct4x4,[8],[16] );
1052 ALIGNED_ARRAY_16( dctcoef, dctscan,[16] );
1053 ALIGNED_4( int16_t mvp[2] );
1054 int i_qp = h->mb.i_qp;
1056 for( int p = 0; p < plane_count; p++ )
1058 int quant_cat = p ? CQM_4PC : CQM_4PY;
1062 mvp[0] = x264_clip3( h->mb.cache.pskip_mv[0], h->mb.mv_min[0], h->mb.mv_max[0] );
1063 mvp[1] = x264_clip3( h->mb.cache.pskip_mv[1], h->mb.mv_min[1], h->mb.mv_max[1] );
1065 /* Motion compensation */
1066 h->mc.mc_luma( h->mb.pic.p_fdec[p], FDEC_STRIDE,
1067 &h->mb.pic.p_fref[0][0][p*4], h->mb.pic.i_stride[p],
1068 mvp[0], mvp[1], 16, 16, &h->sh.weight[0][p] );
1071 for( int i8x8 = 0, i_decimate_mb = 0; i8x8 < 4; i8x8++ )
1073 int fenc_offset = (i8x8&1) * 8 + (i8x8>>1) * FENC_STRIDE * 8;
1074 int fdec_offset = (i8x8&1) * 8 + (i8x8>>1) * FDEC_STRIDE * 8;
1076 h->dctf.sub8x8_dct( dct4x4, h->mb.pic.p_fenc[p] + fenc_offset,
1077 h->mb.pic.p_fdec[p] + fdec_offset );
1078 /* encode one 4x4 block */
1079 for( int i4x4 = 0; i4x4 < 4; i4x4++ )
1081 if( h->mb.b_noise_reduction )
1082 h->quantf.denoise_dct( dct4x4[i4x4], h->nr_residual_sum[0+!!p*2], h->nr_offset[0+!!p*2], 16 );
1083 if( !h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[quant_cat][i_qp], h->quant4_bias[quant_cat][i_qp] ) )
1085 h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
1086 i_decimate_mb += h->quantf.decimate_score16( dctscan );
1087 if( i_decimate_mb >= 6 )
1091 i_qp = h->mb.i_chroma_qp;
1094 if( chroma == CHROMA_420 || chroma == CHROMA_422 )
1096 i_qp = h->mb.i_chroma_qp;
1097 int chroma422 = chroma == CHROMA_422;
1098 int thresh = chroma422 ? (x264_lambda2_tab[i_qp] + 16) >> 5 : (x264_lambda2_tab[i_qp] + 32) >> 6;
1100 ALIGNED_ARRAY_16( dctcoef, dct_dc,[8] );
1104 /* Special case for mv0, which is (of course) very common in P-skip mode. */
1106 h->mc.mc_chroma( h->mb.pic.p_fdec[1], h->mb.pic.p_fdec[2], FDEC_STRIDE,
1107 h->mb.pic.p_fref[0][0][4], h->mb.pic.i_stride[1],
1108 mvp[0], mvp[1]<<chroma422, 8, chroma422?16:8 );
1110 h->mc.load_deinterleave_chroma_fdec( h->mb.pic.p_fdec[1], h->mb.pic.p_fref[0][0][4],
1111 h->mb.pic.i_stride[1], chroma422?16:8 );
1114 for( int ch = 0; ch < 2; ch++ )
1116 pixel *p_src = h->mb.pic.p_fenc[1+ch];
1117 pixel *p_dst = h->mb.pic.p_fdec[1+ch];
1119 if( !b_bidir && h->sh.weight[0][1+ch].weightfn )
1120 h->sh.weight[0][1+ch].weightfn[8>>2]( h->mb.pic.p_fdec[1+ch], FDEC_STRIDE,
1121 h->mb.pic.p_fdec[1+ch], FDEC_STRIDE,
1122 &h->sh.weight[0][1+ch], chroma422?16:8 );
1124 /* there is almost never a termination during chroma, but we can't avoid the check entirely */
1125 /* so instead we check SSD and skip the actual check if the score is low enough. */
1126 ssd = h->pixf.ssd[chroma422?PIXEL_8x16:PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src, FENC_STRIDE );
1130 /* The vast majority of chroma checks will terminate during the DC check or the higher
1131 * threshold check, so we can save time by doing a DC-only DCT. */
1132 if( h->mb.b_noise_reduction )
1134 for( int i = 0; i <= chroma422; i++ )
1135 h->dctf.sub8x8_dct( &dct4x4[4*i], p_src + 8*i*FENC_STRIDE, p_dst + 8*i*FDEC_STRIDE );
1137 for( int i4x4 = 0; i4x4 < (chroma422?8:4); i4x4++ )
1139 h->quantf.denoise_dct( dct4x4[i4x4], h->nr_residual_sum[2], h->nr_offset[2], 16 );
1140 dct_dc[i4x4] = dct4x4[i4x4][0];
1146 h->dctf.sub8x16_dct_dc( dct_dc, p_src, p_dst );
1148 h->dctf.sub8x8_dct_dc( dct_dc, p_src, p_dst );
1151 for( int i = 0; i <= chroma422; i++ )
1152 if( h->quantf.quant_2x2_dc( &dct_dc[4*i], h->quant4_mf[CQM_4PC][i_qp+3*chroma422][0] >> 1,
1153 h->quant4_bias[CQM_4PC][i_qp+3*chroma422][0] << 1 ) )
1156 /* If there wasn't a termination in DC, we can check against a much higher threshold. */
1157 if( ssd < thresh*4 )
1160 if( !h->mb.b_noise_reduction )
1161 for( int i = 0; i <= chroma422; i++ )
1162 h->dctf.sub8x8_dct( &dct4x4[4*i], p_src + 8*i*FENC_STRIDE, p_dst + 8*i*FDEC_STRIDE );
1164 /* calculate dct coeffs */
1165 for( int i4x4 = 0, i_decimate_mb = 0; i4x4 < (chroma422?8:4); i4x4++ )
1167 dct4x4[i4x4][0] = 0;
1168 if( h->mb.b_noise_reduction )
1169 h->quantf.denoise_dct( dct4x4[i4x4], h->nr_residual_sum[2], h->nr_offset[2], 16 );
1170 if( !h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] ) )
1172 h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
1173 i_decimate_mb += h->quantf.decimate_score15( dctscan );
1174 if( i_decimate_mb >= 7 )
1180 h->mb.b_skip_mc = 1;
1184 int x264_macroblock_probe_skip( x264_t *h, int b_bidir )
1186 if( CHROMA_FORMAT == CHROMA_444 )
1187 return x264_macroblock_probe_skip_internal( h, b_bidir, 3, CHROMA_444 );
1188 else if( CHROMA_FORMAT == CHROMA_422 )
1189 return x264_macroblock_probe_skip_internal( h, b_bidir, 1, CHROMA_422 );
1191 return x264_macroblock_probe_skip_internal( h, b_bidir, 1, CHROMA_420 );
1194 /****************************************************************************
1195 * DCT-domain noise reduction / adaptive deadzone
1197 ****************************************************************************/
1199 void x264_noise_reduction_update( x264_t *h )
1201 h->nr_offset = h->nr_offset_denoise;
1202 h->nr_residual_sum = h->nr_residual_sum_buf[0];
1203 h->nr_count = h->nr_count_buf[0];
1204 for( int cat = 0; cat < 3 + CHROMA444; cat++ )
1207 int size = dct8x8 ? 64 : 16;
1208 const uint16_t *weight = dct8x8 ? x264_dct8_weight2_tab : x264_dct4_weight2_tab;
1210 if( h->nr_count[cat] > (dct8x8 ? (1<<16) : (1<<18)) )
1212 for( int i = 0; i < size; i++ )
1213 h->nr_residual_sum[cat][i] >>= 1;
1214 h->nr_count[cat] >>= 1;
1217 for( int i = 0; i < size; i++ )
1218 h->nr_offset[cat][i] =
1219 ((uint64_t)h->param.analyse.i_noise_reduction * h->nr_count[cat]
1220 + h->nr_residual_sum[cat][i]/2)
1221 / ((uint64_t)h->nr_residual_sum[cat][i] * weight[i]/256 + 1);
1223 /* Don't denoise DC coefficients */
1224 h->nr_offset[cat][0] = 0;
1228 /*****************************************************************************
1229 * RD only; 4 calls to this do not make up for one macroblock_encode.
1230 * doesn't transform chroma dc.
1231 *****************************************************************************/
1232 static ALWAYS_INLINE void x264_macroblock_encode_p8x8_internal( x264_t *h, int i8, int plane_count, int chroma )
1234 int b_decimate = h->mb.b_dct_decimate;
1235 int i_qp = h->mb.i_qp;
1239 int chroma422 = chroma == CHROMA_422;
1241 h->mb.i_cbp_chroma = 0;
1242 h->mb.i_cbp_luma &= ~(1 << i8);
1244 if( !h->mb.b_skip_mc )
1245 x264_mb_mc_8x8( h, i8 );
1247 if( h->mb.b_lossless )
1249 for( int p = 0; p < plane_count; p++ )
1251 pixel *p_fenc = h->mb.pic.p_fenc[p] + 8*x + 8*y*FENC_STRIDE;
1252 pixel *p_fdec = h->mb.pic.p_fdec[p] + 8*x + 8*y*FDEC_STRIDE;
1254 if( h->mb.b_transform_8x8 )
1256 nnz8x8 = h->zigzagf.sub_8x8( h->dct.luma8x8[4*p+i8], p_fenc, p_fdec );
1257 STORE_8x8_NNZ( p, i8, nnz8x8 );
1261 for( int i4 = i8*4; i4 < i8*4+4; i4++ )
1263 nz = h->zigzagf.sub_4x4( h->dct.luma4x4[16*p+i4],
1264 h->mb.pic.p_fenc[p]+block_idx_xy_fenc[i4],
1265 h->mb.pic.p_fdec[p]+block_idx_xy_fdec[i4] );
1266 h->mb.cache.non_zero_count[x264_scan8[16*p+i4]] = nz;
1270 h->mb.i_cbp_luma |= nnz8x8 << i8;
1272 if( chroma == CHROMA_420 || chroma == CHROMA_422 )
1274 for( int ch = 0; ch < 2; ch++ )
1277 pixel *p_fenc = h->mb.pic.p_fenc[1+ch] + 4*x + (chroma422?8:4)*y*FENC_STRIDE;
1278 pixel *p_fdec = h->mb.pic.p_fdec[1+ch] + 4*x + (chroma422?8:4)*y*FDEC_STRIDE;
1280 for( int i4x4 = 0; i4x4 <= chroma422; i4x4++ )
1282 int offset = chroma422 ? 8*y + 2*i4x4 + x : i8;
1283 nz = h->zigzagf.sub_4x4ac( h->dct.luma4x4[16+offset+ch*16], p_fenc+4*i4x4*FENC_STRIDE, p_fdec+4*i4x4*FDEC_STRIDE, &dc );
1284 h->mb.cache.non_zero_count[x264_scan8[16+offset+ch*16]] = nz;
1287 h->mb.i_cbp_chroma = 0x02;
1292 if( h->mb.b_transform_8x8 )
1294 for( int p = 0; p < plane_count; p++ )
1296 int quant_cat = p ? CQM_8PC : CQM_8PY;
1297 pixel *p_fenc = h->mb.pic.p_fenc[p] + 8*x + 8*y*FENC_STRIDE;
1298 pixel *p_fdec = h->mb.pic.p_fdec[p] + 8*x + 8*y*FDEC_STRIDE;
1299 ALIGNED_ARRAY_16( dctcoef, dct8x8,[64] );
1300 h->dctf.sub8x8_dct8( dct8x8, p_fenc, p_fdec );
1301 int nnz8x8 = x264_quant_8x8( h, dct8x8, i_qp, ctx_cat_plane[DCT_LUMA_8x8][p], 0, p, i8 );
1304 h->zigzagf.scan_8x8( h->dct.luma8x8[4*p+i8], dct8x8 );
1306 if( b_decimate && !h->mb.b_trellis )
1307 nnz8x8 = 4 <= h->quantf.decimate_score64( h->dct.luma8x8[4*p+i8] );
1311 h->quantf.dequant_8x8( dct8x8, h->dequant8_mf[quant_cat], i_qp );
1312 h->dctf.add8x8_idct8( p_fdec, dct8x8 );
1313 STORE_8x8_NNZ( p, i8, 1 );
1316 STORE_8x8_NNZ( p, i8, 0 );
1319 STORE_8x8_NNZ( p, i8, 0 );
1320 h->mb.i_cbp_luma |= nnz8x8 << i8;
1321 i_qp = h->mb.i_chroma_qp;
1326 for( int p = 0; p < plane_count; p++ )
1328 int quant_cat = p ? CQM_4PC : CQM_4PY;
1329 pixel *p_fenc = h->mb.pic.p_fenc[p] + 8*x + 8*y*FENC_STRIDE;
1330 pixel *p_fdec = h->mb.pic.p_fdec[p] + 8*x + 8*y*FDEC_STRIDE;
1331 int i_decimate_8x8 = 0, nnz8x8 = 0;
1332 ALIGNED_ARRAY_16( dctcoef, dct4x4,[4],[16] );
1333 h->dctf.sub8x8_dct( dct4x4, p_fenc, p_fdec );
1334 for( int i4 = 0; i4 < 4; i4++ )
1336 nz = x264_quant_4x4( h, dct4x4[i4], i_qp, ctx_cat_plane[DCT_LUMA_4x4][p], 0, p, i8*4+i4 );
1337 h->mb.cache.non_zero_count[x264_scan8[p*16+i8*4+i4]] = nz;
1340 h->zigzagf.scan_4x4( h->dct.luma4x4[p*16+i8*4+i4], dct4x4[i4] );
1341 h->quantf.dequant_4x4( dct4x4[i4], h->dequant4_mf[quant_cat], i_qp );
1343 i_decimate_8x8 += h->quantf.decimate_score16( h->dct.luma4x4[p*16+i8*4+i4] );
1348 if( b_decimate && i_decimate_8x8 < 4 )
1352 h->dctf.add8x8_idct( p_fdec, dct4x4 );
1354 STORE_8x8_NNZ( p, i8, 0 );
1356 h->mb.i_cbp_luma |= nnz8x8 << i8;
1357 i_qp = h->mb.i_chroma_qp;
1361 if( chroma == CHROMA_420 || chroma == CHROMA_422 )
1363 i_qp = h->mb.i_chroma_qp;
1364 for( int ch = 0; ch < 2; ch++ )
1366 ALIGNED_ARRAY_16( dctcoef, dct4x4,[2],[16] );
1367 pixel *p_fenc = h->mb.pic.p_fenc[1+ch] + 4*x + (chroma422?8:4)*y*FENC_STRIDE;
1368 pixel *p_fdec = h->mb.pic.p_fdec[1+ch] + 4*x + (chroma422?8:4)*y*FDEC_STRIDE;
1370 for( int i4x4 = 0; i4x4 <= chroma422; i4x4++ )
1372 h->dctf.sub4x4_dct( dct4x4[i4x4], p_fenc + 4*i4x4*FENC_STRIDE, p_fdec + 4*i4x4*FDEC_STRIDE );
1374 if( h->mb.b_noise_reduction )
1375 h->quantf.denoise_dct( dct4x4[i4x4], h->nr_residual_sum[2], h->nr_offset[2], 16 );
1376 dct4x4[i4x4][0] = 0;
1378 if( h->mb.b_trellis )
1379 nz = x264_quant_4x4_trellis( h, dct4x4[i4x4], CQM_4PC, i_qp, DCT_CHROMA_AC, 0, 1, 0 );
1381 nz = h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] );
1383 int offset = chroma422 ? ((5*i8) & 0x09) + 2*i4x4 : i8;
1384 h->mb.cache.non_zero_count[x264_scan8[16+offset+ch*16]] = nz;
1387 h->zigzagf.scan_4x4( h->dct.luma4x4[16+offset+ch*16], dct4x4[i4x4] );
1388 h->quantf.dequant_4x4( dct4x4[i4x4], h->dequant4_mf[CQM_4PC], i_qp );
1389 h->dctf.add4x4_idct( p_fdec + 4*i4x4*FDEC_STRIDE, dct4x4[i4x4] );
1393 h->mb.i_cbp_chroma = 0x02;
1398 void x264_macroblock_encode_p8x8( x264_t *h, int i8 )
1401 x264_macroblock_encode_p8x8_internal( h, i8, 3, CHROMA_444 );
1402 else if( CHROMA_FORMAT == CHROMA_422 )
1403 x264_macroblock_encode_p8x8_internal( h, i8, 1, CHROMA_422 );
1405 x264_macroblock_encode_p8x8_internal( h, i8, 1, CHROMA_420 );
1408 /*****************************************************************************
1409 * RD only, luma only (for 4:2:0)
1410 *****************************************************************************/
1411 static ALWAYS_INLINE void x264_macroblock_encode_p4x4_internal( x264_t *h, int i4, int plane_count )
1413 int i_qp = h->mb.i_qp;
1415 for( int p = 0; p < plane_count; p++ )
1417 int quant_cat = p ? CQM_4PC : CQM_4PY;
1418 pixel *p_fenc = &h->mb.pic.p_fenc[p][block_idx_xy_fenc[i4]];
1419 pixel *p_fdec = &h->mb.pic.p_fdec[p][block_idx_xy_fdec[i4]];
1422 /* Don't need motion compensation as this function is only used in qpel-RD, which caches pixel data. */
1424 if( h->mb.b_lossless )
1426 nz = h->zigzagf.sub_4x4( h->dct.luma4x4[p*16+i4], p_fenc, p_fdec );
1427 h->mb.cache.non_zero_count[x264_scan8[p*16+i4]] = nz;
1431 ALIGNED_ARRAY_16( dctcoef, dct4x4,[16] );
1432 h->dctf.sub4x4_dct( dct4x4, p_fenc, p_fdec );
1433 nz = x264_quant_4x4( h, dct4x4, i_qp, ctx_cat_plane[DCT_LUMA_4x4][p], 0, p, i4 );
1434 h->mb.cache.non_zero_count[x264_scan8[p*16+i4]] = nz;
1437 h->zigzagf.scan_4x4( h->dct.luma4x4[p*16+i4], dct4x4 );
1438 h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[quant_cat], i_qp );
1439 h->dctf.add4x4_idct( p_fdec, dct4x4 );
1442 i_qp = h->mb.i_chroma_qp;
1446 void x264_macroblock_encode_p4x4( x264_t *h, int i8 )
1449 x264_macroblock_encode_p4x4_internal( h, i8, 3 );
1451 x264_macroblock_encode_p4x4_internal( h, i8, 1 );