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( 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 /* How data is organized for chroma weightp:
104 * fenc = ref + offset
105 * v = u + stride * chroma height
106 * We'll need more room if we do 4:2:2 or 4:4:4. */
108 static NOINLINE void x264_weight_cost_init_chroma( x264_t *h, x264_frame_t *fenc, x264_frame_t *ref, pixel *dstu, pixel *dstv )
110 int ref0_distance = fenc->i_frame - ref->i_frame - 1;
111 int i_stride = fenc->i_stride[1];
112 int i_offset = i_stride / 2;
113 int i_lines = fenc->i_lines[1];
114 int i_width = fenc->i_width[1];
115 int cw = h->mb.i_mb_width << 3;
116 int ch = h->mb.i_mb_height << 3;
118 if( fenc->lowres_mvs[0][ref0_distance][0][0] != 0x7FFF )
120 for( int y = 0, mb_xy = 0, pel_offset_y = 0; y < i_lines; y += 8, pel_offset_y = y*i_stride )
121 for( int x = 0, pel_offset_x = 0; x < i_width; x += 8, mb_xy++, pel_offset_x += 8 )
123 pixel *pixu = dstu + pel_offset_y + pel_offset_x;
124 pixel *pixv = dstv + pel_offset_y + pel_offset_x;
125 pixel *src1 = ref->plane[1] + pel_offset_y + pel_offset_x*2; /* NV12 */
126 int mvx = fenc->lowres_mvs[0][ref0_distance][mb_xy][0];
127 int mvy = fenc->lowres_mvs[0][ref0_distance][mb_xy][1];
128 h->mc.mc_chroma( pixu, pixv, i_stride, src1, i_stride, mvx, mvy, 8, 8 );
132 h->mc.plane_copy_deinterleave( dstu, i_stride, dstv, i_stride, ref->plane[1], i_stride, cw, ch );
133 h->mc.plane_copy_deinterleave( dstu+i_offset, i_stride, dstv+i_offset, i_stride, fenc->plane[1], i_stride, cw, ch );
137 static int x264_weight_slice_header_cost( x264_t *h, x264_weight_t *w, int b_chroma )
139 /* Add cost of weights in the slice header. */
140 int lambda = x264_lambda_tab[X264_LOOKAHEAD_QP];
142 if( h->param.i_slice_count )
143 numslices = h->param.i_slice_count;
144 else if( h->param.i_slice_max_mbs )
145 numslices = (h->mb.i_mb_width * h->mb.i_mb_height + h->param.i_slice_max_mbs-1) / h->param.i_slice_max_mbs;
148 /* FIXME: find a way to account for --slice-max-size?
149 * 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.
150 * Cut denom cost in half if chroma, since it's shared between the two chroma planes. */
151 int denom_cost = bs_size_ue( w[0].i_denom ) * (2 - b_chroma);
152 return lambda * numslices * ( 10 + denom_cost + 2 * (bs_size_se( w[0].i_scale ) + bs_size_se( w[0].i_offset )) );
155 static NOINLINE unsigned int x264_weight_cost_luma( x264_t *h, x264_frame_t *fenc, pixel *src, x264_weight_t *w )
157 unsigned int cost = 0;
158 int i_stride = fenc->i_stride_lowres;
159 int i_lines = fenc->i_lines_lowres;
160 int i_width = fenc->i_width_lowres;
161 pixel *fenc_plane = fenc->lowres[0];
162 ALIGNED_ARRAY_16( pixel, buf,[8*8] );
168 for( int y = 0; y < i_lines; y += 8, pixoff = y*i_stride )
169 for( int x = 0; x < i_width; x += 8, i_mb++, pixoff += 8)
171 w->weightfn[8>>2]( buf, 8, &src[pixoff], i_stride, w, 8 );
172 cost += X264_MIN( h->pixf.mbcmp[PIXEL_8x8]( buf, 8, &fenc_plane[pixoff], i_stride ), fenc->i_intra_cost[i_mb] );
174 cost += x264_weight_slice_header_cost( h, w, 0 );
177 for( int y = 0; y < i_lines; y += 8, pixoff = y*i_stride )
178 for( int x = 0; x < i_width; x += 8, i_mb++, pixoff += 8 )
179 cost += X264_MIN( h->pixf.mbcmp[PIXEL_8x8]( &src[pixoff], i_stride, &fenc_plane[pixoff], i_stride ), fenc->i_intra_cost[i_mb] );
184 static NOINLINE unsigned int x264_weight_cost_chroma( x264_t *h, x264_frame_t *fenc, pixel *ref, x264_weight_t *w )
186 unsigned int cost = 0;
187 int i_stride = fenc->i_stride[1];
188 int i_offset = i_stride / 2;
189 int i_lines = fenc->i_lines[1];
190 int i_width = fenc->i_width[1];
191 pixel *src = ref + i_offset;
192 ALIGNED_ARRAY_16( pixel, buf, [8*8] );
194 ALIGNED_16( static pixel flat[8] ) = {0};
197 for( int y = 0; y < i_lines; y += 8, pixoff = y*i_stride )
198 for( int x = 0; x < i_width; x += 8, pixoff += 8 )
200 w->weightfn[8>>2]( buf, 8, &ref[pixoff], i_stride, w, 8 );
201 /* The naive and seemingly sensible algorithm is to use mbcmp as in luma.
202 * But testing shows that for chroma the DC coefficient is by far the most
203 * important part of the coding cost. Thus a more useful chroma weight is
204 * obtained by comparing each block's DC coefficient instead of the actual
207 * FIXME: add a (faster) asm sum function to replace sad. */
208 cost += abs( h->pixf.sad_aligned[PIXEL_8x8]( buf, 8, flat, 0 ) -
209 h->pixf.sad_aligned[PIXEL_8x8]( &src[pixoff], i_stride, flat, 0 ) );
211 cost += x264_weight_slice_header_cost( h, w, 1 );
214 for( int y = 0; y < i_lines; y += 8, pixoff = y*i_stride )
215 for( int x = 0; x < i_width; x += 8, pixoff += 8 )
216 cost += abs( h->pixf.sad_aligned[PIXEL_8x8]( &ref[pixoff], i_stride, flat, 0 ) -
217 h->pixf.sad_aligned[PIXEL_8x8]( &src[pixoff], i_stride, flat, 0 ) );
222 void x264_weights_analyse( x264_t *h, x264_frame_t *fenc, x264_frame_t *ref, int b_lookahead )
224 int i_delta_index = fenc->i_frame - ref->i_frame - 1;
225 /* epsilon is chosen to require at least a numerator of 127 (with denominator = 128) */
226 const float epsilon = 1.f/128.f;
227 x264_weight_t *weights = fenc->weight[0];
228 SET_WEIGHT( weights[0], 0, 1, 0, 0 );
229 SET_WEIGHT( weights[1], 0, 1, 0, 0 );
230 SET_WEIGHT( weights[2], 0, 1, 0, 0 );
231 /* Don't check chroma in lookahead, or if there wasn't a luma weight. */
232 for( int plane = 0; plane <= 2 && !( plane && ( !weights[0].weightfn || b_lookahead ) ); plane++ )
234 int cur_offset, start_offset, end_offset;
235 int minoff, minscale, mindenom;
236 unsigned int minscore, origscore;
238 float fenc_var = fenc->i_pixel_ssd[plane] + !ref->i_pixel_ssd[plane];
239 float ref_var = ref->i_pixel_ssd[plane] + !ref->i_pixel_ssd[plane];
240 float guess_scale = sqrtf( fenc_var / ref_var );
241 float fenc_mean = (float)fenc->i_pixel_sum[plane] / (fenc->i_lines[!!plane] * fenc->i_width[!!plane]) / (1 << (BIT_DEPTH - 8));
242 float ref_mean = (float) ref->i_pixel_sum[plane] / (fenc->i_lines[!!plane] * fenc->i_width[!!plane]) / (1 << (BIT_DEPTH - 8));
245 if( fabsf( ref_mean - fenc_mean ) < 0.5f && fabsf( 1.f - guess_scale ) < epsilon )
247 SET_WEIGHT( weights[plane], 0, 1, 0, 0 );
253 weights[plane].i_denom = 6;
254 weights[plane].i_scale = x264_clip3( round( guess_scale * 64 ), 0, 255 );
255 if( weights[plane].i_scale > 127 )
257 weights[1].weightfn = weights[2].weightfn = NULL;
262 x264_weight_get_h264( round( guess_scale * 128 ), 0, &weights[plane] );
265 mindenom = weights[plane].i_denom;
266 minscale = weights[plane].i_scale;
272 if( !fenc->b_intra_calculated )
274 x264_mb_analysis_t a;
275 x264_lowres_context_init( h, &a );
276 x264_slicetype_frame_cost( h, &a, &fenc, 0, 0, 0, 0 );
278 mcbuf = x264_weight_cost_init_luma( h, fenc, ref, h->mb.p_weight_buf[0] );
279 origscore = minscore = x264_weight_cost_luma( h, fenc, mcbuf, NULL );
283 pixel *dstu = h->mb.p_weight_buf[0];
284 pixel *dstv = h->mb.p_weight_buf[0]+fenc->i_stride[1]*fenc->i_lines[1];
285 /* Only initialize chroma data once. */
287 x264_weight_cost_init_chroma( h, fenc, ref, dstu, dstv );
288 mcbuf = plane == 1 ? dstu : dstv;
289 origscore = minscore = x264_weight_cost_chroma( h, fenc, mcbuf, NULL );
295 // This gives a slight improvement due to rounding errors but only tests one offset in lookahead.
296 // Currently only searches within +/- 1 of the best offset found so far.
297 // TODO: Try other offsets/multipliers/combinations thereof?
298 cur_offset = fenc_mean - ref_mean * minscale / (1 << mindenom) + 0.5f * b_lookahead;
299 start_offset = x264_clip3( cur_offset - !b_lookahead, -128, 127 );
300 end_offset = x264_clip3( cur_offset + !b_lookahead, -128, 127 );
301 for( int i_off = start_offset; i_off <= end_offset; i_off++ )
303 SET_WEIGHT( weights[plane], 1, minscale, mindenom, i_off );
306 s = x264_weight_cost_chroma( h, fenc, mcbuf, &weights[plane] );
308 s = x264_weight_cost_luma( h, fenc, mcbuf, &weights[plane] );
309 COPY3_IF_LT( minscore, s, minoff, i_off, found, 1 );
311 // Don't check any more offsets if the previous one had a lower cost than the current one
312 if( minoff == start_offset && i_off != start_offset )
317 /* FIXME: More analysis can be done here on SAD vs. SATD termination. */
318 /* 0.2% termination derived experimentally to avoid weird weights in frames that are mostly intra. */
319 if( !found || (minscale == 1 << mindenom && minoff == 0) || (float)minscore / origscore > 0.998f )
321 SET_WEIGHT( weights[plane], 0, 1, 0, 0 );
325 SET_WEIGHT( weights[plane], 1, minscale, mindenom, minoff );
327 if( h->param.analyse.i_weighted_pred == X264_WEIGHTP_FAKE && weights[0].weightfn && !plane )
328 fenc->f_weighted_cost_delta[i_delta_index] = (float)minscore / origscore;
331 //FIXME, what is the correct way to deal with this?
332 if( weights[1].weightfn && weights[2].weightfn && weights[1].i_denom != weights[2].i_denom )
334 int denom = X264_MIN( weights[1].i_denom, weights[2].i_denom );
336 for( i = 1; i <= 2; i++ )
338 weights[i].i_scale = x264_clip3( weights[i].i_scale >> ( weights[i].i_denom - denom ), 0, 255 );
339 weights[i].i_denom = denom;
340 h->mc.weight_cache( h, &weights[i] );
344 if( weights[0].weightfn && b_lookahead )
346 //scale lowres in lookahead for slicetype_frame_cost
347 pixel *src = ref->buffer_lowres[0];
348 pixel *dst = h->mb.p_weight_buf[0];
349 int width = ref->i_width_lowres + PADH*2;
350 int height = ref->i_lines_lowres + PADV*2;
351 x264_weight_scale_plane( h, dst, ref->i_stride_lowres, src, ref->i_stride_lowres,
352 width, height, &weights[0] );
353 fenc->weighted[0] = h->mb.p_weight_buf[0] + PADH + ref->i_stride_lowres * PADV;
357 static void x264_slicetype_mb_cost( x264_t *h, x264_mb_analysis_t *a,
358 x264_frame_t **frames, int p0, int p1, int b,
359 int dist_scale_factor, int do_search[2], const x264_weight_t *w )
361 x264_frame_t *fref0 = frames[p0];
362 x264_frame_t *fref1 = frames[p1];
363 x264_frame_t *fenc = frames[b];
364 const int b_bidir = (b < p1);
365 const int i_mb_x = h->mb.i_mb_x;
366 const int i_mb_y = h->mb.i_mb_y;
367 const int i_mb_stride = h->mb.i_mb_width;
368 const int i_mb_xy = i_mb_x + i_mb_y * i_mb_stride;
369 const int i_stride = fenc->i_stride_lowres;
370 const int i_pel_offset = 8 * (i_mb_x + i_mb_y * i_stride);
371 const int i_bipred_weight = h->param.analyse.b_weighted_bipred ? 64 - (dist_scale_factor>>2) : 32;
372 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] };
373 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] };
374 int b_frame_score_mb = (i_mb_x > 0 && i_mb_x < h->mb.i_mb_width - 1 &&
375 i_mb_y > 0 && i_mb_y < h->mb.i_mb_height - 1) ||
376 h->mb.i_mb_width <= 2 || h->mb.i_mb_height <= 2;
378 ALIGNED_ARRAY_16( pixel, pix1,[9*FDEC_STRIDE] );
379 pixel *pix2 = pix1+8;
381 int i_bcost = COST_MAX;
384 h->mb.pic.p_fenc[0] = h->mb.pic.fenc_buf;
385 h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fenc[0], FENC_STRIDE, &fenc->lowres[0][i_pel_offset], i_stride, 8 );
388 goto lowres_intra_mb;
390 // no need for h->mb.mv_min[]
391 h->mb.mv_min_fpel[0] = -8*h->mb.i_mb_x - 4;
392 h->mb.mv_max_fpel[0] = 8*( h->mb.i_mb_width - h->mb.i_mb_x - 1 ) + 4;
393 h->mb.mv_min_spel[0] = 4*( h->mb.mv_min_fpel[0] - 8 );
394 h->mb.mv_max_spel[0] = 4*( h->mb.mv_max_fpel[0] + 8 );
395 if( h->mb.i_mb_x >= h->mb.i_mb_width - 2 )
397 h->mb.mv_min_fpel[1] = -8*h->mb.i_mb_y - 4;
398 h->mb.mv_max_fpel[1] = 8*( h->mb.i_mb_height - h->mb.i_mb_y - 1 ) + 4;
399 h->mb.mv_min_spel[1] = 4*( h->mb.mv_min_fpel[1] - 8 );
400 h->mb.mv_max_spel[1] = 4*( h->mb.mv_max_fpel[1] + 8 );
403 #define LOAD_HPELS_LUMA(dst, src) \
405 (dst)[0] = &(src)[0][i_pel_offset]; \
406 (dst)[1] = &(src)[1][i_pel_offset]; \
407 (dst)[2] = &(src)[2][i_pel_offset]; \
408 (dst)[3] = &(src)[3][i_pel_offset]; \
410 #define LOAD_WPELS_LUMA(dst,src) \
411 (dst) = &(src)[i_pel_offset];
413 #define CLIP_MV( mv ) \
415 mv[0] = x264_clip3( mv[0], h->mb.mv_min_spel[0], h->mb.mv_max_spel[0] ); \
416 mv[1] = x264_clip3( mv[1], h->mb.mv_min_spel[1], h->mb.mv_max_spel[1] ); \
418 #define TRY_BIDIR( mv0, mv1, penalty ) \
421 if( h->param.analyse.i_subpel_refine <= 1 ) \
423 int hpel_idx1 = (((mv0)[0]&2)>>1) + ((mv0)[1]&2); \
424 int hpel_idx2 = (((mv1)[0]&2)>>1) + ((mv1)[1]&2); \
425 pixel *src1 = m[0].p_fref[hpel_idx1] + ((mv0)[0]>>2) + ((mv0)[1]>>2) * m[0].i_stride[0]; \
426 pixel *src2 = m[1].p_fref[hpel_idx2] + ((mv1)[0]>>2) + ((mv1)[1]>>2) * m[1].i_stride[0]; \
427 h->mc.avg[PIXEL_8x8]( pix1, 16, src1, m[0].i_stride[0], src2, m[1].i_stride[0], i_bipred_weight ); \
431 int stride1 = 16, stride2 = 16; \
432 pixel *src1, *src2; \
433 src1 = h->mc.get_ref( pix1, &stride1, m[0].p_fref, m[0].i_stride[0], \
434 (mv0)[0], (mv0)[1], 8, 8, w ); \
435 src2 = h->mc.get_ref( pix2, &stride2, m[1].p_fref, m[1].i_stride[0], \
436 (mv1)[0], (mv1)[1], 8, 8, w ); \
437 h->mc.avg[PIXEL_8x8]( pix1, 16, src1, stride1, src2, stride2, i_bipred_weight ); \
439 i_cost = penalty * a->i_lambda + h->pixf.mbcmp[PIXEL_8x8]( \
440 m[0].p_fenc[0], FENC_STRIDE, pix1, 16 ); \
441 COPY2_IF_LT( i_bcost, i_cost, list_used, 3 ); \
444 m[0].i_pixel = PIXEL_8x8;
445 m[0].p_cost_mv = a->p_cost_mv;
446 m[0].i_stride[0] = i_stride;
447 m[0].p_fenc[0] = h->mb.pic.p_fenc[0];
450 LOAD_HPELS_LUMA( m[0].p_fref, fref0->lowres );
451 m[0].p_fref_w = m[0].p_fref[0];
453 LOAD_WPELS_LUMA( m[0].p_fref_w, fenc->weighted[0] );
457 int16_t *mvr = fref1->lowres_mvs[0][p1-p0-1][i_mb_xy];
458 ALIGNED_ARRAY_8( int16_t, dmv,[2],[2] );
460 m[1].i_pixel = PIXEL_8x8;
461 m[1].p_cost_mv = a->p_cost_mv;
462 m[1].i_stride[0] = i_stride;
463 m[1].p_fenc[0] = h->mb.pic.p_fenc[0];
465 m[1].weight = weight_none;
466 LOAD_HPELS_LUMA( m[1].p_fref, fref1->lowres );
467 m[1].p_fref_w = m[1].p_fref[0];
469 dmv[0][0] = ( mvr[0] * dist_scale_factor + 128 ) >> 8;
470 dmv[0][1] = ( mvr[1] * dist_scale_factor + 128 ) >> 8;
471 dmv[1][0] = dmv[0][0] - mvr[0];
472 dmv[1][1] = dmv[0][1] - mvr[1];
475 if( h->param.analyse.i_subpel_refine <= 1 )
476 M64( dmv ) &= ~0x0001000100010001ULL; /* mv & ~1 */
478 TRY_BIDIR( dmv[0], dmv[1], 0 );
482 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 );
483 i_cost = h->pixf.mbcmp[PIXEL_8x8]( m[0].p_fenc[0], FENC_STRIDE, pix1, 16 );
484 COPY2_IF_LT( i_bcost, i_cost, list_used, 3 );
488 for( int l = 0; l < 1 + b_bidir; l++ )
493 int16_t (*fenc_mv)[2] = fenc_mvs[l];
494 ALIGNED_4( int16_t mvc[4][2] );
496 /* Reverse-order MV prediction. */
499 #define MVC(mv) { CP32( mvc[i_mvc], mv ); i_mvc++; }
500 if( i_mb_x < h->mb.i_mb_width - 1 )
502 if( i_mb_y < h->mb.i_mb_height - 1 )
504 MVC( fenc_mv[i_mb_stride] );
506 MVC( fenc_mv[i_mb_stride-1] );
507 if( i_mb_x < h->mb.i_mb_width - 1 )
508 MVC( fenc_mv[i_mb_stride+1] );
512 CP32( m[l].mvp, mvc[0] );
514 x264_median_mv( m[l].mvp, mvc[0], mvc[1], mvc[2] );
516 /* Fast skip for cases of near-zero residual. Shortcut: don't bother except in the mv0 case,
517 * since anything else is likely to have enough residual to not trigger the skip. */
518 if( !M32( m[l].mvp ) )
520 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] );
528 x264_me_search( h, &m[l], mvc, i_mvc );
529 m[l].cost -= a->p_cost_mv[0]; // remove mvcost from skip mbs
531 m[l].cost += 5 * a->i_lambda;
534 CP32( fenc_mvs[l], m[l].mv );
535 *fenc_costs[l] = m[l].cost;
539 CP32( m[l].mv, fenc_mvs[l] );
540 m[l].cost = *fenc_costs[l];
542 COPY2_IF_LT( i_bcost, m[l].cost, list_used, l+1 );
545 if( b_bidir && ( M32( m[0].mv ) || M32( m[1].mv ) ) )
546 TRY_BIDIR( m[0].mv, m[1].mv, 5 );
549 if( !fenc->b_intra_calculated )
551 ALIGNED_ARRAY_16( pixel, edge,[33] );
552 pixel *pix = &pix1[8+FDEC_STRIDE - 1];
553 pixel *src = &fenc->lowres[0][i_pel_offset - 1];
554 const int intra_penalty = 5 * a->i_lambda;
557 memcpy( pix-FDEC_STRIDE, src-i_stride, 17 * sizeof(pixel) );
558 for( int i = 0; i < 8; i++ )
559 pix[i*FDEC_STRIDE] = src[i*i_stride];
562 h->pixf.intra_mbcmp_x3_8x8c( h->mb.pic.p_fenc[0], pix, satds );
563 int i_icost = X264_MIN3( satds[0], satds[1], satds[2] );
565 if( h->param.analyse.i_subpel_refine > 1 )
567 h->predict_8x8c[I_PRED_CHROMA_P]( pix );
568 int satd = h->pixf.mbcmp[PIXEL_8x8]( pix, FDEC_STRIDE, h->mb.pic.p_fenc[0], FENC_STRIDE );
569 i_icost = X264_MIN( i_icost, satd );
570 h->predict_8x8_filter( pix, edge, ALL_NEIGHBORS, ALL_NEIGHBORS );
571 for( int i = 3; i < 9; i++ )
573 h->predict_8x8[i]( pix, edge );
574 satd = h->pixf.mbcmp[PIXEL_8x8]( pix, FDEC_STRIDE, h->mb.pic.p_fenc[0], FENC_STRIDE );
575 i_icost = X264_MIN( i_icost, satd );
579 i_icost += intra_penalty;
580 fenc->i_intra_cost[i_mb_xy] = i_icost;
581 if( b_frame_score_mb )
583 int *row_satd_intra = frames[b]->i_row_satds[0][0];
584 int i_icost_aq = i_icost;
585 if( h->param.rc.i_aq_mode )
586 i_icost_aq = (i_icost_aq * frames[b]->i_inv_qscale_factor[i_mb_xy] + 128) >> 8;
587 fenc->i_cost_est[0][0] += i_icost;
588 fenc->i_cost_est_aq[0][0] += i_icost_aq;
589 row_satd_intra[h->mb.i_mb_y] += i_icost_aq;
593 /* forbid intra-mbs in B-frames, because it's rare and not worth checking */
594 /* FIXME: Should we still forbid them now that we cache intra scores? */
597 int i_icost = fenc->i_intra_cost[i_mb_xy];
598 int b_intra = i_icost < i_bcost;
604 if( b_frame_score_mb )
605 fenc->i_intra_mbs[b-p0] += b_intra;
608 /* In an I-frame, we've already added the results above in the intra section. */
611 int i_bcost_aq = i_bcost;
612 if( h->param.rc.i_aq_mode )
613 i_bcost_aq = (i_bcost_aq * frames[b]->i_inv_qscale_factor[i_mb_xy] + 128) >> 8;
614 fenc->i_row_satds[b-p0][p1-b][h->mb.i_mb_y] += i_bcost_aq;
615 if( b_frame_score_mb )
617 /* Don't use AQ-weighted costs for slicetype decision, only for ratecontrol. */
618 frames[b]->i_cost_est[b-p0][p1-b] += i_bcost;
619 frames[b]->i_cost_est_aq[b-p0][p1-b] += i_bcost_aq;
623 fenc->lowres_costs[b-p0][p1-b][i_mb_xy] = X264_MIN( i_bcost, LOWRES_COST_MASK ) + (list_used << LOWRES_COST_SHIFT);
628 (h->mb.i_mb_width > 2 && h->mb.i_mb_height > 2 ?\
629 (h->mb.i_mb_width - 2) * (h->mb.i_mb_height - 2) :\
630 h->mb.i_mb_width * h->mb.i_mb_height)
632 static int x264_slicetype_frame_cost( x264_t *h, x264_mb_analysis_t *a,
633 x264_frame_t **frames, int p0, int p1, int b,
634 int b_intra_penalty )
638 const x264_weight_t *w = weight_none;
639 /* Check whether we already evaluated this frame
640 * If we have tried this frame as P, then we have also tried
641 * the preceding frames as B. (is this still true?) */
642 /* Also check that we already calculated the row SATDs for the current frame. */
643 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) )
644 i_score = frames[b]->i_cost_est[b-p0][p1-b];
647 int dist_scale_factor = 128;
648 int *row_satd = frames[b]->i_row_satds[b-p0][p1-b];
649 int *row_satd_intra = frames[b]->i_row_satds[0][0];
651 /* For each list, check to see whether we have lowres motion-searched this reference frame before. */
652 do_search[0] = b != p0 && frames[b]->lowres_mvs[0][b-p0-1][0][0] == 0x7FFF;
653 do_search[1] = b != p1 && frames[b]->lowres_mvs[1][p1-b-1][0][0] == 0x7FFF;
656 if( h->param.analyse.i_weighted_pred && b == p1 )
659 x264_weights_analyse( h, frames[b], frames[p0], 1 );
660 w = frames[b]->weight[0];
662 frames[b]->lowres_mvs[0][b-p0-1][0][0] = 0;
664 if( do_search[1] ) frames[b]->lowres_mvs[1][p1-b-1][0][0] = 0;
667 frames[b]->i_intra_mbs[b-p0] = 0;
668 if( !frames[b]->b_intra_calculated )
670 frames[b]->i_cost_est[0][0] = 0;
671 frames[b]->i_cost_est_aq[0][0] = 0;
674 dist_scale_factor = ( ((b-p0) << 8) + ((p1-p0) >> 1) ) / (p1-p0);
676 frames[b]->i_cost_est[b-p0][p1-b] = 0;
677 frames[b]->i_cost_est_aq[b-p0][p1-b] = 0;
679 /* Lowres lookahead goes backwards because the MVs are used as predictors in the main encode.
680 * This considerably improves MV prediction overall. */
682 /* The edge mbs seem to reduce the predictive quality of the
683 * whole frame's score, but are needed for a spatial distribution. */
684 if( h->param.rc.b_mb_tree || h->param.rc.i_vbv_buffer_size ||
685 h->mb.i_mb_width <= 2 || h->mb.i_mb_height <= 2 )
687 for( h->mb.i_mb_y = h->mb.i_mb_height - 1; h->mb.i_mb_y >= 0; h->mb.i_mb_y-- )
689 row_satd[h->mb.i_mb_y] = 0;
690 if( !frames[b]->b_intra_calculated )
691 row_satd_intra[h->mb.i_mb_y] = 0;
692 for( h->mb.i_mb_x = h->mb.i_mb_width - 1; h->mb.i_mb_x >= 0; h->mb.i_mb_x-- )
693 x264_slicetype_mb_cost( h, a, frames, p0, p1, b, dist_scale_factor, do_search, w );
698 for( h->mb.i_mb_y = h->mb.i_mb_height - 2; h->mb.i_mb_y >= 1; h->mb.i_mb_y-- )
699 for( h->mb.i_mb_x = h->mb.i_mb_width - 2; h->mb.i_mb_x >= 1; h->mb.i_mb_x-- )
700 x264_slicetype_mb_cost( h, a, frames, p0, p1, b, dist_scale_factor, do_search, w );
703 i_score = frames[b]->i_cost_est[b-p0][p1-b];
705 i_score = (uint64_t)i_score * 100 / (120 + h->param.i_bframe_bias);
707 frames[b]->b_intra_calculated = 1;
709 frames[b]->i_cost_est[b-p0][p1-b] = i_score;
713 if( b_intra_penalty )
715 // arbitrary penalty for I-blocks after B-frames
717 i_score += i_score * frames[b]->i_intra_mbs[b-p0] / (nmb * 8);
722 /* If MB-tree changes the quantizers, we need to recalculate the frame cost without
723 * re-running lookahead. */
724 static int x264_slicetype_frame_cost_recalculate( x264_t *h, x264_frame_t **frames, int p0, int p1, int b )
727 int *row_satd = frames[b]->i_row_satds[b-p0][p1-b];
728 float *qp_offset = IS_X264_TYPE_B(frames[b]->i_type) ? frames[b]->f_qp_offset_aq : frames[b]->f_qp_offset;
730 for( h->mb.i_mb_y = h->mb.i_mb_height - 1; h->mb.i_mb_y >= 0; h->mb.i_mb_y-- )
732 row_satd[ h->mb.i_mb_y ] = 0;
733 for( h->mb.i_mb_x = h->mb.i_mb_width - 1; h->mb.i_mb_x >= 0; h->mb.i_mb_x-- )
735 int i_mb_xy = h->mb.i_mb_x + h->mb.i_mb_y*h->mb.i_mb_stride;
736 int i_mb_cost = frames[b]->lowres_costs[b-p0][p1-b][i_mb_xy] & LOWRES_COST_MASK;
737 float qp_adj = qp_offset[i_mb_xy];
738 i_mb_cost = (i_mb_cost * x264_exp2fix8(qp_adj) + 128) >> 8;
739 row_satd[ h->mb.i_mb_y ] += i_mb_cost;
740 if( (h->mb.i_mb_y > 0 && h->mb.i_mb_y < h->mb.i_mb_height - 1 &&
741 h->mb.i_mb_x > 0 && h->mb.i_mb_x < h->mb.i_mb_width - 1) ||
742 h->mb.i_mb_width <= 2 || h->mb.i_mb_height <= 2 )
744 i_score += i_mb_cost;
751 static void x264_macroblock_tree_finish( x264_t *h, x264_frame_t *frame, int ref0_distance )
754 float weightdelta = 0.0;
755 if( ref0_distance && frame->f_weighted_cost_delta[ref0_distance-1] > 0 )
756 weightdelta = (1.0 - frame->f_weighted_cost_delta[ref0_distance-1]);
758 /* Allow the strength to be adjusted via qcompress, since the two
759 * concepts are very similar. */
760 float strength = 5.0f * (1.0f - h->param.rc.f_qcompress);
761 for( int mb_index = 0; mb_index < h->mb.i_mb_count; mb_index++ )
763 int intra_cost = (frame->i_intra_cost[mb_index] * frame->i_inv_qscale_factor[mb_index]+128)>>8;
766 int propagate_cost = frame->i_propagate_cost[mb_index];
767 float log2_ratio = x264_log2(intra_cost + propagate_cost) - x264_log2(intra_cost) + weightdelta;
768 frame->f_qp_offset[mb_index] = frame->f_qp_offset_aq[mb_index] - strength * log2_ratio;
773 static void x264_macroblock_tree_propagate( x264_t *h, x264_frame_t **frames, int p0, int p1, int b, int referenced )
775 uint16_t *ref_costs[2] = {frames[p0]->i_propagate_cost,frames[p1]->i_propagate_cost};
776 int dist_scale_factor = ( ((b-p0) << 8) + ((p1-p0) >> 1) ) / (p1-p0);
777 int i_bipred_weight = h->param.analyse.b_weighted_bipred ? 64 - (dist_scale_factor>>2) : 32;
778 int16_t (*mvs[2])[2] = { frames[b]->lowres_mvs[0][b-p0-1], frames[b]->lowres_mvs[1][p1-b-1] };
779 int bipred_weights[2] = {i_bipred_weight, 64 - i_bipred_weight};
780 int *buf = h->scratch_buffer;
781 uint16_t *propagate_cost = frames[b]->i_propagate_cost;
783 /* For non-reffed frames the source costs are always zero, so just memset one row and re-use it. */
785 memset( frames[b]->i_propagate_cost, 0, h->mb.i_mb_width * sizeof(uint16_t) );
787 for( h->mb.i_mb_y = 0; h->mb.i_mb_y < h->mb.i_mb_height; h->mb.i_mb_y++ )
789 int mb_index = h->mb.i_mb_y*h->mb.i_mb_stride;
790 h->mc.mbtree_propagate_cost( buf, propagate_cost,
791 frames[b]->i_intra_cost+mb_index, frames[b]->lowres_costs[b-p0][p1-b]+mb_index,
792 frames[b]->i_inv_qscale_factor+mb_index, h->mb.i_mb_width );
794 propagate_cost += h->mb.i_mb_width;
795 for( h->mb.i_mb_x = 0; h->mb.i_mb_x < h->mb.i_mb_width; h->mb.i_mb_x++, mb_index++ )
797 int propagate_amount = buf[h->mb.i_mb_x];
798 /* Don't propagate for an intra block. */
799 if( propagate_amount > 0 )
801 /* Access width-2 bitfield. */
802 int lists_used = frames[b]->lowres_costs[b-p0][p1-b][mb_index] >> LOWRES_COST_SHIFT;
803 /* Follow the MVs to the previous frame(s). */
804 for( int list = 0; list < 2; list++ )
805 if( (lists_used >> list)&1 )
807 #define CLIP_ADD(s,x) (s) = X264_MIN((s)+(x),(1<<16)-1)
808 int listamount = propagate_amount;
809 /* Apply bipred weighting. */
810 if( lists_used == 3 )
811 listamount = (listamount * bipred_weights[list] + 32) >> 6;
813 /* Early termination for simple case of mv0. */
814 if( !M32( mvs[list][mb_index] ) )
816 CLIP_ADD( ref_costs[list][mb_index], listamount );
820 int x = mvs[list][mb_index][0];
821 int y = mvs[list][mb_index][1];
822 int mbx = (x>>5)+h->mb.i_mb_x;
823 int mby = (y>>5)+h->mb.i_mb_y;
824 int idx0 = mbx + mby * h->mb.i_mb_stride;
826 int idx2 = idx0 + h->mb.i_mb_stride;
827 int idx3 = idx0 + h->mb.i_mb_stride + 1;
830 int idx0weight = (32-y)*(32-x);
831 int idx1weight = (32-y)*x;
832 int idx2weight = y*(32-x);
833 int idx3weight = y*x;
835 /* We could just clip the MVs, but pixels that lie outside the frame probably shouldn't
837 if( mbx < h->mb.i_mb_width-1 && mby < h->mb.i_mb_height-1 && mbx >= 0 && mby >= 0 )
839 CLIP_ADD( ref_costs[list][idx0], (listamount*idx0weight+512)>>10 );
840 CLIP_ADD( ref_costs[list][idx1], (listamount*idx1weight+512)>>10 );
841 CLIP_ADD( ref_costs[list][idx2], (listamount*idx2weight+512)>>10 );
842 CLIP_ADD( ref_costs[list][idx3], (listamount*idx3weight+512)>>10 );
844 else /* Check offsets individually */
846 if( mbx < h->mb.i_mb_width && mby < h->mb.i_mb_height && mbx >= 0 && mby >= 0 )
847 CLIP_ADD( ref_costs[list][idx0], (listamount*idx0weight+512)>>10 );
848 if( mbx+1 < h->mb.i_mb_width && mby < h->mb.i_mb_height && mbx+1 >= 0 && mby >= 0 )
849 CLIP_ADD( ref_costs[list][idx1], (listamount*idx1weight+512)>>10 );
850 if( mbx < h->mb.i_mb_width && mby+1 < h->mb.i_mb_height && mbx >= 0 && mby+1 >= 0 )
851 CLIP_ADD( ref_costs[list][idx2], (listamount*idx2weight+512)>>10 );
852 if( mbx+1 < h->mb.i_mb_width && mby+1 < h->mb.i_mb_height && mbx+1 >= 0 && mby+1 >= 0 )
853 CLIP_ADD( ref_costs[list][idx3], (listamount*idx3weight+512)>>10 );
860 if( h->param.rc.i_vbv_buffer_size && h->param.rc.i_lookahead && referenced )
861 x264_macroblock_tree_finish( h, frames[b], b == p1 ? b - p0 : 0 );
864 static void x264_macroblock_tree( x264_t *h, x264_mb_analysis_t *a, x264_frame_t **frames, int num_frames, int b_intra )
867 int last_nonb, cur_nonb = 1;
872 x264_slicetype_frame_cost( h, a, frames, 0, 0, 0, 0 );
874 while( i > 0 && frames[i]->i_type == X264_TYPE_B )
878 /* Lookaheadless MB-tree is not a theoretically distinct case; the same extrapolation could
879 * be applied to the end of a lookahead buffer of any size. However, it's most needed when
880 * lookahead=0, so that's what's currently implemented. */
881 if( !h->param.rc.i_lookahead )
885 memset( frames[0]->i_propagate_cost, 0, h->mb.i_mb_count * sizeof(uint16_t) );
886 memcpy( frames[0]->f_qp_offset, frames[0]->f_qp_offset_aq, h->mb.i_mb_count * sizeof(float) );
889 XCHG( uint16_t*, frames[last_nonb]->i_propagate_cost, frames[0]->i_propagate_cost );
890 memset( frames[0]->i_propagate_cost, 0, h->mb.i_mb_count * sizeof(uint16_t) );
894 if( last_nonb < idx )
896 memset( frames[last_nonb]->i_propagate_cost, 0, h->mb.i_mb_count * sizeof(uint16_t) );
902 while( frames[cur_nonb]->i_type == X264_TYPE_B && cur_nonb > 0 )
906 x264_slicetype_frame_cost( h, a, frames, cur_nonb, last_nonb, last_nonb, 0 );
907 memset( frames[cur_nonb]->i_propagate_cost, 0, h->mb.i_mb_count * sizeof(uint16_t) );
908 bframes = last_nonb - cur_nonb - 1;
909 if( h->param.i_bframe_pyramid && bframes > 1 )
911 int middle = (bframes + 1)/2 + cur_nonb;
912 x264_slicetype_frame_cost( h, a, frames, cur_nonb, last_nonb, middle, 0 );
913 memset( frames[middle]->i_propagate_cost, 0, h->mb.i_mb_count * sizeof(uint16_t) );
914 while( i > cur_nonb )
916 int p0 = i > middle ? middle : cur_nonb;
917 int p1 = i < middle ? middle : last_nonb;
920 x264_slicetype_frame_cost( h, a, frames, p0, p1, i, 0 );
921 x264_macroblock_tree_propagate( h, frames, p0, p1, i, 0 );
925 x264_macroblock_tree_propagate( h, frames, cur_nonb, last_nonb, middle, 1 );
929 while( i > cur_nonb )
931 x264_slicetype_frame_cost( h, a, frames, cur_nonb, last_nonb, i, 0 );
932 x264_macroblock_tree_propagate( h, frames, cur_nonb, last_nonb, i, 0 );
936 x264_macroblock_tree_propagate( h, frames, cur_nonb, last_nonb, last_nonb, 1 );
937 last_nonb = cur_nonb;
940 if( !h->param.rc.i_lookahead )
942 x264_macroblock_tree_propagate( h, frames, 0, last_nonb, last_nonb, 1 );
943 XCHG( uint16_t*, frames[last_nonb]->i_propagate_cost, frames[0]->i_propagate_cost );
946 x264_macroblock_tree_finish( h, frames[last_nonb], last_nonb );
947 if( h->param.i_bframe_pyramid && bframes > 1 && !h->param.rc.i_vbv_buffer_size )
948 x264_macroblock_tree_finish( h, frames[last_nonb+(bframes+1)/2], 0 );
951 static int x264_vbv_frame_cost( x264_t *h, x264_mb_analysis_t *a, x264_frame_t **frames, int p0, int p1, int b )
953 int cost = x264_slicetype_frame_cost( h, a, frames, p0, p1, b, 0 );
954 if( h->param.rc.i_aq_mode )
956 if( h->param.rc.b_mb_tree )
957 return x264_slicetype_frame_cost_recalculate( h, frames, p0, p1, b );
959 return frames[b]->i_cost_est_aq[b-p0][p1-b];
964 static void x264_calculate_durations( x264_t *h, x264_frame_t *cur_frame, x264_frame_t *prev_frame, int64_t *i_cpb_delay, int64_t *i_coded_fields )
966 cur_frame->i_cpb_delay = *i_cpb_delay;
967 cur_frame->i_dpb_output_delay = cur_frame->i_field_cnt - *i_coded_fields;
969 // add a correction term for frame reordering
970 cur_frame->i_dpb_output_delay += h->sps->vui.i_num_reorder_frames*2;
972 // fix possible negative dpb_output_delay because of pulldown changes and reordering
973 if( cur_frame->i_dpb_output_delay < 0 )
975 cur_frame->i_cpb_delay += cur_frame->i_dpb_output_delay;
976 cur_frame->i_dpb_output_delay = 0;
978 prev_frame->i_cpb_duration += cur_frame->i_dpb_output_delay;
981 if( cur_frame->b_keyframe )
984 *i_cpb_delay += cur_frame->i_duration;
985 *i_coded_fields += cur_frame->i_duration;
986 cur_frame->i_cpb_duration = cur_frame->i_duration;
989 static void x264_vbv_lookahead( x264_t *h, x264_mb_analysis_t *a, x264_frame_t **frames, int num_frames, int keyframe )
991 int last_nonb = 0, cur_nonb = 1, idx = 0;
992 x264_frame_t *prev_frame = NULL;
993 int prev_frame_idx = 0;
994 while( cur_nonb < num_frames && frames[cur_nonb]->i_type == X264_TYPE_B )
996 int next_nonb = keyframe ? last_nonb : cur_nonb;
998 if( frames[cur_nonb]->i_coded_fields_lookahead >= 0 )
1000 h->i_coded_fields_lookahead = frames[cur_nonb]->i_coded_fields_lookahead;
1001 h->i_cpb_delay_lookahead = frames[cur_nonb]->i_cpb_delay_lookahead;
1004 while( cur_nonb < num_frames )
1006 /* P/I cost: This shouldn't include the cost of next_nonb */
1007 if( next_nonb != cur_nonb )
1009 int p0 = IS_X264_TYPE_I( frames[cur_nonb]->i_type ) ? cur_nonb : last_nonb;
1010 frames[next_nonb]->i_planned_satd[idx] = x264_vbv_frame_cost( h, a, frames, p0, cur_nonb, cur_nonb );
1011 frames[next_nonb]->i_planned_type[idx] = frames[cur_nonb]->i_type;
1012 frames[cur_nonb]->i_coded_fields_lookahead = h->i_coded_fields_lookahead;
1013 frames[cur_nonb]->i_cpb_delay_lookahead = h->i_cpb_delay_lookahead;
1014 x264_calculate_durations( h, frames[cur_nonb], prev_frame, &h->i_cpb_delay_lookahead, &h->i_coded_fields_lookahead );
1017 frames[next_nonb]->f_planned_cpb_duration[prev_frame_idx] = (double)prev_frame->i_cpb_duration *
1018 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1020 frames[next_nonb]->f_planned_cpb_duration[idx] = (double)frames[cur_nonb]->i_cpb_duration *
1021 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1022 prev_frame = frames[cur_nonb];
1023 prev_frame_idx = idx;
1026 /* Handle the B-frames: coded order */
1027 for( int i = last_nonb+1; i < cur_nonb; i++, idx++ )
1029 frames[next_nonb]->i_planned_satd[idx] = x264_vbv_frame_cost( h, a, frames, last_nonb, cur_nonb, i );
1030 frames[next_nonb]->i_planned_type[idx] = X264_TYPE_B;
1031 frames[i]->i_coded_fields_lookahead = h->i_coded_fields_lookahead;
1032 frames[i]->i_cpb_delay_lookahead = h->i_cpb_delay_lookahead;
1033 x264_calculate_durations( h, frames[i], prev_frame, &h->i_cpb_delay_lookahead, &h->i_coded_fields_lookahead );
1036 frames[next_nonb]->f_planned_cpb_duration[prev_frame_idx] = (double)prev_frame->i_cpb_duration *
1037 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1039 frames[next_nonb]->f_planned_cpb_duration[idx] = (double)frames[i]->i_cpb_duration *
1040 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1041 prev_frame = frames[i];
1042 prev_frame_idx = idx;
1044 last_nonb = cur_nonb;
1046 while( cur_nonb <= num_frames && frames[cur_nonb]->i_type == X264_TYPE_B )
1049 frames[next_nonb]->i_planned_type[idx] = X264_TYPE_AUTO;
1052 static int x264_slicetype_path_cost( x264_t *h, x264_mb_analysis_t *a, x264_frame_t **frames, char *path, int threshold )
1057 path--; /* Since the 1st path element is really the second frame */
1061 /* Find the location of the next P-frame. */
1062 while( path[next_p] != 'P' )
1065 /* Add the cost of the P-frame found above */
1066 cost += x264_slicetype_frame_cost( h, a, frames, cur_p, next_p, next_p, 0 );
1067 /* Early terminate if the cost we have found is larger than the best path cost so far */
1068 if( cost > threshold )
1071 if( h->param.i_bframe_pyramid && next_p - cur_p > 2 )
1073 int middle = cur_p + (next_p - cur_p)/2;
1074 cost += x264_slicetype_frame_cost( h, a, frames, cur_p, next_p, middle, 0 );
1075 for( int next_b = loc; next_b < middle && cost < threshold; next_b++ )
1076 cost += x264_slicetype_frame_cost( h, a, frames, cur_p, middle, next_b, 0 );
1077 for( int next_b = middle+1; next_b < next_p && cost < threshold; next_b++ )
1078 cost += x264_slicetype_frame_cost( h, a, frames, middle, next_p, next_b, 0 );
1081 for( int next_b = loc; next_b < next_p && cost < threshold; next_b++ )
1082 cost += x264_slicetype_frame_cost( h, a, frames, cur_p, next_p, next_b, 0 );
1090 /* Viterbi/trellis slicetype decision algorithm. */
1091 /* Uses strings due to the fact that the speed of the control functions is
1092 negligible compared to the cost of running slicetype_frame_cost, and because
1093 it makes debugging easier. */
1094 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] )
1096 char paths[2][X264_LOOKAHEAD_MAX];
1097 int num_paths = X264_MIN( h->param.i_bframe+1, length );
1098 int best_cost = COST_MAX;
1101 /* Iterate over all currently possible paths */
1102 for( int path = 0; path < num_paths; path++ )
1104 /* Add suffixes to the current path */
1105 int len = length - (path + 1);
1106 memcpy( paths[idx], best_paths[len % (X264_BFRAME_MAX+1)], len );
1107 memset( paths[idx]+len, 'B', path );
1108 strcpy( paths[idx]+len+path, "P" );
1110 /* Calculate the actual cost of the current path */
1111 int cost = x264_slicetype_path_cost( h, a, frames, paths[idx], best_cost );
1112 if( cost < best_cost )
1119 /* Store the best path. */
1120 memcpy( best_paths[length % (X264_BFRAME_MAX+1)], paths[idx^1], length );
1123 static int scenecut_internal( x264_t *h, x264_mb_analysis_t *a, x264_frame_t **frames, int p0, int p1, int print )
1125 x264_frame_t *frame = frames[p1];
1126 x264_slicetype_frame_cost( h, a, frames, p0, p1, p1, 0 );
1128 int icost = frame->i_cost_est[0][0];
1129 int pcost = frame->i_cost_est[p1-p0][0];
1131 int i_gop_size = frame->i_frame - h->lookahead->i_last_keyframe;
1132 float f_thresh_max = h->param.i_scenecut_threshold / 100.0;
1133 /* magic numbers pulled out of thin air */
1134 float f_thresh_min = f_thresh_max * 0.25;
1137 if( h->param.i_keyint_min == h->param.i_keyint_max )
1138 f_thresh_min = f_thresh_max;
1139 if( i_gop_size <= h->param.i_keyint_min / 4 || h->param.b_intra_refresh )
1140 f_bias = f_thresh_min / 4;
1141 else if( i_gop_size <= h->param.i_keyint_min )
1142 f_bias = f_thresh_min * i_gop_size / h->param.i_keyint_min;
1145 f_bias = f_thresh_min
1146 + ( f_thresh_max - f_thresh_min )
1147 * ( i_gop_size - h->param.i_keyint_min )
1148 / ( h->param.i_keyint_max - h->param.i_keyint_min );
1151 res = pcost >= (1.0 - f_bias) * icost;
1154 int imb = frame->i_intra_mbs[p1-p0];
1155 int pmb = NUM_MBS - imb;
1156 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",
1158 icost, pcost, 1. - (double)pcost / icost,
1159 f_bias, i_gop_size, imb, pmb );
1164 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, int i_max_search )
1166 /* Only do analysis during a normal scenecut check. */
1167 if( real_scenecut && h->param.i_bframe )
1169 int origmaxp1 = p0 + 1;
1170 /* Look ahead to avoid coding short flashes as scenecuts. */
1171 if( h->param.i_bframe_adaptive == X264_B_ADAPT_TRELLIS )
1172 /* Don't analyse any more frames than the trellis would have covered. */
1173 origmaxp1 += h->param.i_bframe;
1176 int maxp1 = X264_MIN( origmaxp1, num_frames );
1178 /* Where A and B are scenes: AAAAAABBBAAAAAA
1179 * If BBB is shorter than (maxp1-p0), it is detected as a flash
1180 * and not considered a scenecut. */
1181 for( int curp1 = p1; curp1 <= maxp1; curp1++ )
1182 if( !scenecut_internal( h, a, frames, p0, curp1, 0 ) )
1183 /* Any frame in between p0 and cur_p1 cannot be a real scenecut. */
1184 for( int i = curp1; i > p0; i-- )
1185 frames[i]->b_scenecut = 0;
1187 /* Where A-F are scenes: AAAAABBCCDDEEFFFFFF
1188 * If each of BB ... EE are shorter than (maxp1-p0), they are
1189 * detected as flashes and not considered scenecuts.
1190 * Instead, the first F frame becomes a scenecut.
1191 * If the video ends before F, no frame becomes a scenecut. */
1192 for( int curp0 = p0; curp0 <= maxp1; curp0++ )
1193 if( origmaxp1 > i_max_search || (curp0 < maxp1 && scenecut_internal( h, a, frames, curp0, maxp1, 0 )) )
1194 /* If cur_p0 is the p0 of a scenecut, it cannot be the p1 of a scenecut. */
1195 frames[curp0]->b_scenecut = 0;
1198 /* Ignore frames that are part of a flash, i.e. cannot be real scenecuts. */
1199 if( !frames[p1]->b_scenecut )
1201 return scenecut_internal( h, a, frames, p0, p1, real_scenecut );
1204 void x264_slicetype_analyse( x264_t *h, int keyframe )
1206 x264_mb_analysis_t a;
1207 x264_frame_t *frames[X264_LOOKAHEAD_MAX+3] = { NULL, };
1208 int num_frames, orig_num_frames, keyint_limit, framecnt;
1209 int i_mb_count = NUM_MBS;
1210 int cost1p0, cost2p0, cost1b1, cost2p1;
1211 int i_max_search = X264_MIN( h->lookahead->next.i_size, X264_LOOKAHEAD_MAX );
1212 int vbv_lookahead = h->param.rc.i_vbv_buffer_size && h->param.rc.i_lookahead;
1213 if( h->param.b_deterministic )
1214 i_max_search = X264_MIN( i_max_search, h->lookahead->i_slicetype_length + !keyframe );
1216 assert( h->frames.b_have_lowres );
1218 if( !h->lookahead->last_nonb )
1220 frames[0] = h->lookahead->last_nonb;
1221 for( framecnt = 0; framecnt < i_max_search && h->lookahead->next.list[framecnt]->i_type == X264_TYPE_AUTO; framecnt++ )
1222 frames[framecnt+1] = h->lookahead->next.list[framecnt];
1224 x264_lowres_context_init( h, &a );
1228 if( h->param.rc.b_mb_tree )
1229 x264_macroblock_tree( h, &a, frames, 0, keyframe );
1233 keyint_limit = h->param.i_keyint_max - frames[0]->i_frame + h->lookahead->i_last_keyframe - 1;
1234 orig_num_frames = num_frames = h->param.b_intra_refresh ? framecnt : X264_MIN( framecnt, keyint_limit );
1236 /* This is important psy-wise: if we have a non-scenecut keyframe,
1237 * there will be significant visual artifacts if the frames just before
1238 * go down in quality due to being referenced less, despite it being
1239 * more RD-optimal. */
1240 if( (h->param.analyse.b_psy && h->param.rc.b_mb_tree) || vbv_lookahead )
1241 num_frames = framecnt;
1242 else if( num_frames == 0 )
1244 frames[1]->i_type = X264_TYPE_I;
1248 int num_bframes = 0;
1249 int num_analysed_frames = num_frames;
1251 if( h->param.i_scenecut_threshold && scenecut( h, &a, frames, 0, 1, 1, orig_num_frames, i_max_search ) )
1253 frames[1]->i_type = X264_TYPE_I;
1257 if( h->param.i_bframe )
1259 if( h->param.i_bframe_adaptive == X264_B_ADAPT_TRELLIS )
1261 if( num_frames > 1 )
1263 char best_paths[X264_BFRAME_MAX+1][X264_LOOKAHEAD_MAX] = {"","P"};
1264 int best_path_index = (num_frames-1) % (X264_BFRAME_MAX+1);
1266 /* Perform the frametype analysis. */
1267 for( int j = 2; j < num_frames; j++ )
1268 x264_slicetype_path( h, &a, frames, j, best_paths );
1270 num_bframes = strspn( best_paths[best_path_index], "B" );
1271 /* Load the results of the analysis into the frame types. */
1272 for( int j = 1; j < num_frames; j++ )
1273 frames[j]->i_type = best_paths[best_path_index][j-1] == 'B' ? X264_TYPE_B : X264_TYPE_P;
1275 frames[num_frames]->i_type = X264_TYPE_P;
1277 else if( h->param.i_bframe_adaptive == X264_B_ADAPT_FAST )
1279 for( int i = 0; i <= num_frames-2; )
1281 cost2p1 = x264_slicetype_frame_cost( h, &a, frames, i+0, i+2, i+2, 1 );
1282 if( frames[i+2]->i_intra_mbs[2] > i_mb_count / 2 )
1284 frames[i+1]->i_type = X264_TYPE_P;
1285 frames[i+2]->i_type = X264_TYPE_P;
1290 cost1b1 = x264_slicetype_frame_cost( h, &a, frames, i+0, i+2, i+1, 0 );
1291 cost1p0 = x264_slicetype_frame_cost( h, &a, frames, i+0, i+1, i+1, 0 );
1292 cost2p0 = x264_slicetype_frame_cost( h, &a, frames, i+1, i+2, i+2, 0 );
1294 if( cost1p0 + cost2p0 < cost1b1 + cost2p1 )
1296 frames[i+1]->i_type = X264_TYPE_P;
1301 // arbitrary and untuned
1302 #define INTER_THRESH 300
1303 #define P_SENS_BIAS (50 - h->param.i_bframe_bias)
1304 frames[i+1]->i_type = X264_TYPE_B;
1307 for( j = i+2; j <= X264_MIN( i+h->param.i_bframe, num_frames-1 ); j++ )
1309 int pthresh = X264_MAX(INTER_THRESH - P_SENS_BIAS * (j-i-1), INTER_THRESH/10);
1310 int pcost = x264_slicetype_frame_cost( h, &a, frames, i+0, j+1, j+1, 1 );
1311 if( pcost > pthresh*i_mb_count || frames[j+1]->i_intra_mbs[j-i+1] > i_mb_count/3 )
1313 frames[j]->i_type = X264_TYPE_B;
1315 frames[j]->i_type = X264_TYPE_P;
1318 frames[num_frames]->i_type = X264_TYPE_P;
1320 while( num_bframes < num_frames && frames[num_bframes+1]->i_type == X264_TYPE_B )
1325 num_bframes = X264_MIN(num_frames-1, h->param.i_bframe);
1326 for( int j = 1; j < num_frames; j++ )
1327 frames[j]->i_type = (j%(num_bframes+1)) ? X264_TYPE_B : X264_TYPE_P;
1328 frames[num_frames]->i_type = X264_TYPE_P;
1331 /* Check scenecut on the first minigop. */
1332 for( int j = 1; j < num_bframes+1; j++ )
1333 if( h->param.i_scenecut_threshold && scenecut( h, &a, frames, j, j+1, 0, orig_num_frames, i_max_search ) )
1335 frames[j]->i_type = X264_TYPE_P;
1336 num_analysed_frames = j;
1340 reset_start = keyframe ? 1 : X264_MIN( num_bframes+2, num_analysed_frames+1 );
1344 for( int j = 1; j <= num_frames; j++ )
1345 frames[j]->i_type = X264_TYPE_P;
1346 reset_start = !keyframe + 1;
1350 /* Perform the actual macroblock tree analysis.
1351 * Don't go farther than the maximum keyframe interval; this helps in short GOPs. */
1352 if( h->param.rc.b_mb_tree )
1353 x264_macroblock_tree( h, &a, frames, X264_MIN(num_frames, h->param.i_keyint_max), keyframe );
1355 /* Enforce keyframe limit. */
1356 if( !h->param.b_intra_refresh )
1357 for( int i = keyint_limit+1; i <= num_frames; i += h->param.i_keyint_max )
1359 frames[i]->i_type = X264_TYPE_I;
1360 reset_start = X264_MIN( reset_start, i+1 );
1361 if( h->param.i_open_gop == X264_OPEN_GOP_BLURAY )
1362 while( IS_X264_TYPE_B( frames[i-1]->i_type ) )
1367 x264_vbv_lookahead( h, &a, frames, num_frames, keyframe );
1369 /* Restore frametypes for all frames that haven't actually been decided yet. */
1370 for( int j = reset_start; j <= num_frames; j++ )
1371 frames[j]->i_type = X264_TYPE_AUTO;
1374 void x264_slicetype_decide( x264_t *h )
1376 x264_frame_t *frames[X264_BFRAME_MAX+2];
1381 if( !h->lookahead->next.i_size )
1384 int lookahead_size = h->lookahead->next.i_size;
1386 for( int i = 0; i < h->lookahead->next.i_size; i++ )
1388 if( h->param.b_vfr_input )
1390 if( lookahead_size-- > 1 )
1391 h->lookahead->next.list[i]->i_duration = 2 * (h->lookahead->next.list[i+1]->i_pts - h->lookahead->next.list[i]->i_pts);
1393 h->lookahead->next.list[i]->i_duration = h->i_prev_duration;
1396 h->lookahead->next.list[i]->i_duration = delta_tfi_divisor[h->lookahead->next.list[i]->i_pic_struct];
1397 h->i_prev_duration = h->lookahead->next.list[i]->i_duration;
1398 h->lookahead->next.list[i]->f_duration = (double)h->lookahead->next.list[i]->i_duration
1399 * h->sps->vui.i_num_units_in_tick
1400 / h->sps->vui.i_time_scale;
1402 if( h->lookahead->next.list[i]->i_frame > h->i_disp_fields_last_frame && lookahead_size > 0 )
1404 h->lookahead->next.list[i]->i_field_cnt = h->i_disp_fields;
1405 h->i_disp_fields += h->lookahead->next.list[i]->i_duration;
1406 h->i_disp_fields_last_frame = h->lookahead->next.list[i]->i_frame;
1408 else if( lookahead_size == 0 )
1410 h->lookahead->next.list[i]->i_field_cnt = h->i_disp_fields;
1411 h->lookahead->next.list[i]->i_duration = h->i_prev_duration;
1415 if( h->param.rc.b_stat_read )
1417 /* Use the frame types from the first pass */
1418 for( int i = 0; i < h->lookahead->next.i_size; i++ )
1419 h->lookahead->next.list[i]->i_type =
1420 x264_ratecontrol_slice_type( h, h->lookahead->next.list[i]->i_frame );
1422 else if( (h->param.i_bframe && h->param.i_bframe_adaptive)
1423 || h->param.i_scenecut_threshold
1424 || h->param.rc.b_mb_tree
1425 || (h->param.rc.i_vbv_buffer_size && h->param.rc.i_lookahead) )
1426 x264_slicetype_analyse( h, 0 );
1428 for( bframes = 0, brefs = 0;; bframes++ )
1430 frm = h->lookahead->next.list[bframes];
1431 if( frm->i_type == X264_TYPE_BREF && h->param.i_bframe_pyramid < X264_B_PYRAMID_NORMAL &&
1432 brefs == h->param.i_bframe_pyramid )
1434 frm->i_type = X264_TYPE_B;
1435 x264_log( h, X264_LOG_WARNING, "B-ref at frame %d incompatible with B-pyramid %s \n",
1436 frm->i_frame, x264_b_pyramid_names[h->param.i_bframe_pyramid] );
1438 /* pyramid with multiple B-refs needs a big enough dpb that the preceding P-frame stays available.
1439 smaller dpb could be supported by smart enough use of mmco, but it's easier just to forbid it. */
1440 else if( frm->i_type == X264_TYPE_BREF && h->param.i_bframe_pyramid == X264_B_PYRAMID_NORMAL &&
1441 brefs && h->param.i_frame_reference <= (brefs+3) )
1443 frm->i_type = X264_TYPE_B;
1444 x264_log( h, X264_LOG_WARNING, "B-ref at frame %d incompatible with B-pyramid %s and %d reference frames\n",
1445 frm->i_frame, x264_b_pyramid_names[h->param.i_bframe_pyramid], h->param.i_frame_reference );
1448 if( frm->i_type == X264_TYPE_KEYFRAME )
1449 frm->i_type = h->param.i_open_gop ? X264_TYPE_I : X264_TYPE_IDR;
1451 /* Limit GOP size */
1452 if( (!h->param.b_intra_refresh || frm->i_frame == 0) && frm->i_frame - h->lookahead->i_last_keyframe >= h->param.i_keyint_max )
1454 if( frm->i_type == X264_TYPE_AUTO || frm->i_type == X264_TYPE_I )
1455 frm->i_type = h->param.i_open_gop && h->lookahead->i_last_keyframe >= 0 ? X264_TYPE_I : X264_TYPE_IDR;
1456 int warn = frm->i_type != X264_TYPE_IDR;
1457 if( warn && h->param.i_open_gop )
1458 warn &= frm->i_type != X264_TYPE_I;
1460 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 );
1462 if( frm->i_type == X264_TYPE_I && frm->i_frame - h->lookahead->i_last_keyframe >= h->param.i_keyint_min )
1464 if( h->param.i_open_gop )
1466 h->lookahead->i_last_keyframe = frm->i_frame; // Use display order
1467 if( h->param.i_open_gop == X264_OPEN_GOP_BLURAY )
1468 h->lookahead->i_last_keyframe -= bframes; // Use bluray order
1469 frm->b_keyframe = 1;
1472 frm->i_type = X264_TYPE_IDR;
1474 if( frm->i_type == X264_TYPE_IDR )
1477 h->lookahead->i_last_keyframe = frm->i_frame;
1478 frm->b_keyframe = 1;
1482 h->lookahead->next.list[bframes]->i_type = X264_TYPE_P;
1486 if( bframes == h->param.i_bframe ||
1487 !h->lookahead->next.list[bframes+1] )
1489 if( IS_X264_TYPE_B( frm->i_type ) )
1490 x264_log( h, X264_LOG_WARNING, "specified frame type is not compatible with max B-frames\n" );
1491 if( frm->i_type == X264_TYPE_AUTO
1492 || IS_X264_TYPE_B( frm->i_type ) )
1493 frm->i_type = X264_TYPE_P;
1496 if( frm->i_type == X264_TYPE_BREF )
1499 if( frm->i_type == X264_TYPE_AUTO )
1500 frm->i_type = X264_TYPE_B;
1502 else if( !IS_X264_TYPE_B( frm->i_type ) ) break;
1506 h->lookahead->next.list[bframes-1]->b_last_minigop_bframe = 1;
1507 h->lookahead->next.list[bframes]->i_bframes = bframes;
1509 /* insert a bref into the sequence */
1510 if( h->param.i_bframe_pyramid && bframes > 1 && !brefs )
1512 h->lookahead->next.list[bframes/2]->i_type = X264_TYPE_BREF;
1516 /* calculate the frame costs ahead of time for x264_rc_analyse_slice while we still have lowres */
1517 if( h->param.rc.i_rc_method != X264_RC_CQP )
1519 x264_mb_analysis_t a;
1521 p1 = b = bframes + 1;
1523 x264_lowres_context_init( h, &a );
1525 frames[0] = h->lookahead->last_nonb;
1526 memcpy( &frames[1], h->lookahead->next.list, (bframes+1) * sizeof(x264_frame_t*) );
1527 if( IS_X264_TYPE_I( h->lookahead->next.list[bframes]->i_type ) )
1532 x264_slicetype_frame_cost( h, &a, frames, p0, p1, b, 0 );
1534 if( (p0 != p1 || bframes) && h->param.rc.i_vbv_buffer_size )
1536 /* We need the intra costs for row SATDs. */
1537 x264_slicetype_frame_cost( h, &a, frames, b, b, b, 0 );
1539 /* We need B-frame costs for row SATDs. */
1541 for( b = 1; b <= bframes; b++ )
1543 if( frames[b]->i_type == X264_TYPE_B )
1544 for( p1 = b; frames[p1]->i_type == X264_TYPE_B; )
1548 x264_slicetype_frame_cost( h, &a, frames, p0, p1, b, 0 );
1549 if( frames[b]->i_type == X264_TYPE_BREF )
1555 /* Analyse for weighted P frames */
1556 if( !h->param.rc.b_stat_read && h->lookahead->next.list[bframes]->i_type == X264_TYPE_P
1557 && h->param.analyse.i_weighted_pred >= X264_WEIGHTP_SIMPLE )
1560 x264_weights_analyse( h, h->lookahead->next.list[bframes], h->lookahead->last_nonb, 0 );
1563 /* shift sequence to coded order.
1564 use a small temporary list to avoid shifting the entire next buffer around */
1565 int i_coded = h->lookahead->next.list[0]->i_frame;
1568 int idx_list[] = { brefs+1, 1 };
1569 for( int i = 0; i < bframes; i++ )
1571 int idx = idx_list[h->lookahead->next.list[i]->i_type == X264_TYPE_BREF]++;
1572 frames[idx] = h->lookahead->next.list[i];
1573 frames[idx]->i_reordered_pts = h->lookahead->next.list[idx]->i_pts;
1575 frames[0] = h->lookahead->next.list[bframes];
1576 frames[0]->i_reordered_pts = h->lookahead->next.list[0]->i_pts;
1577 memcpy( h->lookahead->next.list, frames, (bframes+1) * sizeof(x264_frame_t*) );
1580 for( int i = 0; i <= bframes; i++ )
1582 h->lookahead->next.list[i]->i_coded = i_coded++;
1585 x264_calculate_durations( h, h->lookahead->next.list[i], h->lookahead->next.list[i-1], &h->i_cpb_delay, &h->i_coded_fields );
1586 h->lookahead->next.list[0]->f_planned_cpb_duration[i-1] = (double)h->lookahead->next.list[i-1]->i_cpb_duration *
1587 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1590 x264_calculate_durations( h, h->lookahead->next.list[i], NULL, &h->i_cpb_delay, &h->i_coded_fields );
1592 h->lookahead->next.list[0]->f_planned_cpb_duration[i] = (double)h->lookahead->next.list[i]->i_cpb_duration *
1593 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1597 int x264_rc_analyse_slice( x264_t *h )
1603 if( IS_X264_TYPE_I(h->fenc->i_type) )
1605 else if( h->fenc->i_type == X264_TYPE_P )
1606 p1 = b = h->fenc->i_bframes + 1;
1609 p1 = (h->fref1[0]->i_poc - h->fref0[0]->i_poc)/2;
1610 b = (h->fenc->i_poc - h->fref0[0]->i_poc)/2;
1612 /* We don't need to assign p0/p1 since we are not performing any real analysis here. */
1613 x264_frame_t **frames = &h->fenc - b;
1615 /* cost should have been already calculated by x264_slicetype_decide */
1616 cost = frames[b]->i_cost_est[b-p0][p1-b];
1617 assert( cost >= 0 );
1619 if( h->param.rc.b_mb_tree && !h->param.rc.b_stat_read )
1621 cost = x264_slicetype_frame_cost_recalculate( h, frames, p0, p1, b );
1622 if( b && h->param.rc.i_vbv_buffer_size )
1623 x264_slicetype_frame_cost_recalculate( h, frames, b, b, b );
1625 /* In AQ, use the weighted score instead. */
1626 else if( h->param.rc.i_aq_mode )
1627 cost = frames[b]->i_cost_est_aq[b-p0][p1-b];
1629 h->fenc->i_row_satd = h->fenc->i_row_satds[b-p0][p1-b];
1630 h->fdec->i_row_satd = h->fdec->i_row_satds[b-p0][p1-b];
1631 h->fdec->i_satd = cost;
1632 memcpy( h->fdec->i_row_satd, h->fenc->i_row_satd, h->mb.i_mb_height * sizeof(int) );
1633 if( !IS_X264_TYPE_I(h->fenc->i_type) )
1634 memcpy( h->fdec->i_row_satds[0][0], h->fenc->i_row_satds[0][0], h->mb.i_mb_height * sizeof(int) );
1636 if( h->param.b_intra_refresh && h->param.rc.i_vbv_buffer_size && h->fenc->i_type == X264_TYPE_P )
1638 int ip_factor = 256 * h->param.rc.f_ip_factor; /* fix8 */
1639 for( int y = 0; y < h->mb.i_mb_height; y++ )
1641 int mb_xy = y * h->mb.i_mb_stride;
1642 for( int x = h->fdec->i_pir_start_col; x <= h->fdec->i_pir_end_col; x++, mb_xy++ )
1644 int intra_cost = (h->fenc->i_intra_cost[mb_xy] * ip_factor + 128) >> 8;
1645 int inter_cost = h->fenc->lowres_costs[b-p0][p1-b][mb_xy] & LOWRES_COST_MASK;
1646 int diff = intra_cost - inter_cost;
1647 if( h->param.rc.i_aq_mode )
1648 h->fdec->i_row_satd[y] += (diff * frames[b]->i_inv_qscale_factor[mb_xy] + 128) >> 8;
1650 h->fdec->i_row_satd[y] += diff;
1657 for( int y = 0; y < h->mb.i_mb_height; y++ )
1658 h->fdec->i_row_satd[y] >>= (BIT_DEPTH - 8);
1660 return cost >> (BIT_DEPTH - 8);