1 /*****************************************************************************
2 * slicetype.c: lookahead analysis
3 *****************************************************************************
4 * Copyright (C) 2005-2010 x264 project
6 * Authors: Fiona Glaser <fiona@x264.com>
7 * Loren Merritt <lorenm@u.washington.edu>
8 * Dylan Yudaken <dyudaken@gmail.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.
24 * This program is also available under a commercial proprietary license.
25 * For more information, contact us at licensing@x264.com.
26 *****************************************************************************/
30 #include "common/common.h"
31 #include "macroblock.h"
34 // Indexed by pic_struct values
35 static const uint8_t delta_tfi_divisor[10] = { 0, 2, 1, 1, 2, 2, 3, 3, 4, 6 };
37 static int x264_slicetype_frame_cost( x264_t *h, x264_mb_analysis_t *a,
38 x264_frame_t **frames, int p0, int p1, int b,
39 int b_intra_penalty );
41 static void x264_lowres_context_init( x264_t *h, x264_mb_analysis_t *a )
43 a->i_qp = X264_LOOKAHEAD_QP;
44 a->i_lambda = x264_lambda_tab[ a->i_qp ];
45 x264_mb_analyse_load_costs( h, a );
46 if( h->param.analyse.i_subpel_refine > 1 )
48 h->mb.i_me_method = X264_MIN( X264_ME_HEX, h->param.analyse.i_me_method );
49 h->mb.i_subpel_refine = 4;
53 h->mb.i_me_method = X264_ME_DIA;
54 h->mb.i_subpel_refine = 2;
56 h->mb.b_chroma_me = 0;
59 /* makes a non-h264 weight (i.e. fix7), into an h264 weight */
60 static void x264_weight_get_h264( unsigned int weight_nonh264, int offset, x264_weight_t *w )
64 w->i_scale = weight_nonh264;
65 while( w->i_denom > 0 && (w->i_scale > 127 || !(w->i_scale & 1)) )
70 w->i_scale = X264_MIN( w->i_scale, 127 );
73 static NOINLINE pixel *x264_weight_cost_init_luma( x264_t *h, x264_frame_t *fenc, x264_frame_t *ref, pixel *dest )
75 int ref0_distance = fenc->i_frame - ref->i_frame - 1;
76 /* Note: this will never run during lookahead as weights_analyse is only called if no
77 * motion search has been done. */
78 if( fenc->lowres_mvs[0][ref0_distance][0][0] != 0x7FFF )
80 int i_stride = fenc->i_stride_lowres;
81 int i_lines = fenc->i_lines_lowres;
82 int i_width = fenc->i_width_lowres;
86 for( int y = 0; y < i_lines; y += 8, p += i_stride*8 )
87 for( int x = 0; x < i_width; x += 8, i_mb_xy++ )
89 int mvx = fenc->lowres_mvs[0][ref0_distance][i_mb_xy][0];
90 int mvy = fenc->lowres_mvs[0][ref0_distance][i_mb_xy][1];
91 h->mc.mc_luma( p+x, i_stride, ref->lowres, i_stride,
92 mvx+(x<<2), mvy+(y<<2), 8, 8, weight_none );
98 return ref->lowres[0];
101 static NOINLINE unsigned int x264_weight_cost( x264_t *h, x264_frame_t *fenc, pixel *src, x264_weight_t *w )
103 unsigned int cost = 0;
104 int i_stride = fenc->i_stride_lowres;
105 int i_lines = fenc->i_lines_lowres;
106 int i_width = fenc->i_width_lowres;
107 pixel *fenc_plane = fenc->lowres[0];
108 ALIGNED_ARRAY_8( pixel, buf,[8*8] );
114 for( int y = 0; y < i_lines; y += 8, pixoff = y*i_stride )
115 for( int x = 0; x < i_width; x += 8, i_mb++, pixoff += 8)
117 w->weightfn[8>>2]( buf, 8, &src[pixoff], i_stride, w, 8 );
118 cost += X264_MIN( h->pixf.mbcmp[PIXEL_8x8]( buf, 8, &fenc_plane[pixoff], i_stride ), fenc->i_intra_cost[i_mb] );
120 /* Add cost of weights in the slice header. */
122 if( h->param.i_slice_count )
123 numslices = h->param.i_slice_count;
124 else if( h->param.i_slice_max_mbs )
125 numslices = (h->mb.i_mb_width * h->mb.i_mb_height + h->param.i_slice_max_mbs-1) / h->param.i_slice_max_mbs;
128 /* FIXME: find a way to account for --slice-max-size?
129 * Multiply by 2 as there will be a duplicate. 10 bits added as if there is a weighted frame, then an additional duplicate is used.
130 * Since using lowres frames, assume lambda = 1. */
131 cost += numslices * ( 10 + 2 * ( bs_size_ue( w[0].i_denom ) + bs_size_se( w[0].i_scale ) + bs_size_se( w[0].i_offset ) ) );
134 for( int y = 0; y < i_lines; y += 8, pixoff = y*i_stride )
135 for( int x = 0; x < i_width; x += 8, i_mb++, pixoff += 8 )
136 cost += X264_MIN( h->pixf.mbcmp[PIXEL_8x8]( &src[pixoff], i_stride, &fenc_plane[pixoff], i_stride ), fenc->i_intra_cost[i_mb] );
141 void x264_weights_analyse( x264_t *h, x264_frame_t *fenc, x264_frame_t *ref, int b_lookahead )
143 float fenc_mean, ref_mean, fenc_var, ref_var;
145 int minoff, minscale, mindenom;
146 unsigned int minscore, origscore;
147 int i_delta_index = fenc->i_frame - ref->i_frame - 1;
148 /* epsilon is chosen to require at least a numerator of 127 (with denominator = 128) */
149 const float epsilon = 1.0/128.0;
152 x264_weight_t *weights = fenc->weight[0];
154 fenc_var = round( sqrt( fenc->i_pixel_ssd[0] ) );
155 ref_var = round( sqrt( ref->i_pixel_ssd[0] ) );
156 fenc_mean = (float)fenc->i_pixel_sum[0] / (fenc->i_lines[0] * fenc->i_width[0]);
157 ref_mean = (float) ref->i_pixel_sum[0] / (fenc->i_lines[0] * fenc->i_width[0]);
160 if( fabs( ref_mean - fenc_mean ) < 0.5 && fabs( 1 - fenc_var / ref_var ) < epsilon )
162 SET_WEIGHT( weights[0], 0, 1, 0, 0 );
166 guess_scale = ref_var ? fenc_var/ref_var : 0;
167 x264_weight_get_h264( round( guess_scale * 128 ), 0, &weights[0] );
170 mindenom = weights[0].i_denom;
171 minscale = weights[0].i_scale;
173 offset_search = x264_clip3( floor( fenc_mean - ref_mean * minscale / (1 << mindenom) + 0.5f*b_lookahead ), -128, 126 );
175 if( !fenc->b_intra_calculated )
177 x264_mb_analysis_t a;
178 x264_lowres_context_init( h, &a );
179 x264_slicetype_frame_cost( h, &a, &fenc, 0, 0, 0, 0 );
181 pixel *mcbuf = x264_weight_cost_init_luma( h, fenc, ref, h->mb.p_weight_buf[0] );
182 origscore = minscore = x264_weight_cost( h, fenc, mcbuf, 0 );
186 SET_WEIGHT( weights[0], 0, 1, 0, 0 );
190 // This gives a slight improvement due to rounding errors but only tests
191 // one offset on lookahead.
192 // TODO: currently searches only offset +1. try other offsets/multipliers/combinations thereof?
193 for( int i_off = offset_search; i_off <= offset_search+!b_lookahead; i_off++ )
195 SET_WEIGHT( weights[0], 1, minscale, mindenom, i_off );
196 unsigned int s = x264_weight_cost( h, fenc, mcbuf, &weights[0] );
197 COPY3_IF_LT( minscore, s, minoff, i_off, found, 1 );
201 /* FIXME: More analysis can be done here on SAD vs. SATD termination. */
202 /* 0.2% termination derived experimentally to avoid weird weights in frames that are mostly intra. */
203 if( !found || (minscale == 1<<mindenom && minoff == 0) || (float)minscore / origscore > 0.998 )
205 SET_WEIGHT( weights[0], 0, 1, 0, 0 );
209 SET_WEIGHT( weights[0], 1, minscale, mindenom, minoff );
211 if( h->param.analyse.i_weighted_pred == X264_WEIGHTP_FAKE && weights[0].weightfn )
212 fenc->f_weighted_cost_delta[i_delta_index] = (float)minscore / origscore;
214 if( weights[0].weightfn && b_lookahead )
216 //scale lowres in lookahead for slicetype_frame_cost
217 pixel *src = ref->buffer_lowres[0];
218 pixel *dst = h->mb.p_weight_buf[0];
219 int width = ref->i_width_lowres + PADH*2;
220 int height = ref->i_lines_lowres + PADV*2;
221 x264_weight_scale_plane( h, dst, ref->i_stride_lowres, src, ref->i_stride_lowres,
222 width, height, &weights[0] );
223 fenc->weighted[0] = h->mb.p_weight_buf[0] + PADH + ref->i_stride_lowres * PADV;
227 static void x264_slicetype_mb_cost( x264_t *h, x264_mb_analysis_t *a,
228 x264_frame_t **frames, int p0, int p1, int b,
229 int dist_scale_factor, int do_search[2], const x264_weight_t *w )
231 x264_frame_t *fref0 = frames[p0];
232 x264_frame_t *fref1 = frames[p1];
233 x264_frame_t *fenc = frames[b];
234 const int b_bidir = (b < p1);
235 const int i_mb_x = h->mb.i_mb_x;
236 const int i_mb_y = h->mb.i_mb_y;
237 const int i_mb_stride = h->mb.i_mb_width;
238 const int i_mb_xy = i_mb_x + i_mb_y * i_mb_stride;
239 const int i_stride = fenc->i_stride_lowres;
240 const int i_pel_offset = 8 * (i_mb_x + i_mb_y * i_stride);
241 const int i_bipred_weight = h->param.analyse.b_weighted_bipred ? 64 - (dist_scale_factor>>2) : 32;
242 int16_t (*fenc_mvs[2])[2] = { &frames[b]->lowres_mvs[0][b-p0-1][i_mb_xy], &frames[b]->lowres_mvs[1][p1-b-1][i_mb_xy] };
243 int (*fenc_costs[2]) = { &frames[b]->lowres_mv_costs[0][b-p0-1][i_mb_xy], &frames[b]->lowres_mv_costs[1][p1-b-1][i_mb_xy] };
244 int b_frame_score_mb = (i_mb_x > 0 && i_mb_x < h->mb.i_mb_width - 1 &&
245 i_mb_y > 0 && i_mb_y < h->mb.i_mb_height - 1) ||
246 h->mb.i_mb_width <= 2 || h->mb.i_mb_height <= 2;
248 ALIGNED_ARRAY_8( pixel, pix1,[9*FDEC_STRIDE] );
249 pixel *pix2 = pix1+8;
251 int i_bcost = COST_MAX;
254 h->mb.pic.p_fenc[0] = h->mb.pic.fenc_buf;
255 h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fenc[0], FENC_STRIDE, &fenc->lowres[0][i_pel_offset], i_stride, 8 );
258 goto lowres_intra_mb;
260 // no need for h->mb.mv_min[]
261 h->mb.mv_min_fpel[0] = -8*h->mb.i_mb_x - 4;
262 h->mb.mv_max_fpel[0] = 8*( h->mb.i_mb_width - h->mb.i_mb_x - 1 ) + 4;
263 h->mb.mv_min_spel[0] = 4*( h->mb.mv_min_fpel[0] - 8 );
264 h->mb.mv_max_spel[0] = 4*( h->mb.mv_max_fpel[0] + 8 );
265 if( h->mb.i_mb_x >= h->mb.i_mb_width - 2 )
267 h->mb.mv_min_fpel[1] = -8*h->mb.i_mb_y - 4;
268 h->mb.mv_max_fpel[1] = 8*( h->mb.i_mb_height - h->mb.i_mb_y - 1 ) + 4;
269 h->mb.mv_min_spel[1] = 4*( h->mb.mv_min_fpel[1] - 8 );
270 h->mb.mv_max_spel[1] = 4*( h->mb.mv_max_fpel[1] + 8 );
273 #define LOAD_HPELS_LUMA(dst, src) \
275 (dst)[0] = &(src)[0][i_pel_offset]; \
276 (dst)[1] = &(src)[1][i_pel_offset]; \
277 (dst)[2] = &(src)[2][i_pel_offset]; \
278 (dst)[3] = &(src)[3][i_pel_offset]; \
280 #define LOAD_WPELS_LUMA(dst,src) \
281 (dst) = &(src)[i_pel_offset];
283 #define CLIP_MV( mv ) \
285 mv[0] = x264_clip3( mv[0], h->mb.mv_min_spel[0], h->mb.mv_max_spel[0] ); \
286 mv[1] = x264_clip3( mv[1], h->mb.mv_min_spel[1], h->mb.mv_max_spel[1] ); \
288 #define TRY_BIDIR( mv0, mv1, penalty ) \
291 if( h->param.analyse.i_subpel_refine <= 1 ) \
293 int hpel_idx1 = (((mv0)[0]&2)>>1) + ((mv0)[1]&2); \
294 int hpel_idx2 = (((mv1)[0]&2)>>1) + ((mv1)[1]&2); \
295 pixel *src1 = m[0].p_fref[hpel_idx1] + ((mv0)[0]>>2) + ((mv0)[1]>>2) * m[0].i_stride[0]; \
296 pixel *src2 = m[1].p_fref[hpel_idx2] + ((mv1)[0]>>2) + ((mv1)[1]>>2) * m[1].i_stride[0]; \
297 h->mc.avg[PIXEL_8x8]( pix1, 16, src1, m[0].i_stride[0], src2, m[1].i_stride[0], i_bipred_weight ); \
301 int stride1 = 16, stride2 = 16; \
302 pixel *src1, *src2; \
303 src1 = h->mc.get_ref( pix1, &stride1, m[0].p_fref, m[0].i_stride[0], \
304 (mv0)[0], (mv0)[1], 8, 8, w ); \
305 src2 = h->mc.get_ref( pix2, &stride2, m[1].p_fref, m[1].i_stride[0], \
306 (mv1)[0], (mv1)[1], 8, 8, w ); \
307 h->mc.avg[PIXEL_8x8]( pix1, 16, src1, stride1, src2, stride2, i_bipred_weight ); \
309 i_cost = penalty * a->i_lambda + h->pixf.mbcmp[PIXEL_8x8]( \
310 m[0].p_fenc[0], FENC_STRIDE, pix1, 16 ); \
311 COPY2_IF_LT( i_bcost, i_cost, list_used, 3 ); \
314 m[0].i_pixel = PIXEL_8x8;
315 m[0].p_cost_mv = a->p_cost_mv;
316 m[0].i_stride[0] = i_stride;
317 m[0].p_fenc[0] = h->mb.pic.p_fenc[0];
320 LOAD_HPELS_LUMA( m[0].p_fref, fref0->lowres );
321 m[0].p_fref_w = m[0].p_fref[0];
323 LOAD_WPELS_LUMA( m[0].p_fref_w, fenc->weighted[0] );
327 int16_t *mvr = fref1->lowres_mvs[0][p1-p0-1][i_mb_xy];
328 ALIGNED_ARRAY_8( int16_t, dmv,[2],[2] );
330 m[1].i_pixel = PIXEL_8x8;
331 m[1].p_cost_mv = a->p_cost_mv;
332 m[1].i_stride[0] = i_stride;
333 m[1].p_fenc[0] = h->mb.pic.p_fenc[0];
335 m[1].weight = weight_none;
336 LOAD_HPELS_LUMA( m[1].p_fref, fref1->lowres );
337 m[1].p_fref_w = m[1].p_fref[0];
339 dmv[0][0] = ( mvr[0] * dist_scale_factor + 128 ) >> 8;
340 dmv[0][1] = ( mvr[1] * dist_scale_factor + 128 ) >> 8;
341 dmv[1][0] = dmv[0][0] - mvr[0];
342 dmv[1][1] = dmv[0][1] - mvr[1];
345 if( h->param.analyse.i_subpel_refine <= 1 )
346 M64( dmv ) &= ~0x0001000100010001ULL; /* mv & ~1 */
348 TRY_BIDIR( dmv[0], dmv[1], 0 );
352 h->mc.avg[PIXEL_8x8]( pix1, 16, m[0].p_fref[0], m[0].i_stride[0], m[1].p_fref[0], m[1].i_stride[0], i_bipred_weight );
353 i_cost = h->pixf.mbcmp[PIXEL_8x8]( m[0].p_fenc[0], FENC_STRIDE, pix1, 16 );
354 COPY2_IF_LT( i_bcost, i_cost, list_used, 3 );
358 for( int l = 0; l < 1 + b_bidir; l++ )
363 int16_t (*fenc_mv)[2] = fenc_mvs[l];
364 ALIGNED_4( int16_t mvc[4][2] );
366 /* Reverse-order MV prediction. */
369 #define MVC(mv) { CP32( mvc[i_mvc], mv ); i_mvc++; }
370 if( i_mb_x < h->mb.i_mb_width - 1 )
372 if( i_mb_y < h->mb.i_mb_height - 1 )
374 MVC( fenc_mv[i_mb_stride] );
376 MVC( fenc_mv[i_mb_stride-1] );
377 if( i_mb_x < h->mb.i_mb_width - 1 )
378 MVC( fenc_mv[i_mb_stride+1] );
382 CP32( m[l].mvp, mvc[0] );
384 x264_median_mv( m[l].mvp, mvc[0], mvc[1], mvc[2] );
386 /* Fast skip for cases of near-zero residual. Shortcut: don't bother except in the mv0 case,
387 * since anything else is likely to have enough residual to not trigger the skip. */
388 if( !M32( m[l].mvp ) )
390 m[l].cost = h->pixf.mbcmp[PIXEL_8x8]( m[l].p_fenc[0], FENC_STRIDE, m[l].p_fref[0], m[l].i_stride[0] );
398 x264_me_search( h, &m[l], mvc, i_mvc );
399 m[l].cost -= 2 * a->i_lambda; // remove mvcost from skip mbs
401 m[l].cost += 5 * a->i_lambda;
404 CP32( fenc_mvs[l], m[l].mv );
405 *fenc_costs[l] = m[l].cost;
409 CP32( m[l].mv, fenc_mvs[l] );
410 m[l].cost = *fenc_costs[l];
412 COPY2_IF_LT( i_bcost, m[l].cost, list_used, l+1 );
415 if( b_bidir && ( M32( m[0].mv ) || M32( m[1].mv ) ) )
416 TRY_BIDIR( m[0].mv, m[1].mv, 5 );
419 if( !fenc->b_intra_calculated )
421 ALIGNED_ARRAY_16( pixel, edge,[33] );
422 pixel *pix = &pix1[8+FDEC_STRIDE - 1];
423 pixel *src = &fenc->lowres[0][i_pel_offset - 1];
424 const int intra_penalty = 5 * a->i_lambda;
427 memcpy( pix-FDEC_STRIDE, src-i_stride, 17 * sizeof(pixel) );
428 for( int i = 0; i < 8; i++ )
429 pix[i*FDEC_STRIDE] = src[i*i_stride];
432 h->pixf.intra_mbcmp_x3_8x8c( h->mb.pic.p_fenc[0], pix, satds );
433 int i_icost = X264_MIN3( satds[0], satds[1], satds[2] );
435 if( h->param.analyse.i_subpel_refine > 1 )
437 h->predict_8x8c[I_PRED_CHROMA_P]( pix );
438 int satd = h->pixf.mbcmp[PIXEL_8x8]( pix, FDEC_STRIDE, h->mb.pic.p_fenc[0], FENC_STRIDE );
439 i_icost = X264_MIN( i_icost, satd );
440 h->predict_8x8_filter( pix, edge, ALL_NEIGHBORS, ALL_NEIGHBORS );
441 for( int i = 3; i < 9; i++ )
443 h->predict_8x8[i]( pix, edge );
444 satd = h->pixf.mbcmp[PIXEL_8x8]( pix, FDEC_STRIDE, h->mb.pic.p_fenc[0], FENC_STRIDE );
445 i_icost = X264_MIN( i_icost, satd );
449 i_icost += intra_penalty;
450 fenc->i_intra_cost[i_mb_xy] = i_icost;
451 if( b_frame_score_mb )
453 int *row_satd_intra = frames[b]->i_row_satds[0][0];
454 int i_icost_aq = i_icost;
455 if( h->param.rc.i_aq_mode )
456 i_icost_aq = (i_icost_aq * frames[b]->i_inv_qscale_factor[i_mb_xy] + 128) >> 8;
457 fenc->i_cost_est[0][0] += i_icost;
458 fenc->i_cost_est_aq[0][0] += i_icost_aq;
459 row_satd_intra[h->mb.i_mb_y] += i_icost_aq;
463 /* forbid intra-mbs in B-frames, because it's rare and not worth checking */
464 /* FIXME: Should we still forbid them now that we cache intra scores? */
467 int i_icost = fenc->i_intra_cost[i_mb_xy];
468 int b_intra = i_icost < i_bcost;
474 if( b_frame_score_mb )
475 fenc->i_intra_mbs[b-p0] += b_intra;
478 /* In an I-frame, we've already added the results above in the intra section. */
481 int i_bcost_aq = i_bcost;
482 if( h->param.rc.i_aq_mode )
483 i_bcost_aq = (i_bcost_aq * frames[b]->i_inv_qscale_factor[i_mb_xy] + 128) >> 8;
484 fenc->i_row_satds[b-p0][p1-b][h->mb.i_mb_y] += i_bcost_aq;
485 if( b_frame_score_mb )
487 /* Don't use AQ-weighted costs for slicetype decision, only for ratecontrol. */
488 frames[b]->i_cost_est[b-p0][p1-b] += i_bcost;
489 frames[b]->i_cost_est_aq[b-p0][p1-b] += i_bcost_aq;
493 fenc->lowres_costs[b-p0][p1-b][i_mb_xy] = X264_MIN( i_bcost, LOWRES_COST_MASK ) + (list_used << LOWRES_COST_SHIFT);
498 (h->mb.i_mb_width > 2 && h->mb.i_mb_height > 2 ?\
499 (h->mb.i_mb_width - 2) * (h->mb.i_mb_height - 2) :\
500 h->mb.i_mb_width * h->mb.i_mb_height)
502 static int x264_slicetype_frame_cost( x264_t *h, x264_mb_analysis_t *a,
503 x264_frame_t **frames, int p0, int p1, int b,
504 int b_intra_penalty )
508 const x264_weight_t *w = weight_none;
509 /* Check whether we already evaluated this frame
510 * If we have tried this frame as P, then we have also tried
511 * the preceding frames as B. (is this still true?) */
512 /* Also check that we already calculated the row SATDs for the current frame. */
513 if( frames[b]->i_cost_est[b-p0][p1-b] >= 0 && (!h->param.rc.i_vbv_buffer_size || frames[b]->i_row_satds[b-p0][p1-b][0] != -1) )
514 i_score = frames[b]->i_cost_est[b-p0][p1-b];
517 int dist_scale_factor = 128;
518 int *row_satd = frames[b]->i_row_satds[b-p0][p1-b];
519 int *row_satd_intra = frames[b]->i_row_satds[0][0];
521 /* For each list, check to see whether we have lowres motion-searched this reference frame before. */
522 do_search[0] = b != p0 && frames[b]->lowres_mvs[0][b-p0-1][0][0] == 0x7FFF;
523 do_search[1] = b != p1 && frames[b]->lowres_mvs[1][p1-b-1][0][0] == 0x7FFF;
526 if( ( h->param.analyse.i_weighted_pred == X264_WEIGHTP_SMART ||
527 h->param.analyse.i_weighted_pred == X264_WEIGHTP_FAKE ) && b == p1 )
530 x264_weights_analyse( h, frames[b], frames[p0], 1 );
531 w = frames[b]->weight[0];
533 frames[b]->lowres_mvs[0][b-p0-1][0][0] = 0;
535 if( do_search[1] ) frames[b]->lowres_mvs[1][p1-b-1][0][0] = 0;
538 frames[b]->i_intra_mbs[b-p0] = 0;
539 if( !frames[b]->b_intra_calculated )
541 frames[b]->i_cost_est[0][0] = 0;
542 frames[b]->i_cost_est_aq[0][0] = 0;
545 dist_scale_factor = ( ((b-p0) << 8) + ((p1-p0) >> 1) ) / (p1-p0);
547 frames[b]->i_cost_est[b-p0][p1-b] = 0;
548 frames[b]->i_cost_est_aq[b-p0][p1-b] = 0;
550 /* Lowres lookahead goes backwards because the MVs are used as predictors in the main encode.
551 * This considerably improves MV prediction overall. */
553 /* The edge mbs seem to reduce the predictive quality of the
554 * whole frame's score, but are needed for a spatial distribution. */
555 if( h->param.rc.b_mb_tree || h->param.rc.i_vbv_buffer_size ||
556 h->mb.i_mb_width <= 2 || h->mb.i_mb_height <= 2 )
558 for( h->mb.i_mb_y = h->mb.i_mb_height - 1; h->mb.i_mb_y >= 0; h->mb.i_mb_y-- )
560 row_satd[h->mb.i_mb_y] = 0;
561 if( !frames[b]->b_intra_calculated )
562 row_satd_intra[h->mb.i_mb_y] = 0;
563 for( h->mb.i_mb_x = h->mb.i_mb_width - 1; h->mb.i_mb_x >= 0; h->mb.i_mb_x-- )
564 x264_slicetype_mb_cost( h, a, frames, p0, p1, b, dist_scale_factor, do_search, w );
569 for( h->mb.i_mb_y = h->mb.i_mb_height - 2; h->mb.i_mb_y >= 1; h->mb.i_mb_y-- )
570 for( h->mb.i_mb_x = h->mb.i_mb_width - 2; h->mb.i_mb_x >= 1; h->mb.i_mb_x-- )
571 x264_slicetype_mb_cost( h, a, frames, p0, p1, b, dist_scale_factor, do_search, w );
574 i_score = frames[b]->i_cost_est[b-p0][p1-b];
576 i_score = (uint64_t)i_score * 100 / (120 + h->param.i_bframe_bias);
578 frames[b]->b_intra_calculated = 1;
580 frames[b]->i_cost_est[b-p0][p1-b] = i_score;
584 if( b_intra_penalty )
586 // arbitrary penalty for I-blocks after B-frames
588 i_score += i_score * frames[b]->i_intra_mbs[b-p0] / (nmb * 8);
593 /* If MB-tree changes the quantizers, we need to recalculate the frame cost without
594 * re-running lookahead. */
595 static int x264_slicetype_frame_cost_recalculate( x264_t *h, x264_frame_t **frames, int p0, int p1, int b )
598 int *row_satd = frames[b]->i_row_satds[b-p0][p1-b];
599 float *qp_offset = IS_X264_TYPE_B(frames[b]->i_type) ? frames[b]->f_qp_offset_aq : frames[b]->f_qp_offset;
601 for( h->mb.i_mb_y = h->mb.i_mb_height - 1; h->mb.i_mb_y >= 0; h->mb.i_mb_y-- )
603 row_satd[ h->mb.i_mb_y ] = 0;
604 for( h->mb.i_mb_x = h->mb.i_mb_width - 1; h->mb.i_mb_x >= 0; h->mb.i_mb_x-- )
606 int i_mb_xy = h->mb.i_mb_x + h->mb.i_mb_y*h->mb.i_mb_stride;
607 int i_mb_cost = frames[b]->lowres_costs[b-p0][p1-b][i_mb_xy] & LOWRES_COST_MASK;
608 float qp_adj = qp_offset[i_mb_xy];
609 i_mb_cost = (i_mb_cost * x264_exp2fix8(qp_adj) + 128) >> 8;
610 row_satd[ h->mb.i_mb_y ] += i_mb_cost;
611 if( (h->mb.i_mb_y > 0 && h->mb.i_mb_y < h->mb.i_mb_height - 1 &&
612 h->mb.i_mb_x > 0 && h->mb.i_mb_x < h->mb.i_mb_width - 1) ||
613 h->mb.i_mb_width <= 2 || h->mb.i_mb_height <= 2 )
615 i_score += i_mb_cost;
622 static void x264_macroblock_tree_finish( x264_t *h, x264_frame_t *frame, int ref0_distance )
625 float weightdelta = 0.0;
626 if( ref0_distance && frame->f_weighted_cost_delta[ref0_distance-1] > 0 )
627 weightdelta = (1.0 - frame->f_weighted_cost_delta[ref0_distance-1]);
629 /* Allow the strength to be adjusted via qcompress, since the two
630 * concepts are very similar. */
631 float strength = 5.0f * (1.0f - h->param.rc.f_qcompress);
632 for( int mb_index = 0; mb_index < h->mb.i_mb_count; mb_index++ )
634 int intra_cost = (frame->i_intra_cost[mb_index] * frame->i_inv_qscale_factor[mb_index]+128)>>8;
637 int propagate_cost = frame->i_propagate_cost[mb_index];
638 float log2_ratio = x264_log2(intra_cost + propagate_cost) - x264_log2(intra_cost) + weightdelta;
639 frame->f_qp_offset[mb_index] = frame->f_qp_offset_aq[mb_index] - strength * log2_ratio;
644 static void x264_macroblock_tree_propagate( x264_t *h, x264_frame_t **frames, int p0, int p1, int b, int referenced )
646 uint16_t *ref_costs[2] = {frames[p0]->i_propagate_cost,frames[p1]->i_propagate_cost};
647 int dist_scale_factor = ( ((b-p0) << 8) + ((p1-p0) >> 1) ) / (p1-p0);
648 int i_bipred_weight = h->param.analyse.b_weighted_bipred ? 64 - (dist_scale_factor>>2) : 32;
649 int16_t (*mvs[2])[2] = { frames[b]->lowres_mvs[0][b-p0-1], frames[b]->lowres_mvs[1][p1-b-1] };
650 int bipred_weights[2] = {i_bipred_weight, 64 - i_bipred_weight};
651 int *buf = h->scratch_buffer;
652 uint16_t *propagate_cost = frames[b]->i_propagate_cost;
654 /* For non-reffed frames the source costs are always zero, so just memset one row and re-use it. */
656 memset( frames[b]->i_propagate_cost, 0, h->mb.i_mb_width * sizeof(uint16_t) );
658 for( h->mb.i_mb_y = 0; h->mb.i_mb_y < h->mb.i_mb_height; h->mb.i_mb_y++ )
660 int mb_index = h->mb.i_mb_y*h->mb.i_mb_stride;
661 h->mc.mbtree_propagate_cost( buf, propagate_cost,
662 frames[b]->i_intra_cost+mb_index, frames[b]->lowres_costs[b-p0][p1-b]+mb_index,
663 frames[b]->i_inv_qscale_factor+mb_index, h->mb.i_mb_width );
665 propagate_cost += h->mb.i_mb_width;
666 for( h->mb.i_mb_x = 0; h->mb.i_mb_x < h->mb.i_mb_width; h->mb.i_mb_x++, mb_index++ )
668 int propagate_amount = buf[h->mb.i_mb_x];
669 /* Don't propagate for an intra block. */
670 if( propagate_amount > 0 )
672 /* Access width-2 bitfield. */
673 int lists_used = frames[b]->lowres_costs[b-p0][p1-b][mb_index] >> LOWRES_COST_SHIFT;
674 /* Follow the MVs to the previous frame(s). */
675 for( int list = 0; list < 2; list++ )
676 if( (lists_used >> list)&1 )
678 #define CLIP_ADD(s,x) (s) = X264_MIN((s)+(x),(1<<16)-1)
679 int listamount = propagate_amount;
680 /* Apply bipred weighting. */
681 if( lists_used == 3 )
682 listamount = (listamount * bipred_weights[list] + 32) >> 6;
684 /* Early termination for simple case of mv0. */
685 if( !M32( mvs[list][mb_index] ) )
687 CLIP_ADD( ref_costs[list][mb_index], listamount );
691 int x = mvs[list][mb_index][0];
692 int y = mvs[list][mb_index][1];
693 int mbx = (x>>5)+h->mb.i_mb_x;
694 int mby = (y>>5)+h->mb.i_mb_y;
695 int idx0 = mbx + mby * h->mb.i_mb_stride;
697 int idx2 = idx0 + h->mb.i_mb_stride;
698 int idx3 = idx0 + h->mb.i_mb_stride + 1;
701 int idx0weight = (32-y)*(32-x);
702 int idx1weight = (32-y)*x;
703 int idx2weight = y*(32-x);
704 int idx3weight = y*x;
706 /* We could just clip the MVs, but pixels that lie outside the frame probably shouldn't
708 if( mbx < h->mb.i_mb_width-1 && mby < h->mb.i_mb_height-1 && mbx >= 0 && mby >= 0 )
710 CLIP_ADD( ref_costs[list][idx0], (listamount*idx0weight+512)>>10 );
711 CLIP_ADD( ref_costs[list][idx1], (listamount*idx1weight+512)>>10 );
712 CLIP_ADD( ref_costs[list][idx2], (listamount*idx2weight+512)>>10 );
713 CLIP_ADD( ref_costs[list][idx3], (listamount*idx3weight+512)>>10 );
715 else /* Check offsets individually */
717 if( mbx < h->mb.i_mb_width && mby < h->mb.i_mb_height && mbx >= 0 && mby >= 0 )
718 CLIP_ADD( ref_costs[list][idx0], (listamount*idx0weight+512)>>10 );
719 if( mbx+1 < h->mb.i_mb_width && mby < h->mb.i_mb_height && mbx+1 >= 0 && mby >= 0 )
720 CLIP_ADD( ref_costs[list][idx1], (listamount*idx1weight+512)>>10 );
721 if( mbx < h->mb.i_mb_width && mby+1 < h->mb.i_mb_height && mbx >= 0 && mby+1 >= 0 )
722 CLIP_ADD( ref_costs[list][idx2], (listamount*idx2weight+512)>>10 );
723 if( mbx+1 < h->mb.i_mb_width && mby+1 < h->mb.i_mb_height && mbx+1 >= 0 && mby+1 >= 0 )
724 CLIP_ADD( ref_costs[list][idx3], (listamount*idx3weight+512)>>10 );
731 if( h->param.rc.i_vbv_buffer_size && h->param.rc.i_lookahead && referenced )
732 x264_macroblock_tree_finish( h, frames[b], b == p1 ? b - p0 : 0 );
735 static void x264_macroblock_tree( x264_t *h, x264_mb_analysis_t *a, x264_frame_t **frames, int num_frames, int b_intra )
738 int last_nonb, cur_nonb = 1;
743 x264_slicetype_frame_cost( h, a, frames, 0, 0, 0, 0 );
745 while( i > 0 && frames[i]->i_type == X264_TYPE_B )
749 /* Lookaheadless MB-tree is not a theoretically distinct case; the same extrapolation could
750 * be applied to the end of a lookahead buffer of any size. However, it's most needed when
751 * lookahead=0, so that's what's currently implemented. */
752 if( !h->param.rc.i_lookahead )
756 memset( frames[0]->i_propagate_cost, 0, h->mb.i_mb_count * sizeof(uint16_t) );
757 memcpy( frames[0]->f_qp_offset, frames[0]->f_qp_offset_aq, h->mb.i_mb_count * sizeof(float) );
760 XCHG( uint16_t*, frames[last_nonb]->i_propagate_cost, frames[0]->i_propagate_cost );
761 memset( frames[0]->i_propagate_cost, 0, h->mb.i_mb_count * sizeof(uint16_t) );
765 if( last_nonb < idx )
767 memset( frames[last_nonb]->i_propagate_cost, 0, h->mb.i_mb_count * sizeof(uint16_t) );
773 while( frames[cur_nonb]->i_type == X264_TYPE_B && cur_nonb > 0 )
777 x264_slicetype_frame_cost( h, a, frames, cur_nonb, last_nonb, last_nonb, 0 );
778 memset( frames[cur_nonb]->i_propagate_cost, 0, h->mb.i_mb_count * sizeof(uint16_t) );
779 bframes = last_nonb - cur_nonb - 1;
780 if( h->param.i_bframe_pyramid && bframes > 1 )
782 int middle = (bframes + 1)/2 + cur_nonb;
783 x264_slicetype_frame_cost( h, a, frames, cur_nonb, last_nonb, middle, 0 );
784 memset( frames[middle]->i_propagate_cost, 0, h->mb.i_mb_count * sizeof(uint16_t) );
785 while( i > cur_nonb )
787 int p0 = i > middle ? middle : cur_nonb;
788 int p1 = i < middle ? middle : last_nonb;
791 x264_slicetype_frame_cost( h, a, frames, p0, p1, i, 0 );
792 x264_macroblock_tree_propagate( h, frames, p0, p1, i, 0 );
796 x264_macroblock_tree_propagate( h, frames, cur_nonb, last_nonb, middle, 1 );
800 while( i > cur_nonb )
802 x264_slicetype_frame_cost( h, a, frames, cur_nonb, last_nonb, i, 0 );
803 x264_macroblock_tree_propagate( h, frames, cur_nonb, last_nonb, i, 0 );
807 x264_macroblock_tree_propagate( h, frames, cur_nonb, last_nonb, last_nonb, 1 );
808 last_nonb = cur_nonb;
811 if( !h->param.rc.i_lookahead )
813 x264_macroblock_tree_propagate( h, frames, 0, last_nonb, last_nonb, 1 );
814 XCHG( uint16_t*, frames[last_nonb]->i_propagate_cost, frames[0]->i_propagate_cost );
817 x264_macroblock_tree_finish( h, frames[last_nonb], last_nonb );
818 if( h->param.i_bframe_pyramid && bframes > 1 && !h->param.rc.i_vbv_buffer_size )
819 x264_macroblock_tree_finish( h, frames[last_nonb+(bframes+1)/2], 0 );
822 static int x264_vbv_frame_cost( x264_t *h, x264_mb_analysis_t *a, x264_frame_t **frames, int p0, int p1, int b )
824 int cost = x264_slicetype_frame_cost( h, a, frames, p0, p1, b, 0 );
825 if( h->param.rc.i_aq_mode )
827 if( h->param.rc.b_mb_tree )
828 return x264_slicetype_frame_cost_recalculate( h, frames, p0, p1, b );
830 return frames[b]->i_cost_est_aq[b-p0][p1-b];
835 static void x264_calculate_durations( x264_t *h, x264_frame_t *cur_frame, x264_frame_t *prev_frame, int *i_cpb_delay, int *i_coded_fields )
837 cur_frame->i_cpb_delay = *i_cpb_delay;
838 cur_frame->i_dpb_output_delay = cur_frame->i_field_cnt - *i_coded_fields;
840 // add a correction term for frame reordering
841 cur_frame->i_dpb_output_delay += h->sps->vui.i_num_reorder_frames*2;
843 // fix possible negative dpb_output_delay because of pulldown changes and reordering
844 if( cur_frame->i_dpb_output_delay < 0 )
846 cur_frame->i_cpb_delay += cur_frame->i_dpb_output_delay;
847 cur_frame->i_dpb_output_delay = 0;
849 prev_frame->i_cpb_duration += cur_frame->i_dpb_output_delay;
852 if( cur_frame->b_keyframe )
855 *i_cpb_delay += cur_frame->i_duration;
856 *i_coded_fields += cur_frame->i_duration;
857 cur_frame->i_cpb_duration = cur_frame->i_duration;
860 static void x264_vbv_lookahead( x264_t *h, x264_mb_analysis_t *a, x264_frame_t **frames, int num_frames, int keyframe )
862 int last_nonb = 0, cur_nonb = 1, idx = 0;
863 x264_frame_t *prev_frame = NULL;
864 int prev_frame_idx = 0;
865 while( cur_nonb < num_frames && frames[cur_nonb]->i_type == X264_TYPE_B )
867 int next_nonb = keyframe ? last_nonb : cur_nonb;
869 if( frames[cur_nonb]->i_coded_fields_lookahead >= 0 )
871 h->i_coded_fields_lookahead = frames[cur_nonb]->i_coded_fields_lookahead;
872 h->i_cpb_delay_lookahead = frames[cur_nonb]->i_cpb_delay_lookahead;
875 while( cur_nonb < num_frames )
877 /* P/I cost: This shouldn't include the cost of next_nonb */
878 if( next_nonb != cur_nonb )
880 int p0 = IS_X264_TYPE_I( frames[cur_nonb]->i_type ) ? cur_nonb : last_nonb;
881 frames[next_nonb]->i_planned_satd[idx] = x264_vbv_frame_cost( h, a, frames, p0, cur_nonb, cur_nonb );
882 frames[next_nonb]->i_planned_type[idx] = frames[cur_nonb]->i_type;
883 frames[cur_nonb]->i_coded_fields_lookahead = h->i_coded_fields_lookahead;
884 frames[cur_nonb]->i_cpb_delay_lookahead = h->i_cpb_delay_lookahead;
885 x264_calculate_durations( h, frames[cur_nonb], prev_frame, &h->i_cpb_delay_lookahead, &h->i_coded_fields_lookahead );
888 frames[next_nonb]->f_planned_cpb_duration[prev_frame_idx] = (double)prev_frame->i_cpb_duration *
889 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
891 frames[next_nonb]->f_planned_cpb_duration[idx] = (double)frames[cur_nonb]->i_cpb_duration *
892 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
893 prev_frame = frames[cur_nonb];
894 prev_frame_idx = idx;
897 /* Handle the B-frames: coded order */
898 for( int i = last_nonb+1; i < cur_nonb; i++, idx++ )
900 frames[next_nonb]->i_planned_satd[idx] = x264_vbv_frame_cost( h, a, frames, last_nonb, cur_nonb, i );
901 frames[next_nonb]->i_planned_type[idx] = X264_TYPE_B;
902 frames[i]->i_coded_fields_lookahead = h->i_coded_fields_lookahead;
903 frames[i]->i_cpb_delay_lookahead = h->i_cpb_delay_lookahead;
904 x264_calculate_durations( h, frames[i], prev_frame, &h->i_cpb_delay_lookahead, &h->i_coded_fields_lookahead );
907 frames[next_nonb]->f_planned_cpb_duration[prev_frame_idx] = (double)prev_frame->i_cpb_duration *
908 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
910 frames[next_nonb]->f_planned_cpb_duration[idx] = (double)frames[i]->i_cpb_duration *
911 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
912 prev_frame = frames[i];
913 prev_frame_idx = idx;
915 last_nonb = cur_nonb;
917 while( cur_nonb <= num_frames && frames[cur_nonb]->i_type == X264_TYPE_B )
920 frames[next_nonb]->i_planned_type[idx] = X264_TYPE_AUTO;
923 static int x264_slicetype_path_cost( x264_t *h, x264_mb_analysis_t *a, x264_frame_t **frames, char *path, int threshold )
928 path--; /* Since the 1st path element is really the second frame */
932 /* Find the location of the next P-frame. */
933 while( path[next_p] != 'P' )
936 /* Add the cost of the P-frame found above */
937 cost += x264_slicetype_frame_cost( h, a, frames, cur_p, next_p, next_p, 0 );
938 /* Early terminate if the cost we have found is larger than the best path cost so far */
939 if( cost > threshold )
942 if( h->param.i_bframe_pyramid && next_p - cur_p > 2 )
944 int middle = cur_p + (next_p - cur_p)/2;
945 cost += x264_slicetype_frame_cost( h, a, frames, cur_p, next_p, middle, 0 );
946 for( int next_b = loc; next_b < middle && cost < threshold; next_b++ )
947 cost += x264_slicetype_frame_cost( h, a, frames, cur_p, middle, next_b, 0 );
948 for( int next_b = middle+1; next_b < next_p && cost < threshold; next_b++ )
949 cost += x264_slicetype_frame_cost( h, a, frames, middle, next_p, next_b, 0 );
952 for( int next_b = loc; next_b < next_p && cost < threshold; next_b++ )
953 cost += x264_slicetype_frame_cost( h, a, frames, cur_p, next_p, next_b, 0 );
961 /* Viterbi/trellis slicetype decision algorithm. */
962 /* Uses strings due to the fact that the speed of the control functions is
963 negligible compared to the cost of running slicetype_frame_cost, and because
964 it makes debugging easier. */
965 static void x264_slicetype_path( x264_t *h, x264_mb_analysis_t *a, x264_frame_t **frames, int length, char (*best_paths)[X264_LOOKAHEAD_MAX] )
967 char paths[2][X264_LOOKAHEAD_MAX];
968 int num_paths = X264_MIN( h->param.i_bframe+1, length );
969 int best_cost = COST_MAX;
972 /* Iterate over all currently possible paths */
973 for( int path = 0; path < num_paths; path++ )
975 /* Add suffixes to the current path */
976 int len = length - (path + 1);
977 memcpy( paths[idx], best_paths[len % (X264_BFRAME_MAX+1)], len );
978 memset( paths[idx]+len, 'B', path );
979 strcpy( paths[idx]+len+path, "P" );
981 /* Calculate the actual cost of the current path */
982 int cost = x264_slicetype_path_cost( h, a, frames, paths[idx], best_cost );
983 if( cost < best_cost )
990 /* Store the best path. */
991 memcpy( best_paths[length % (X264_BFRAME_MAX+1)], paths[idx^1], length );
994 static int scenecut_internal( x264_t *h, x264_mb_analysis_t *a, x264_frame_t **frames, int p0, int p1, int print )
996 x264_frame_t *frame = frames[p1];
997 x264_slicetype_frame_cost( h, a, frames, p0, p1, p1, 0 );
999 int icost = frame->i_cost_est[0][0];
1000 int pcost = frame->i_cost_est[p1-p0][0];
1002 int i_gop_size = frame->i_frame - h->lookahead->i_last_keyframe;
1003 float f_thresh_max = h->param.i_scenecut_threshold / 100.0;
1004 /* magic numbers pulled out of thin air */
1005 float f_thresh_min = f_thresh_max * 0.25;
1008 if( h->param.i_keyint_min == h->param.i_keyint_max )
1009 f_thresh_min = f_thresh_max;
1010 if( i_gop_size <= h->param.i_keyint_min / 4 || h->param.b_intra_refresh )
1011 f_bias = f_thresh_min / 4;
1012 else if( i_gop_size <= h->param.i_keyint_min )
1013 f_bias = f_thresh_min * i_gop_size / h->param.i_keyint_min;
1016 f_bias = f_thresh_min
1017 + ( f_thresh_max - f_thresh_min )
1018 * ( i_gop_size - h->param.i_keyint_min )
1019 / ( h->param.i_keyint_max - h->param.i_keyint_min );
1022 res = pcost >= (1.0 - f_bias) * icost;
1025 int imb = frame->i_intra_mbs[p1-p0];
1026 int pmb = NUM_MBS - imb;
1027 x264_log( h, X264_LOG_DEBUG, "scene cut at %d Icost:%d Pcost:%d ratio:%.4f bias:%.4f gop:%d (imb:%d pmb:%d)\n",
1029 icost, pcost, 1. - (double)pcost / icost,
1030 f_bias, i_gop_size, imb, pmb );
1035 static int scenecut( x264_t *h, x264_mb_analysis_t *a, x264_frame_t **frames, int p0, int p1, int real_scenecut, int num_frames )
1037 /* Only do analysis during a normal scenecut check. */
1038 if( real_scenecut && h->param.i_bframe )
1041 /* Look ahead to avoid coding short flashes as scenecuts. */
1042 if( h->param.i_bframe_adaptive == X264_B_ADAPT_TRELLIS )
1043 /* Don't analyse any more frames than the trellis would have covered. */
1044 maxp1 += h->param.i_bframe;
1047 maxp1 = X264_MIN( maxp1, num_frames );
1049 /* Where A and B are scenes: AAAAAABBBAAAAAA
1050 * If BBB is shorter than (maxp1-p0), it is detected as a flash
1051 * and not considered a scenecut. */
1052 for( int curp1 = p1; curp1 <= maxp1; curp1++ )
1053 if( !scenecut_internal( h, a, frames, p0, curp1, 0 ) )
1054 /* Any frame in between p0 and cur_p1 cannot be a real scenecut. */
1055 for( int i = curp1; i > p0; i-- )
1056 frames[i]->b_scenecut = 0;
1058 /* Where A-F are scenes: AAAAABBCCDDEEFFFFFF
1059 * If each of BB ... EE are shorter than (maxp1-p0), they are
1060 * detected as flashes and not considered scenecuts.
1061 * Instead, the first F frame becomes a scenecut. */
1062 for( int curp0 = p0; curp0 < maxp1; curp0++ )
1063 if( scenecut_internal( h, a, frames, curp0, maxp1, 0 ) )
1064 /* If cur_p0 is the p0 of a scenecut, it cannot be the p1 of a scenecut. */
1065 frames[curp0]->b_scenecut = 0;
1068 /* Ignore frames that are part of a flash, i.e. cannot be real scenecuts. */
1069 if( !frames[p1]->b_scenecut )
1071 return scenecut_internal( h, a, frames, p0, p1, real_scenecut );
1074 void x264_slicetype_analyse( x264_t *h, int keyframe )
1076 x264_mb_analysis_t a;
1077 x264_frame_t *frames[X264_LOOKAHEAD_MAX+3] = { NULL, };
1078 int num_frames, orig_num_frames, keyint_limit, framecnt;
1079 int i_mb_count = NUM_MBS;
1080 int cost1p0, cost2p0, cost1b1, cost2p1;
1081 int i_max_search = X264_MIN( h->lookahead->next.i_size, X264_LOOKAHEAD_MAX );
1082 int vbv_lookahead = h->param.rc.i_vbv_buffer_size && h->param.rc.i_lookahead;
1083 if( h->param.b_deterministic )
1084 i_max_search = X264_MIN( i_max_search, h->lookahead->i_slicetype_length + !keyframe );
1086 assert( h->frames.b_have_lowres );
1088 if( !h->lookahead->last_nonb )
1090 frames[0] = h->lookahead->last_nonb;
1091 for( framecnt = 0; framecnt < i_max_search && h->lookahead->next.list[framecnt]->i_type == X264_TYPE_AUTO; framecnt++ )
1092 frames[framecnt+1] = h->lookahead->next.list[framecnt];
1096 if( h->param.rc.b_mb_tree )
1097 x264_macroblock_tree( h, &a, frames, 0, keyframe );
1101 keyint_limit = h->param.i_keyint_max - frames[0]->i_frame + h->lookahead->i_last_keyframe - 1;
1102 orig_num_frames = num_frames = h->param.b_intra_refresh ? framecnt : X264_MIN( framecnt, keyint_limit );
1104 x264_lowres_context_init( h, &a );
1106 /* This is important psy-wise: if we have a non-scenecut keyframe,
1107 * there will be significant visual artifacts if the frames just before
1108 * go down in quality due to being referenced less, despite it being
1109 * more RD-optimal. */
1110 if( (h->param.analyse.b_psy && h->param.rc.b_mb_tree) || vbv_lookahead )
1111 num_frames = framecnt;
1112 else if( num_frames == 0 )
1114 frames[1]->i_type = X264_TYPE_I;
1118 int num_bframes = 0;
1119 int num_analysed_frames = num_frames;
1121 if( h->param.i_scenecut_threshold && scenecut( h, &a, frames, 0, 1, 1, orig_num_frames ) )
1123 frames[1]->i_type = X264_TYPE_I;
1127 if( h->param.i_bframe )
1129 if( h->param.i_bframe_adaptive == X264_B_ADAPT_TRELLIS )
1131 if( num_frames > 1 )
1133 char best_paths[X264_BFRAME_MAX+1][X264_LOOKAHEAD_MAX] = {"","P"};
1134 int best_path_index = (num_frames-1) % (X264_BFRAME_MAX+1);
1136 /* Perform the frametype analysis. */
1137 for( int j = 2; j < num_frames; j++ )
1138 x264_slicetype_path( h, &a, frames, j, best_paths );
1140 num_bframes = strspn( best_paths[best_path_index], "B" );
1141 /* Load the results of the analysis into the frame types. */
1142 for( int j = 1; j < num_frames; j++ )
1143 frames[j]->i_type = best_paths[best_path_index][j-1] == 'B' ? X264_TYPE_B : X264_TYPE_P;
1145 frames[num_frames]->i_type = X264_TYPE_P;
1147 else if( h->param.i_bframe_adaptive == X264_B_ADAPT_FAST )
1149 for( int i = 0; i <= num_frames-2; )
1151 cost2p1 = x264_slicetype_frame_cost( h, &a, frames, i+0, i+2, i+2, 1 );
1152 if( frames[i+2]->i_intra_mbs[2] > i_mb_count / 2 )
1154 frames[i+1]->i_type = X264_TYPE_P;
1155 frames[i+2]->i_type = X264_TYPE_P;
1160 cost1b1 = x264_slicetype_frame_cost( h, &a, frames, i+0, i+2, i+1, 0 );
1161 cost1p0 = x264_slicetype_frame_cost( h, &a, frames, i+0, i+1, i+1, 0 );
1162 cost2p0 = x264_slicetype_frame_cost( h, &a, frames, i+1, i+2, i+2, 0 );
1164 if( cost1p0 + cost2p0 < cost1b1 + cost2p1 )
1166 frames[i+1]->i_type = X264_TYPE_P;
1171 // arbitrary and untuned
1172 #define INTER_THRESH 300
1173 #define P_SENS_BIAS (50 - h->param.i_bframe_bias)
1174 frames[i+1]->i_type = X264_TYPE_B;
1177 for( j = i+2; j <= X264_MIN( i+h->param.i_bframe, num_frames-1 ); j++ )
1179 int pthresh = X264_MAX(INTER_THRESH - P_SENS_BIAS * (j-i-1), INTER_THRESH/10);
1180 int pcost = x264_slicetype_frame_cost( h, &a, frames, i+0, j+1, j+1, 1 );
1181 if( pcost > pthresh*i_mb_count || frames[j+1]->i_intra_mbs[j-i+1] > i_mb_count/3 )
1183 frames[j]->i_type = X264_TYPE_B;
1185 frames[j]->i_type = X264_TYPE_P;
1188 frames[num_frames]->i_type = X264_TYPE_P;
1190 while( num_bframes < num_frames && frames[num_bframes+1]->i_type == X264_TYPE_B )
1195 num_bframes = X264_MIN(num_frames-1, h->param.i_bframe);
1196 for( int j = 1; j < num_frames; j++ )
1197 frames[j]->i_type = (j%(num_bframes+1)) ? X264_TYPE_B : X264_TYPE_P;
1198 frames[num_frames]->i_type = X264_TYPE_P;
1201 /* Check scenecut on the first minigop. */
1202 for( int j = 1; j < num_bframes+1; j++ )
1203 if( h->param.i_scenecut_threshold && scenecut( h, &a, frames, j, j+1, 0, orig_num_frames ) )
1205 frames[j]->i_type = X264_TYPE_P;
1206 num_analysed_frames = j;
1210 reset_start = keyframe ? 1 : X264_MIN( num_bframes+2, num_analysed_frames+1 );
1214 for( int j = 1; j <= num_frames; j++ )
1215 frames[j]->i_type = X264_TYPE_P;
1216 reset_start = !keyframe + 1;
1220 /* Perform the actual macroblock tree analysis.
1221 * Don't go farther than the maximum keyframe interval; this helps in short GOPs. */
1222 if( h->param.rc.b_mb_tree )
1223 x264_macroblock_tree( h, &a, frames, X264_MIN(num_frames, h->param.i_keyint_max), keyframe );
1225 /* Enforce keyframe limit. */
1226 if( !h->param.b_intra_refresh )
1227 for( int i = keyint_limit+1; i <= num_frames; i += h->param.i_keyint_max )
1229 frames[i]->i_type = X264_TYPE_I;
1230 reset_start = X264_MIN( reset_start, i+1 );
1231 if( h->param.i_open_gop == X264_OPEN_GOP_BLURAY )
1232 while( IS_X264_TYPE_B( frames[i-1]->i_type ) )
1237 x264_vbv_lookahead( h, &a, frames, num_frames, keyframe );
1239 /* Restore frametypes for all frames that haven't actually been decided yet. */
1240 for( int j = reset_start; j <= num_frames; j++ )
1241 frames[j]->i_type = X264_TYPE_AUTO;
1244 void x264_slicetype_decide( x264_t *h )
1246 x264_frame_t *frames[X264_BFRAME_MAX+2];
1251 if( !h->lookahead->next.i_size )
1254 int lookahead_size = h->lookahead->next.i_size;
1256 for( int i = 0; i < h->lookahead->next.i_size; i++ )
1258 if( h->param.b_vfr_input )
1260 if( lookahead_size-- > 1 )
1261 h->lookahead->next.list[i]->i_duration = 2 * (h->lookahead->next.list[i+1]->i_pts - h->lookahead->next.list[i]->i_pts);
1263 h->lookahead->next.list[i]->i_duration = h->i_prev_duration;
1266 h->lookahead->next.list[i]->i_duration = delta_tfi_divisor[h->lookahead->next.list[i]->i_pic_struct];
1267 h->i_prev_duration = h->lookahead->next.list[i]->i_duration;
1268 h->lookahead->next.list[i]->f_duration = (double)h->lookahead->next.list[i]->i_duration
1269 * h->sps->vui.i_num_units_in_tick
1270 / h->sps->vui.i_time_scale;
1272 if( h->lookahead->next.list[i]->i_frame > h->i_disp_fields_last_frame && lookahead_size > 0 )
1274 h->lookahead->next.list[i]->i_field_cnt = h->i_disp_fields;
1275 h->i_disp_fields += h->lookahead->next.list[i]->i_duration;
1276 h->i_disp_fields_last_frame = h->lookahead->next.list[i]->i_frame;
1278 else if( lookahead_size == 0 )
1280 h->lookahead->next.list[i]->i_field_cnt = h->i_disp_fields;
1281 h->lookahead->next.list[i]->i_duration = h->i_prev_duration;
1285 if( h->param.rc.b_stat_read )
1287 /* Use the frame types from the first pass */
1288 for( int i = 0; i < h->lookahead->next.i_size; i++ )
1289 h->lookahead->next.list[i]->i_type =
1290 x264_ratecontrol_slice_type( h, h->lookahead->next.list[i]->i_frame );
1292 else if( (h->param.i_bframe && h->param.i_bframe_adaptive)
1293 || h->param.i_scenecut_threshold
1294 || h->param.rc.b_mb_tree
1295 || (h->param.rc.i_vbv_buffer_size && h->param.rc.i_lookahead) )
1296 x264_slicetype_analyse( h, 0 );
1298 for( bframes = 0, brefs = 0;; bframes++ )
1300 frm = h->lookahead->next.list[bframes];
1301 if( frm->i_type == X264_TYPE_BREF && h->param.i_bframe_pyramid < X264_B_PYRAMID_NORMAL &&
1302 brefs == h->param.i_bframe_pyramid )
1304 frm->i_type = X264_TYPE_B;
1305 x264_log( h, X264_LOG_WARNING, "B-ref at frame %d incompatible with B-pyramid %s \n",
1306 frm->i_frame, x264_b_pyramid_names[h->param.i_bframe_pyramid] );
1308 /* pyramid with multiple B-refs needs a big enough dpb that the preceding P-frame stays available.
1309 smaller dpb could be supported by smart enough use of mmco, but it's easier just to forbid it. */
1310 else if( frm->i_type == X264_TYPE_BREF && h->param.i_bframe_pyramid == X264_B_PYRAMID_NORMAL &&
1311 brefs && h->param.i_frame_reference <= (brefs+3) )
1313 frm->i_type = X264_TYPE_B;
1314 x264_log( h, X264_LOG_WARNING, "B-ref at frame %d incompatible with B-pyramid %s and %d reference frames\n",
1315 frm->i_frame, x264_b_pyramid_names[h->param.i_bframe_pyramid], h->param.i_frame_reference );
1318 if( frm->i_type == X264_TYPE_KEYFRAME )
1319 frm->i_type = h->param.i_open_gop ? X264_TYPE_I : X264_TYPE_IDR;
1321 /* Limit GOP size */
1322 if( (!h->param.b_intra_refresh || frm->i_frame == 0) && frm->i_frame - h->lookahead->i_last_keyframe >= h->param.i_keyint_max )
1324 if( frm->i_type == X264_TYPE_AUTO || frm->i_type == X264_TYPE_I )
1325 frm->i_type = h->param.i_open_gop && h->lookahead->i_last_keyframe >= 0 ? X264_TYPE_I : X264_TYPE_IDR;
1326 int warn = frm->i_type != X264_TYPE_IDR;
1327 if( warn && h->param.i_open_gop )
1328 warn &= frm->i_type != X264_TYPE_I;
1330 x264_log( h, X264_LOG_WARNING, "specified frame type (%d) at %d is not compatible with keyframe interval\n", frm->i_type, frm->i_frame );
1332 if( frm->i_type == X264_TYPE_I && frm->i_frame - h->lookahead->i_last_keyframe >= h->param.i_keyint_min )
1334 if( h->param.i_open_gop )
1336 h->lookahead->i_last_keyframe = frm->i_frame; // Use display order
1337 if( h->param.i_open_gop == X264_OPEN_GOP_BLURAY )
1338 h->lookahead->i_last_keyframe -= bframes; // Use bluray order
1339 frm->b_keyframe = 1;
1342 frm->i_type = X264_TYPE_IDR;
1344 if( frm->i_type == X264_TYPE_IDR )
1347 h->lookahead->i_last_keyframe = frm->i_frame;
1348 frm->b_keyframe = 1;
1352 h->lookahead->next.list[bframes]->i_type = X264_TYPE_P;
1356 if( bframes == h->param.i_bframe ||
1357 !h->lookahead->next.list[bframes+1] )
1359 if( IS_X264_TYPE_B( frm->i_type ) )
1360 x264_log( h, X264_LOG_WARNING, "specified frame type is not compatible with max B-frames\n" );
1361 if( frm->i_type == X264_TYPE_AUTO
1362 || IS_X264_TYPE_B( frm->i_type ) )
1363 frm->i_type = X264_TYPE_P;
1366 if( frm->i_type == X264_TYPE_BREF )
1369 if( frm->i_type == X264_TYPE_AUTO )
1370 frm->i_type = X264_TYPE_B;
1372 else if( !IS_X264_TYPE_B( frm->i_type ) ) break;
1376 h->lookahead->next.list[bframes-1]->b_last_minigop_bframe = 1;
1377 h->lookahead->next.list[bframes]->i_bframes = bframes;
1379 /* insert a bref into the sequence */
1380 if( h->param.i_bframe_pyramid && bframes > 1 && !brefs )
1382 h->lookahead->next.list[bframes/2]->i_type = X264_TYPE_BREF;
1386 /* calculate the frame costs ahead of time for x264_rc_analyse_slice while we still have lowres */
1387 if( h->param.rc.i_rc_method != X264_RC_CQP )
1389 x264_mb_analysis_t a;
1391 p1 = b = bframes + 1;
1393 x264_lowres_context_init( h, &a );
1395 frames[0] = h->lookahead->last_nonb;
1396 memcpy( &frames[1], h->lookahead->next.list, (bframes+1) * sizeof(x264_frame_t*) );
1397 if( IS_X264_TYPE_I( h->lookahead->next.list[bframes]->i_type ) )
1402 x264_slicetype_frame_cost( h, &a, frames, p0, p1, b, 0 );
1404 if( (p0 != p1 || bframes) && h->param.rc.i_vbv_buffer_size )
1406 /* We need the intra costs for row SATDs. */
1407 x264_slicetype_frame_cost( h, &a, frames, b, b, b, 0 );
1409 /* We need B-frame costs for row SATDs. */
1411 for( b = 1; b <= bframes; b++ )
1413 if( frames[b]->i_type == X264_TYPE_B )
1414 for( p1 = b; frames[p1]->i_type == X264_TYPE_B; )
1418 x264_slicetype_frame_cost( h, &a, frames, p0, p1, b, 0 );
1419 if( frames[b]->i_type == X264_TYPE_BREF )
1425 /* Analyse for weighted P frames */
1426 if( !h->param.rc.b_stat_read && h->lookahead->next.list[bframes]->i_type == X264_TYPE_P
1427 && h->param.analyse.i_weighted_pred == X264_WEIGHTP_SMART )
1430 x264_weights_analyse( h, h->lookahead->next.list[bframes], h->lookahead->last_nonb, 0 );
1433 /* shift sequence to coded order.
1434 use a small temporary list to avoid shifting the entire next buffer around */
1435 int i_coded = h->lookahead->next.list[0]->i_frame;
1438 int idx_list[] = { brefs+1, 1 };
1439 for( int i = 0; i < bframes; i++ )
1441 int idx = idx_list[h->lookahead->next.list[i]->i_type == X264_TYPE_BREF]++;
1442 frames[idx] = h->lookahead->next.list[i];
1443 frames[idx]->i_reordered_pts = h->lookahead->next.list[idx]->i_pts;
1445 frames[0] = h->lookahead->next.list[bframes];
1446 frames[0]->i_reordered_pts = h->lookahead->next.list[0]->i_pts;
1447 memcpy( h->lookahead->next.list, frames, (bframes+1) * sizeof(x264_frame_t*) );
1450 for( int i = 0; i <= bframes; i++ )
1452 h->lookahead->next.list[i]->i_coded = i_coded++;
1455 x264_calculate_durations( h, h->lookahead->next.list[i], h->lookahead->next.list[i-1], &h->i_cpb_delay, &h->i_coded_fields );
1456 h->lookahead->next.list[0]->f_planned_cpb_duration[i-1] = (double)h->lookahead->next.list[i-1]->i_cpb_duration *
1457 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1460 x264_calculate_durations( h, h->lookahead->next.list[i], NULL, &h->i_cpb_delay, &h->i_coded_fields );
1462 h->lookahead->next.list[0]->f_planned_cpb_duration[i] = (double)h->lookahead->next.list[i]->i_cpb_duration *
1463 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1467 int x264_rc_analyse_slice( x264_t *h )
1473 if( IS_X264_TYPE_I(h->fenc->i_type) )
1475 else if( h->fenc->i_type == X264_TYPE_P )
1476 p1 = b = h->fenc->i_bframes + 1;
1479 p1 = (h->fref1[0]->i_poc - h->fref0[0]->i_poc)/2;
1480 b = (h->fenc->i_poc - h->fref0[0]->i_poc)/2;
1482 /* We don't need to assign p0/p1 since we are not performing any real analysis here. */
1483 x264_frame_t **frames = &h->fenc - b;
1485 /* cost should have been already calculated by x264_slicetype_decide */
1486 cost = frames[b]->i_cost_est[b-p0][p1-b];
1487 assert( cost >= 0 );
1489 if( h->param.rc.b_mb_tree && !h->param.rc.b_stat_read )
1491 cost = x264_slicetype_frame_cost_recalculate( h, frames, p0, p1, b );
1492 if( b && h->param.rc.i_vbv_buffer_size )
1493 x264_slicetype_frame_cost_recalculate( h, frames, b, b, b );
1495 /* In AQ, use the weighted score instead. */
1496 else if( h->param.rc.i_aq_mode )
1497 cost = frames[b]->i_cost_est_aq[b-p0][p1-b];
1499 h->fenc->i_row_satd = h->fenc->i_row_satds[b-p0][p1-b];
1500 h->fdec->i_row_satd = h->fdec->i_row_satds[b-p0][p1-b];
1501 h->fdec->i_satd = cost;
1502 memcpy( h->fdec->i_row_satd, h->fenc->i_row_satd, h->mb.i_mb_height * sizeof(int) );
1503 if( !IS_X264_TYPE_I(h->fenc->i_type) )
1504 memcpy( h->fdec->i_row_satds[0][0], h->fenc->i_row_satds[0][0], h->mb.i_mb_height * sizeof(int) );
1506 if( h->param.b_intra_refresh && h->param.rc.i_vbv_buffer_size && h->fenc->i_type == X264_TYPE_P )
1508 int ip_factor = 256 * h->param.rc.f_ip_factor; /* fix8 */
1509 for( int y = 0; y < h->mb.i_mb_height; y++ )
1511 int mb_xy = y * h->mb.i_mb_stride;
1512 for( int x = h->fdec->i_pir_start_col; x <= h->fdec->i_pir_end_col; x++, mb_xy++ )
1514 int intra_cost = (h->fenc->i_intra_cost[mb_xy] * ip_factor + 128) >> 8;
1515 int inter_cost = h->fenc->lowres_costs[b-p0][p1-b][mb_xy] & LOWRES_COST_MASK;
1516 int diff = intra_cost - inter_cost;
1517 if( h->param.rc.i_aq_mode )
1518 h->fdec->i_row_satd[y] += (diff * frames[b]->i_inv_qscale_factor[mb_xy] + 128) >> 8;
1520 h->fdec->i_row_satd[y] += diff;