1 /*****************************************************************************
2 * slicetype.c: lookahead analysis
3 *****************************************************************************
4 * Copyright (C) 2005-2011 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 *****************************************************************************/
28 #include "common/common.h"
29 #include "macroblock.h"
32 // Indexed by pic_struct values
33 static const uint8_t delta_tfi_divisor[10] = { 0, 2, 1, 1, 2, 2, 3, 3, 4, 6 };
35 static int x264_slicetype_frame_cost( x264_t *h, x264_mb_analysis_t *a,
36 x264_frame_t **frames, int p0, int p1, int b,
37 int b_intra_penalty );
39 static void x264_lowres_context_init( x264_t *h, x264_mb_analysis_t *a )
41 a->i_qp = X264_LOOKAHEAD_QP;
42 a->i_lambda = x264_lambda_tab[ a->i_qp ];
43 x264_mb_analyse_load_costs( h, a );
44 if( h->param.analyse.i_subpel_refine > 1 )
46 h->mb.i_me_method = X264_MIN( X264_ME_HEX, h->param.analyse.i_me_method );
47 h->mb.i_subpel_refine = 4;
51 h->mb.i_me_method = X264_ME_DIA;
52 h->mb.i_subpel_refine = 2;
54 h->mb.b_chroma_me = 0;
57 /* makes a non-h264 weight (i.e. fix7), into an h264 weight */
58 static void x264_weight_get_h264( int weight_nonh264, int offset, x264_weight_t *w )
62 w->i_scale = weight_nonh264;
63 while( w->i_denom > 0 && (w->i_scale > 127 || !(w->i_scale & 1)) )
68 w->i_scale = X264_MIN( w->i_scale, 127 );
71 static NOINLINE pixel *x264_weight_cost_init_luma( x264_t *h, x264_frame_t *fenc, x264_frame_t *ref, pixel *dest )
73 int ref0_distance = fenc->i_frame - ref->i_frame - 1;
74 /* Note: this will never run during lookahead as weights_analyse is only called if no
75 * motion search has been done. */
76 if( fenc->lowres_mvs[0][ref0_distance][0][0] != 0x7FFF )
78 int i_stride = fenc->i_stride_lowres;
79 int i_lines = fenc->i_lines_lowres;
80 int i_width = fenc->i_width_lowres;
84 for( int y = 0; y < i_lines; y += 8, p += i_stride*8 )
85 for( int x = 0; x < i_width; x += 8, i_mb_xy++ )
87 int mvx = fenc->lowres_mvs[0][ref0_distance][i_mb_xy][0];
88 int mvy = fenc->lowres_mvs[0][ref0_distance][i_mb_xy][1];
89 h->mc.mc_luma( p+x, i_stride, ref->lowres, i_stride,
90 mvx+(x<<2), mvy+(y<<2), 8, 8, x264_weight_none );
96 return ref->lowres[0];
99 /* How data is organized for 4:2:0/4:2:2 chroma weightp:
102 * fenc = ref + offset
103 * v = u + stride * chroma height */
105 static NOINLINE void x264_weight_cost_init_chroma( x264_t *h, x264_frame_t *fenc, x264_frame_t *ref, pixel *dstu, pixel *dstv )
107 int ref0_distance = fenc->i_frame - ref->i_frame - 1;
108 int i_stride = fenc->i_stride[1];
109 int i_offset = i_stride / 2;
110 int i_lines = fenc->i_lines[1];
111 int i_width = fenc->i_width[1];
112 int v_shift = h->mb.chroma_v_shift;
113 int cw = 8*h->mb.i_mb_width;
114 int ch = 16*h->mb.i_mb_height >> v_shift;
115 int height = 16 >> v_shift;
117 if( fenc->lowres_mvs[0][ref0_distance][0][0] != 0x7FFF )
119 x264_frame_expand_border_chroma( h, ref, 1 );
120 for( int y = 0, mb_xy = 0, pel_offset_y = 0; y < i_lines; y += height, 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/NV16 */
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, 2*mvy>>v_shift, 8, height );
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 NOINLINE pixel *x264_weight_cost_init_chroma444( x264_t *h, x264_frame_t *fenc, x264_frame_t *ref, pixel *dst, int p )
139 int ref0_distance = fenc->i_frame - ref->i_frame - 1;
140 int i_stride = fenc->i_stride[p];
141 int i_lines = fenc->i_lines[p];
142 int i_width = fenc->i_width[p];
144 if( fenc->lowres_mvs[0][ref0_distance][0][0] != 0x7FFF )
146 x264_frame_expand_border_chroma( h, ref, p );
147 for( int y = 0, mb_xy = 0, pel_offset_y = 0; y < i_lines; y += 16, pel_offset_y = y*i_stride )
148 for( int x = 0, pel_offset_x = 0; x < i_width; x += 16, mb_xy++, pel_offset_x += 16 )
150 pixel *pix = dst + pel_offset_y + pel_offset_x;
151 pixel *src = ref->plane[p] + pel_offset_y + pel_offset_x;
152 int mvx = fenc->lowres_mvs[0][ref0_distance][mb_xy][0] / 2;
153 int mvy = fenc->lowres_mvs[0][ref0_distance][mb_xy][1] / 2;
154 /* We don't want to calculate hpels for fenc frames, so we round the motion
155 * vectors to fullpel here. It's not too bad, I guess? */
156 h->mc.copy_16x16_unaligned( pix, i_stride, src+mvx+mvy*i_stride, i_stride, 16 );
162 return ref->plane[p];
165 static int x264_weight_slice_header_cost( x264_t *h, x264_weight_t *w, int b_chroma )
167 /* Add cost of weights in the slice header. */
168 int lambda = x264_lambda_tab[X264_LOOKAHEAD_QP];
169 /* 4 times higher, because chroma is analyzed at full resolution. */
173 if( h->param.i_slice_count )
174 numslices = h->param.i_slice_count;
175 else if( h->param.i_slice_max_mbs )
176 numslices = (h->mb.i_mb_width * h->mb.i_mb_height + h->param.i_slice_max_mbs-1) / h->param.i_slice_max_mbs;
179 /* FIXME: find a way to account for --slice-max-size?
180 * 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.
181 * Cut denom cost in half if chroma, since it's shared between the two chroma planes. */
182 int denom_cost = bs_size_ue( w[0].i_denom ) * (2 - b_chroma);
183 return lambda * numslices * ( 10 + denom_cost + 2 * (bs_size_se( w[0].i_scale ) + bs_size_se( w[0].i_offset )) );
186 static NOINLINE unsigned int x264_weight_cost_luma( x264_t *h, x264_frame_t *fenc, pixel *src, x264_weight_t *w )
188 unsigned int cost = 0;
189 int i_stride = fenc->i_stride_lowres;
190 int i_lines = fenc->i_lines_lowres;
191 int i_width = fenc->i_width_lowres;
192 pixel *fenc_plane = fenc->lowres[0];
193 ALIGNED_ARRAY_16( pixel, buf,[8*8] );
199 for( int y = 0; y < i_lines; y += 8, pixoff = y*i_stride )
200 for( int x = 0; x < i_width; x += 8, i_mb++, pixoff += 8)
202 w->weightfn[8>>2]( buf, 8, &src[pixoff], i_stride, w, 8 );
203 int cmp = h->pixf.mbcmp[PIXEL_8x8]( buf, 8, &fenc_plane[pixoff], i_stride );
204 cost += X264_MIN( cmp, fenc->i_intra_cost[i_mb] );
206 cost += x264_weight_slice_header_cost( h, w, 0 );
209 for( int y = 0; y < i_lines; y += 8, pixoff = y*i_stride )
210 for( int x = 0; x < i_width; x += 8, i_mb++, pixoff += 8 )
212 int cmp = h->pixf.mbcmp[PIXEL_8x8]( &src[pixoff], i_stride, &fenc_plane[pixoff], i_stride );
213 cost += X264_MIN( cmp, fenc->i_intra_cost[i_mb] );
219 static NOINLINE unsigned int x264_weight_cost_chroma( x264_t *h, x264_frame_t *fenc, pixel *ref, x264_weight_t *w )
221 unsigned int cost = 0;
222 int i_stride = fenc->i_stride[1];
223 int i_offset = i_stride / 2;
224 int i_lines = fenc->i_lines[1];
225 int i_width = fenc->i_width[1];
226 pixel *src = ref + i_offset;
227 ALIGNED_ARRAY_16( pixel, buf, [8*16] );
229 int chromapix = h->luma2chroma_pixel[PIXEL_16x16];
230 int height = 16 >> h->mb.chroma_v_shift;
231 ALIGNED_16( static pixel flat[8] ) = {0};
234 for( int y = 0; y < i_lines; y += height, pixoff = y*i_stride )
235 for( int x = 0; x < i_width; x += 8, pixoff += 8 )
237 w->weightfn[8>>2]( buf, 8, &ref[pixoff], i_stride, w, height );
238 /* The naive and seemingly sensible algorithm is to use mbcmp as in luma.
239 * But testing shows that for chroma the DC coefficient is by far the most
240 * important part of the coding cost. Thus a more useful chroma weight is
241 * obtained by comparing each block's DC coefficient instead of the actual
244 * FIXME: add a (faster) asm sum function to replace sad. */
245 cost += abs( h->pixf.sad_aligned[chromapix]( buf, 8, flat, 0 ) -
246 h->pixf.sad_aligned[chromapix]( &src[pixoff], i_stride, flat, 0 ) );
248 cost += x264_weight_slice_header_cost( h, w, 1 );
251 for( int y = 0; y < i_lines; y += height, pixoff = y*i_stride )
252 for( int x = 0; x < i_width; x += 8, pixoff += 8 )
253 cost += abs( h->pixf.sad_aligned[chromapix]( &ref[pixoff], i_stride, flat, 0 ) -
254 h->pixf.sad_aligned[chromapix]( &src[pixoff], i_stride, flat, 0 ) );
259 static NOINLINE unsigned int x264_weight_cost_chroma444( x264_t *h, x264_frame_t *fenc, pixel *ref, x264_weight_t *w, int p )
261 unsigned int cost = 0;
262 int i_stride = fenc->i_stride[p];
263 int i_lines = fenc->i_lines[p];
264 int i_width = fenc->i_width[p];
265 pixel *src = fenc->plane[p];
266 ALIGNED_ARRAY_16( pixel, buf, [16*16] );
270 for( int y = 0; y < i_lines; y += 16, pixoff = y*i_stride )
271 for( int x = 0; x < i_width; x += 16, pixoff += 16 )
273 w->weightfn[16>>2]( buf, 16, &ref[pixoff], i_stride, w, 16 );
274 cost += h->pixf.mbcmp[PIXEL_16x16]( buf, 16, &src[pixoff], i_stride );
276 cost += x264_weight_slice_header_cost( h, w, 1 );
279 for( int y = 0; y < i_lines; y += 16, pixoff = y*i_stride )
280 for( int x = 0; x < i_width; x += 16, pixoff += 16 )
281 cost += h->pixf.mbcmp[PIXEL_16x16]( &ref[pixoff], i_stride, &src[pixoff], i_stride );
286 void x264_weights_analyse( x264_t *h, x264_frame_t *fenc, x264_frame_t *ref, int b_lookahead )
288 int i_delta_index = fenc->i_frame - ref->i_frame - 1;
289 /* epsilon is chosen to require at least a numerator of 127 (with denominator = 128) */
290 const float epsilon = 1.f/128.f;
291 x264_weight_t *weights = fenc->weight[0];
292 SET_WEIGHT( weights[0], 0, 1, 0, 0 );
293 SET_WEIGHT( weights[1], 0, 1, 0, 0 );
294 SET_WEIGHT( weights[2], 0, 1, 0, 0 );
295 int chroma_initted = 0;
296 /* Don't check chroma in lookahead, or if there wasn't a luma weight. */
297 for( int plane = 0; plane <= 2 && !( plane && ( !weights[0].weightfn || b_lookahead ) ); plane++ )
299 int cur_offset, start_offset, end_offset;
300 int minoff, minscale, mindenom;
301 unsigned int minscore, origscore;
303 float fenc_var = fenc->i_pixel_ssd[plane] + !ref->i_pixel_ssd[plane];
304 float ref_var = ref->i_pixel_ssd[plane] + !ref->i_pixel_ssd[plane];
305 float guess_scale = sqrtf( fenc_var / ref_var );
306 float fenc_mean = (float)fenc->i_pixel_sum[plane] / (fenc->i_lines[!!plane] * fenc->i_width[!!plane]) / (1 << (BIT_DEPTH - 8));
307 float ref_mean = (float) ref->i_pixel_sum[plane] / (fenc->i_lines[!!plane] * fenc->i_width[!!plane]) / (1 << (BIT_DEPTH - 8));
310 if( fabsf( ref_mean - fenc_mean ) < 0.5f && fabsf( 1.f - guess_scale ) < epsilon )
312 SET_WEIGHT( weights[plane], 0, 1, 0, 0 );
318 weights[plane].i_denom = 6;
319 weights[plane].i_scale = x264_clip3( round( guess_scale * 64 ), 0, 255 );
320 if( weights[plane].i_scale > 127 )
322 weights[1].weightfn = weights[2].weightfn = NULL;
327 x264_weight_get_h264( round( guess_scale * 128 ), 0, &weights[plane] );
330 mindenom = weights[plane].i_denom;
331 minscale = weights[plane].i_scale;
337 if( !fenc->b_intra_calculated )
339 x264_mb_analysis_t a;
340 x264_lowres_context_init( h, &a );
341 x264_slicetype_frame_cost( h, &a, &fenc, 0, 0, 0, 0 );
343 mcbuf = x264_weight_cost_init_luma( h, fenc, ref, h->mb.p_weight_buf[0] );
344 origscore = minscore = x264_weight_cost_luma( h, fenc, mcbuf, NULL );
350 mcbuf = x264_weight_cost_init_chroma444( h, fenc, ref, h->mb.p_weight_buf[0], plane );
351 origscore = minscore = x264_weight_cost_chroma444( h, fenc, mcbuf, NULL, plane );
355 pixel *dstu = h->mb.p_weight_buf[0];
356 pixel *dstv = h->mb.p_weight_buf[0]+fenc->i_stride[1]*fenc->i_lines[1];
357 if( !chroma_initted++ )
358 x264_weight_cost_init_chroma( h, fenc, ref, dstu, dstv );
359 mcbuf = plane == 1 ? dstu : dstv;
360 origscore = minscore = x264_weight_cost_chroma( h, fenc, mcbuf, NULL );
367 // This gives a slight improvement due to rounding errors but only tests one offset in lookahead.
368 // Currently only searches within +/- 1 of the best offset found so far.
369 // TODO: Try other offsets/multipliers/combinations thereof?
370 cur_offset = fenc_mean - ref_mean * minscale / (1 << mindenom) + 0.5f * b_lookahead;
371 start_offset = x264_clip3( cur_offset - !b_lookahead, -128, 127 );
372 end_offset = x264_clip3( cur_offset + !b_lookahead, -128, 127 );
373 for( int i_off = start_offset; i_off <= end_offset; i_off++ )
375 SET_WEIGHT( weights[plane], 1, minscale, mindenom, i_off );
380 s = x264_weight_cost_chroma444( h, fenc, mcbuf, &weights[plane], plane );
382 s = x264_weight_cost_chroma( h, fenc, mcbuf, &weights[plane] );
385 s = x264_weight_cost_luma( h, fenc, mcbuf, &weights[plane] );
386 COPY3_IF_LT( minscore, s, minoff, i_off, found, 1 );
388 // Don't check any more offsets if the previous one had a lower cost than the current one
389 if( minoff == start_offset && i_off != start_offset )
394 /* FIXME: More analysis can be done here on SAD vs. SATD termination. */
395 /* 0.2% termination derived experimentally to avoid weird weights in frames that are mostly intra. */
396 if( !found || (minscale == 1 << mindenom && minoff == 0) || (float)minscore / origscore > 0.998f )
398 SET_WEIGHT( weights[plane], 0, 1, 0, 0 );
402 SET_WEIGHT( weights[plane], 1, minscale, mindenom, minoff );
404 if( h->param.analyse.i_weighted_pred == X264_WEIGHTP_FAKE && weights[0].weightfn && !plane )
405 fenc->f_weighted_cost_delta[i_delta_index] = (float)minscore / origscore;
408 //FIXME, what is the correct way to deal with this?
409 if( weights[1].weightfn && weights[2].weightfn && weights[1].i_denom != weights[2].i_denom )
411 int denom = X264_MIN( weights[1].i_denom, weights[2].i_denom );
413 for( i = 1; i <= 2; i++ )
415 weights[i].i_scale = x264_clip3( weights[i].i_scale >> ( weights[i].i_denom - denom ), 0, 255 );
416 weights[i].i_denom = denom;
417 h->mc.weight_cache( h, &weights[i] );
421 if( weights[0].weightfn && b_lookahead )
423 //scale lowres in lookahead for slicetype_frame_cost
424 pixel *src = ref->buffer_lowres[0];
425 pixel *dst = h->mb.p_weight_buf[0];
426 int width = ref->i_width_lowres + PADH*2;
427 int height = ref->i_lines_lowres + PADV*2;
428 x264_weight_scale_plane( h, dst, ref->i_stride_lowres, src, ref->i_stride_lowres,
429 width, height, &weights[0] );
430 fenc->weighted[0] = h->mb.p_weight_buf[0] + PADH + ref->i_stride_lowres * PADV;
434 static void x264_slicetype_mb_cost( x264_t *h, x264_mb_analysis_t *a,
435 x264_frame_t **frames, int p0, int p1, int b,
436 int dist_scale_factor, int do_search[2], const x264_weight_t *w )
438 x264_frame_t *fref0 = frames[p0];
439 x264_frame_t *fref1 = frames[p1];
440 x264_frame_t *fenc = frames[b];
441 const int b_bidir = (b < p1);
442 const int i_mb_x = h->mb.i_mb_x;
443 const int i_mb_y = h->mb.i_mb_y;
444 const int i_mb_stride = h->mb.i_mb_width;
445 const int i_mb_xy = i_mb_x + i_mb_y * i_mb_stride;
446 const int i_stride = fenc->i_stride_lowres;
447 const int i_pel_offset = 8 * (i_mb_x + i_mb_y * i_stride);
448 const int i_bipred_weight = h->param.analyse.b_weighted_bipred ? 64 - (dist_scale_factor>>2) : 32;
449 int16_t (*fenc_mvs[2])[2] = { &fenc->lowres_mvs[0][b-p0-1][i_mb_xy], &fenc->lowres_mvs[1][p1-b-1][i_mb_xy] };
450 int (*fenc_costs[2]) = { &fenc->lowres_mv_costs[0][b-p0-1][i_mb_xy], &fenc->lowres_mv_costs[1][p1-b-1][i_mb_xy] };
451 int b_frame_score_mb = (i_mb_x > 0 && i_mb_x < h->mb.i_mb_width - 1 &&
452 i_mb_y > 0 && i_mb_y < h->mb.i_mb_height - 1) ||
453 h->mb.i_mb_width <= 2 || h->mb.i_mb_height <= 2;
455 ALIGNED_ARRAY_16( pixel, pix1,[9*FDEC_STRIDE] );
456 pixel *pix2 = pix1+8;
458 int i_bcost = COST_MAX;
461 h->mb.pic.p_fenc[0] = h->mb.pic.fenc_buf;
462 h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fenc[0], FENC_STRIDE, &fenc->lowres[0][i_pel_offset], i_stride, 8 );
465 goto lowres_intra_mb;
467 // no need for h->mb.mv_min[]
468 h->mb.mv_min_fpel[0] = -8*h->mb.i_mb_x - 4;
469 h->mb.mv_max_fpel[0] = 8*( h->mb.i_mb_width - h->mb.i_mb_x - 1 ) + 4;
470 h->mb.mv_min_spel[0] = 4*( h->mb.mv_min_fpel[0] - 8 );
471 h->mb.mv_max_spel[0] = 4*( h->mb.mv_max_fpel[0] + 8 );
472 if( h->mb.i_mb_x >= h->mb.i_mb_width - 2 )
474 h->mb.mv_min_fpel[1] = -8*h->mb.i_mb_y - 4;
475 h->mb.mv_max_fpel[1] = 8*( h->mb.i_mb_height - h->mb.i_mb_y - 1 ) + 4;
476 h->mb.mv_min_spel[1] = 4*( h->mb.mv_min_fpel[1] - 8 );
477 h->mb.mv_max_spel[1] = 4*( h->mb.mv_max_fpel[1] + 8 );
480 #define LOAD_HPELS_LUMA(dst, src) \
482 (dst)[0] = &(src)[0][i_pel_offset]; \
483 (dst)[1] = &(src)[1][i_pel_offset]; \
484 (dst)[2] = &(src)[2][i_pel_offset]; \
485 (dst)[3] = &(src)[3][i_pel_offset]; \
487 #define LOAD_WPELS_LUMA(dst,src) \
488 (dst) = &(src)[i_pel_offset];
490 #define CLIP_MV( mv ) \
492 mv[0] = x264_clip3( mv[0], h->mb.mv_min_spel[0], h->mb.mv_max_spel[0] ); \
493 mv[1] = x264_clip3( mv[1], h->mb.mv_min_spel[1], h->mb.mv_max_spel[1] ); \
495 #define TRY_BIDIR( mv0, mv1, penalty ) \
498 if( h->param.analyse.i_subpel_refine <= 1 ) \
500 int hpel_idx1 = (((mv0)[0]&2)>>1) + ((mv0)[1]&2); \
501 int hpel_idx2 = (((mv1)[0]&2)>>1) + ((mv1)[1]&2); \
502 pixel *src1 = m[0].p_fref[hpel_idx1] + ((mv0)[0]>>2) + ((mv0)[1]>>2) * m[0].i_stride[0]; \
503 pixel *src2 = m[1].p_fref[hpel_idx2] + ((mv1)[0]>>2) + ((mv1)[1]>>2) * m[1].i_stride[0]; \
504 h->mc.avg[PIXEL_8x8]( pix1, 16, src1, m[0].i_stride[0], src2, m[1].i_stride[0], i_bipred_weight ); \
508 int stride1 = 16, stride2 = 16; \
509 pixel *src1, *src2; \
510 src1 = h->mc.get_ref( pix1, &stride1, m[0].p_fref, m[0].i_stride[0], \
511 (mv0)[0], (mv0)[1], 8, 8, w ); \
512 src2 = h->mc.get_ref( pix2, &stride2, m[1].p_fref, m[1].i_stride[0], \
513 (mv1)[0], (mv1)[1], 8, 8, w ); \
514 h->mc.avg[PIXEL_8x8]( pix1, 16, src1, stride1, src2, stride2, i_bipred_weight ); \
516 i_cost = penalty * a->i_lambda + h->pixf.mbcmp[PIXEL_8x8]( \
517 m[0].p_fenc[0], FENC_STRIDE, pix1, 16 ); \
518 COPY2_IF_LT( i_bcost, i_cost, list_used, 3 ); \
521 m[0].i_pixel = PIXEL_8x8;
522 m[0].p_cost_mv = a->p_cost_mv;
523 m[0].i_stride[0] = i_stride;
524 m[0].p_fenc[0] = h->mb.pic.p_fenc[0];
527 LOAD_HPELS_LUMA( m[0].p_fref, fref0->lowres );
528 m[0].p_fref_w = m[0].p_fref[0];
530 LOAD_WPELS_LUMA( m[0].p_fref_w, fenc->weighted[0] );
534 int16_t *mvr = fref1->lowres_mvs[0][p1-p0-1][i_mb_xy];
535 ALIGNED_ARRAY_8( int16_t, dmv,[2],[2] );
537 m[1].i_pixel = PIXEL_8x8;
538 m[1].p_cost_mv = a->p_cost_mv;
539 m[1].i_stride[0] = i_stride;
540 m[1].p_fenc[0] = h->mb.pic.p_fenc[0];
542 m[1].weight = x264_weight_none;
543 LOAD_HPELS_LUMA( m[1].p_fref, fref1->lowres );
544 m[1].p_fref_w = m[1].p_fref[0];
546 dmv[0][0] = ( mvr[0] * dist_scale_factor + 128 ) >> 8;
547 dmv[0][1] = ( mvr[1] * dist_scale_factor + 128 ) >> 8;
548 dmv[1][0] = dmv[0][0] - mvr[0];
549 dmv[1][1] = dmv[0][1] - mvr[1];
552 if( h->param.analyse.i_subpel_refine <= 1 )
553 M64( dmv ) &= ~0x0001000100010001ULL; /* mv & ~1 */
555 TRY_BIDIR( dmv[0], dmv[1], 0 );
559 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 );
560 i_cost = h->pixf.mbcmp[PIXEL_8x8]( m[0].p_fenc[0], FENC_STRIDE, pix1, 16 );
561 COPY2_IF_LT( i_bcost, i_cost, list_used, 3 );
565 for( int l = 0; l < 1 + b_bidir; l++ )
570 int16_t (*fenc_mv)[2] = fenc_mvs[l];
571 ALIGNED_4( int16_t mvc[4][2] );
573 /* Reverse-order MV prediction. */
576 #define MVC(mv) { CP32( mvc[i_mvc], mv ); i_mvc++; }
577 if( i_mb_x < h->mb.i_mb_width - 1 )
579 if( i_mb_y < h->mb.i_mb_height - 1 )
581 MVC( fenc_mv[i_mb_stride] );
583 MVC( fenc_mv[i_mb_stride-1] );
584 if( i_mb_x < h->mb.i_mb_width - 1 )
585 MVC( fenc_mv[i_mb_stride+1] );
589 CP32( m[l].mvp, mvc[0] );
591 x264_median_mv( m[l].mvp, mvc[0], mvc[1], mvc[2] );
593 /* Fast skip for cases of near-zero residual. Shortcut: don't bother except in the mv0 case,
594 * since anything else is likely to have enough residual to not trigger the skip. */
595 if( !M32( m[l].mvp ) )
597 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] );
605 x264_me_search( h, &m[l], mvc, i_mvc );
606 m[l].cost -= a->p_cost_mv[0]; // remove mvcost from skip mbs
608 m[l].cost += 5 * a->i_lambda;
611 CP32( fenc_mvs[l], m[l].mv );
612 *fenc_costs[l] = m[l].cost;
616 CP32( m[l].mv, fenc_mvs[l] );
617 m[l].cost = *fenc_costs[l];
619 COPY2_IF_LT( i_bcost, m[l].cost, list_used, l+1 );
622 if( b_bidir && ( M32( m[0].mv ) || M32( m[1].mv ) ) )
623 TRY_BIDIR( m[0].mv, m[1].mv, 5 );
626 if( !fenc->b_intra_calculated )
628 ALIGNED_ARRAY_16( pixel, edge,[36] );
629 pixel *pix = &pix1[8+FDEC_STRIDE - 1];
630 pixel *src = &fenc->lowres[0][i_pel_offset - 1];
631 const int intra_penalty = 5 * a->i_lambda;
634 memcpy( pix-FDEC_STRIDE, src-i_stride, 17 * sizeof(pixel) );
635 for( int i = 0; i < 8; i++ )
636 pix[i*FDEC_STRIDE] = src[i*i_stride];
639 h->pixf.intra_mbcmp_x3_8x8c( h->mb.pic.p_fenc[0], pix, satds );
640 int i_icost = X264_MIN3( satds[0], satds[1], satds[2] );
642 if( h->param.analyse.i_subpel_refine > 1 )
644 h->predict_8x8c[I_PRED_CHROMA_P]( pix );
645 int satd = h->pixf.mbcmp[PIXEL_8x8]( pix, FDEC_STRIDE, h->mb.pic.p_fenc[0], FENC_STRIDE );
646 i_icost = X264_MIN( i_icost, satd );
647 h->predict_8x8_filter( pix, edge, ALL_NEIGHBORS, ALL_NEIGHBORS );
648 for( int i = 3; i < 9; i++ )
650 h->predict_8x8[i]( pix, edge );
651 satd = h->pixf.mbcmp[PIXEL_8x8]( pix, FDEC_STRIDE, h->mb.pic.p_fenc[0], FENC_STRIDE );
652 i_icost = X264_MIN( i_icost, satd );
656 i_icost += intra_penalty;
657 fenc->i_intra_cost[i_mb_xy] = i_icost;
658 int i_icost_aq = i_icost;
659 if( h->param.rc.i_aq_mode )
660 i_icost_aq = (i_icost_aq * fenc->i_inv_qscale_factor[i_mb_xy] + 128) >> 8;
661 fenc->i_row_satds[0][0][h->mb.i_mb_y] += i_icost_aq;
662 if( b_frame_score_mb )
664 fenc->i_cost_est[0][0] += i_icost;
665 fenc->i_cost_est_aq[0][0] += i_icost_aq;
669 /* forbid intra-mbs in B-frames, because it's rare and not worth checking */
670 /* FIXME: Should we still forbid them now that we cache intra scores? */
673 int i_icost = fenc->i_intra_cost[i_mb_xy];
674 int b_intra = i_icost < i_bcost;
680 if( b_frame_score_mb )
681 fenc->i_intra_mbs[b-p0] += b_intra;
684 /* In an I-frame, we've already added the results above in the intra section. */
687 int i_bcost_aq = i_bcost;
688 if( h->param.rc.i_aq_mode )
689 i_bcost_aq = (i_bcost_aq * fenc->i_inv_qscale_factor[i_mb_xy] + 128) >> 8;
690 fenc->i_row_satds[b-p0][p1-b][h->mb.i_mb_y] += i_bcost_aq;
691 if( b_frame_score_mb )
693 /* Don't use AQ-weighted costs for slicetype decision, only for ratecontrol. */
694 fenc->i_cost_est[b-p0][p1-b] += i_bcost;
695 fenc->i_cost_est_aq[b-p0][p1-b] += i_bcost_aq;
699 fenc->lowres_costs[b-p0][p1-b][i_mb_xy] = X264_MIN( i_bcost, LOWRES_COST_MASK ) + (list_used << LOWRES_COST_SHIFT);
704 (h->mb.i_mb_width > 2 && h->mb.i_mb_height > 2 ?\
705 (h->mb.i_mb_width - 2) * (h->mb.i_mb_height - 2) :\
706 h->mb.i_mb_width * h->mb.i_mb_height)
708 static int x264_slicetype_frame_cost( x264_t *h, x264_mb_analysis_t *a,
709 x264_frame_t **frames, int p0, int p1, int b,
710 int b_intra_penalty )
714 const x264_weight_t *w = x264_weight_none;
715 /* Check whether we already evaluated this frame
716 * If we have tried this frame as P, then we have also tried
717 * the preceding frames as B. (is this still true?) */
718 /* Also check that we already calculated the row SATDs for the current frame. */
719 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) )
720 i_score = frames[b]->i_cost_est[b-p0][p1-b];
723 int dist_scale_factor = 128;
724 int *row_satd = frames[b]->i_row_satds[b-p0][p1-b];
725 int *row_satd_intra = frames[b]->i_row_satds[0][0];
727 /* For each list, check to see whether we have lowres motion-searched this reference frame before. */
728 do_search[0] = b != p0 && frames[b]->lowres_mvs[0][b-p0-1][0][0] == 0x7FFF;
729 do_search[1] = b != p1 && frames[b]->lowres_mvs[1][p1-b-1][0][0] == 0x7FFF;
732 if( h->param.analyse.i_weighted_pred && b == p1 )
735 x264_weights_analyse( h, frames[b], frames[p0], 1 );
736 w = frames[b]->weight[0];
738 frames[b]->lowres_mvs[0][b-p0-1][0][0] = 0;
740 if( do_search[1] ) frames[b]->lowres_mvs[1][p1-b-1][0][0] = 0;
743 frames[b]->i_intra_mbs[b-p0] = 0;
744 if( !frames[b]->b_intra_calculated )
746 frames[b]->i_cost_est[0][0] = 0;
747 frames[b]->i_cost_est_aq[0][0] = 0;
750 dist_scale_factor = ( ((b-p0) << 8) + ((p1-p0) >> 1) ) / (p1-p0);
752 frames[b]->i_cost_est[b-p0][p1-b] = 0;
753 frames[b]->i_cost_est_aq[b-p0][p1-b] = 0;
755 /* Lowres lookahead goes backwards because the MVs are used as predictors in the main encode.
756 * This considerably improves MV prediction overall. */
758 /* The edge mbs seem to reduce the predictive quality of the
759 * whole frame's score, but are needed for a spatial distribution. */
760 if( h->param.rc.b_mb_tree || h->param.rc.i_vbv_buffer_size ||
761 h->mb.i_mb_width <= 2 || h->mb.i_mb_height <= 2 )
763 for( h->mb.i_mb_y = h->mb.i_mb_height - 1; h->mb.i_mb_y >= 0; h->mb.i_mb_y-- )
765 row_satd[h->mb.i_mb_y] = 0;
766 if( !frames[b]->b_intra_calculated )
767 row_satd_intra[h->mb.i_mb_y] = 0;
768 for( h->mb.i_mb_x = h->mb.i_mb_width - 1; h->mb.i_mb_x >= 0; h->mb.i_mb_x-- )
769 x264_slicetype_mb_cost( h, a, frames, p0, p1, b, dist_scale_factor, do_search, w );
774 for( h->mb.i_mb_y = h->mb.i_mb_height - 2; h->mb.i_mb_y >= 1; h->mb.i_mb_y-- )
775 for( h->mb.i_mb_x = h->mb.i_mb_width - 2; h->mb.i_mb_x >= 1; h->mb.i_mb_x-- )
776 x264_slicetype_mb_cost( h, a, frames, p0, p1, b, dist_scale_factor, do_search, w );
779 i_score = frames[b]->i_cost_est[b-p0][p1-b];
781 i_score = (uint64_t)i_score * 100 / (120 + h->param.i_bframe_bias);
783 frames[b]->b_intra_calculated = 1;
785 frames[b]->i_cost_est[b-p0][p1-b] = i_score;
789 if( b_intra_penalty )
791 // arbitrary penalty for I-blocks after B-frames
793 i_score += (uint64_t)i_score * frames[b]->i_intra_mbs[b-p0] / (nmb * 8);
798 /* If MB-tree changes the quantizers, we need to recalculate the frame cost without
799 * re-running lookahead. */
800 static int x264_slicetype_frame_cost_recalculate( x264_t *h, x264_frame_t **frames, int p0, int p1, int b )
803 int *row_satd = frames[b]->i_row_satds[b-p0][p1-b];
804 float *qp_offset = IS_X264_TYPE_B(frames[b]->i_type) ? frames[b]->f_qp_offset_aq : frames[b]->f_qp_offset;
806 for( h->mb.i_mb_y = h->mb.i_mb_height - 1; h->mb.i_mb_y >= 0; h->mb.i_mb_y-- )
808 row_satd[ h->mb.i_mb_y ] = 0;
809 for( h->mb.i_mb_x = h->mb.i_mb_width - 1; h->mb.i_mb_x >= 0; h->mb.i_mb_x-- )
811 int i_mb_xy = h->mb.i_mb_x + h->mb.i_mb_y*h->mb.i_mb_stride;
812 int i_mb_cost = frames[b]->lowres_costs[b-p0][p1-b][i_mb_xy] & LOWRES_COST_MASK;
813 float qp_adj = qp_offset[i_mb_xy];
814 i_mb_cost = (i_mb_cost * x264_exp2fix8(qp_adj) + 128) >> 8;
815 row_satd[ h->mb.i_mb_y ] += i_mb_cost;
816 if( (h->mb.i_mb_y > 0 && h->mb.i_mb_y < h->mb.i_mb_height - 1 &&
817 h->mb.i_mb_x > 0 && h->mb.i_mb_x < h->mb.i_mb_width - 1) ||
818 h->mb.i_mb_width <= 2 || h->mb.i_mb_height <= 2 )
820 i_score += i_mb_cost;
827 static void x264_macroblock_tree_finish( x264_t *h, x264_frame_t *frame, float average_duration, int ref0_distance )
829 int fps_factor = round( CLIP_DURATION(average_duration) / CLIP_DURATION(frame->f_duration) * 256 );
830 float weightdelta = 0.0;
831 if( ref0_distance && frame->f_weighted_cost_delta[ref0_distance-1] > 0 )
832 weightdelta = (1.0 - frame->f_weighted_cost_delta[ref0_distance-1]);
834 /* Allow the strength to be adjusted via qcompress, since the two
835 * concepts are very similar. */
836 float strength = 5.0f * (1.0f - h->param.rc.f_qcompress);
837 for( int mb_index = 0; mb_index < h->mb.i_mb_count; mb_index++ )
839 int intra_cost = (frame->i_intra_cost[mb_index] * frame->i_inv_qscale_factor[mb_index] + 128) >> 8;
842 int propagate_cost = (frame->i_propagate_cost[mb_index] * fps_factor + 128) >> 8;
843 float log2_ratio = x264_log2(intra_cost + propagate_cost) - x264_log2(intra_cost) + weightdelta;
844 frame->f_qp_offset[mb_index] = frame->f_qp_offset_aq[mb_index] - strength * log2_ratio;
849 static void x264_macroblock_tree_propagate( x264_t *h, x264_frame_t **frames, float average_duration, int p0, int p1, int b, int referenced )
851 uint16_t *ref_costs[2] = {frames[p0]->i_propagate_cost,frames[p1]->i_propagate_cost};
852 int dist_scale_factor = ( ((b-p0) << 8) + ((p1-p0) >> 1) ) / (p1-p0);
853 int i_bipred_weight = h->param.analyse.b_weighted_bipred ? 64 - (dist_scale_factor>>2) : 32;
854 int16_t (*mvs[2])[2] = { frames[b]->lowres_mvs[0][b-p0-1], frames[b]->lowres_mvs[1][p1-b-1] };
855 int bipred_weights[2] = {i_bipred_weight, 64 - i_bipred_weight};
856 int *buf = h->scratch_buffer;
857 uint16_t *propagate_cost = frames[b]->i_propagate_cost;
860 float fps_factor = CLIP_DURATION(frames[b]->f_duration) / CLIP_DURATION(average_duration);
862 /* For non-reffed frames the source costs are always zero, so just memset one row and re-use it. */
864 memset( frames[b]->i_propagate_cost, 0, h->mb.i_mb_width * sizeof(uint16_t) );
866 for( h->mb.i_mb_y = 0; h->mb.i_mb_y < h->mb.i_mb_height; h->mb.i_mb_y++ )
868 int mb_index = h->mb.i_mb_y*h->mb.i_mb_stride;
869 h->mc.mbtree_propagate_cost( buf, propagate_cost,
870 frames[b]->i_intra_cost+mb_index, frames[b]->lowres_costs[b-p0][p1-b]+mb_index,
871 frames[b]->i_inv_qscale_factor+mb_index, &fps_factor, h->mb.i_mb_width );
873 propagate_cost += h->mb.i_mb_width;
874 for( h->mb.i_mb_x = 0; h->mb.i_mb_x < h->mb.i_mb_width; h->mb.i_mb_x++, mb_index++ )
876 int propagate_amount = buf[h->mb.i_mb_x];
877 /* Don't propagate for an intra block. */
878 if( propagate_amount > 0 )
880 /* Access width-2 bitfield. */
881 int lists_used = frames[b]->lowres_costs[b-p0][p1-b][mb_index] >> LOWRES_COST_SHIFT;
882 /* Follow the MVs to the previous frame(s). */
883 for( int list = 0; list < 2; list++ )
884 if( (lists_used >> list)&1 )
886 #define CLIP_ADD(s,x) (s) = X264_MIN((s)+(x),(1<<16)-1)
887 int listamount = propagate_amount;
888 /* Apply bipred weighting. */
889 if( lists_used == 3 )
890 listamount = (listamount * bipred_weights[list] + 32) >> 6;
892 /* Early termination for simple case of mv0. */
893 if( !M32( mvs[list][mb_index] ) )
895 CLIP_ADD( ref_costs[list][mb_index], listamount );
899 int x = mvs[list][mb_index][0];
900 int y = mvs[list][mb_index][1];
901 int mbx = (x>>5)+h->mb.i_mb_x;
902 int mby = (y>>5)+h->mb.i_mb_y;
903 int idx0 = mbx + mby * h->mb.i_mb_stride;
905 int idx2 = idx0 + h->mb.i_mb_stride;
906 int idx3 = idx0 + h->mb.i_mb_stride + 1;
909 int idx0weight = (32-y)*(32-x);
910 int idx1weight = (32-y)*x;
911 int idx2weight = y*(32-x);
912 int idx3weight = y*x;
914 /* We could just clip the MVs, but pixels that lie outside the frame probably shouldn't
916 if( mbx < h->mb.i_mb_width-1 && mby < h->mb.i_mb_height-1 && mbx >= 0 && mby >= 0 )
918 CLIP_ADD( ref_costs[list][idx0], (listamount*idx0weight+512)>>10 );
919 CLIP_ADD( ref_costs[list][idx1], (listamount*idx1weight+512)>>10 );
920 CLIP_ADD( ref_costs[list][idx2], (listamount*idx2weight+512)>>10 );
921 CLIP_ADD( ref_costs[list][idx3], (listamount*idx3weight+512)>>10 );
923 else /* Check offsets individually */
925 if( mbx < h->mb.i_mb_width && mby < h->mb.i_mb_height && mbx >= 0 && mby >= 0 )
926 CLIP_ADD( ref_costs[list][idx0], (listamount*idx0weight+512)>>10 );
927 if( mbx+1 < h->mb.i_mb_width && mby < h->mb.i_mb_height && mbx+1 >= 0 && mby >= 0 )
928 CLIP_ADD( ref_costs[list][idx1], (listamount*idx1weight+512)>>10 );
929 if( mbx < h->mb.i_mb_width && mby+1 < h->mb.i_mb_height && mbx >= 0 && mby+1 >= 0 )
930 CLIP_ADD( ref_costs[list][idx2], (listamount*idx2weight+512)>>10 );
931 if( mbx+1 < h->mb.i_mb_width && mby+1 < h->mb.i_mb_height && mbx+1 >= 0 && mby+1 >= 0 )
932 CLIP_ADD( ref_costs[list][idx3], (listamount*idx3weight+512)>>10 );
939 if( h->param.rc.i_vbv_buffer_size && h->param.rc.i_lookahead && referenced )
940 x264_macroblock_tree_finish( h, frames[b], average_duration, b == p1 ? b - p0 : 0 );
943 static void x264_macroblock_tree( x264_t *h, x264_mb_analysis_t *a, x264_frame_t **frames, int num_frames, int b_intra )
946 int last_nonb, cur_nonb = 1;
950 float total_duration = 0.0;
951 for( int j = 0; j <= num_frames; j++ )
952 total_duration += frames[j]->f_duration;
953 float average_duration = total_duration / (num_frames + 1);
958 x264_slicetype_frame_cost( h, a, frames, 0, 0, 0, 0 );
960 while( i > 0 && frames[i]->i_type == X264_TYPE_B )
964 /* Lookaheadless MB-tree is not a theoretically distinct case; the same extrapolation could
965 * be applied to the end of a lookahead buffer of any size. However, it's most needed when
966 * lookahead=0, so that's what's currently implemented. */
967 if( !h->param.rc.i_lookahead )
971 memset( frames[0]->i_propagate_cost, 0, h->mb.i_mb_count * sizeof(uint16_t) );
972 memcpy( frames[0]->f_qp_offset, frames[0]->f_qp_offset_aq, h->mb.i_mb_count * sizeof(float) );
975 XCHG( uint16_t*, frames[last_nonb]->i_propagate_cost, frames[0]->i_propagate_cost );
976 memset( frames[0]->i_propagate_cost, 0, h->mb.i_mb_count * sizeof(uint16_t) );
980 if( last_nonb < idx )
982 memset( frames[last_nonb]->i_propagate_cost, 0, h->mb.i_mb_count * sizeof(uint16_t) );
988 while( frames[cur_nonb]->i_type == X264_TYPE_B && cur_nonb > 0 )
992 x264_slicetype_frame_cost( h, a, frames, cur_nonb, last_nonb, last_nonb, 0 );
993 memset( frames[cur_nonb]->i_propagate_cost, 0, h->mb.i_mb_count * sizeof(uint16_t) );
994 bframes = last_nonb - cur_nonb - 1;
995 if( h->param.i_bframe_pyramid && bframes > 1 )
997 int middle = (bframes + 1)/2 + cur_nonb;
998 x264_slicetype_frame_cost( h, a, frames, cur_nonb, last_nonb, middle, 0 );
999 memset( frames[middle]->i_propagate_cost, 0, h->mb.i_mb_count * sizeof(uint16_t) );
1000 while( i > cur_nonb )
1002 int p0 = i > middle ? middle : cur_nonb;
1003 int p1 = i < middle ? middle : last_nonb;
1006 x264_slicetype_frame_cost( h, a, frames, p0, p1, i, 0 );
1007 x264_macroblock_tree_propagate( h, frames, average_duration, p0, p1, i, 0 );
1011 x264_macroblock_tree_propagate( h, frames, average_duration, cur_nonb, last_nonb, middle, 1 );
1015 while( i > cur_nonb )
1017 x264_slicetype_frame_cost( h, a, frames, cur_nonb, last_nonb, i, 0 );
1018 x264_macroblock_tree_propagate( h, frames, average_duration, cur_nonb, last_nonb, i, 0 );
1022 x264_macroblock_tree_propagate( h, frames, average_duration, cur_nonb, last_nonb, last_nonb, 1 );
1023 last_nonb = cur_nonb;
1026 if( !h->param.rc.i_lookahead )
1028 x264_macroblock_tree_propagate( h, frames, average_duration, 0, last_nonb, last_nonb, 1 );
1029 XCHG( uint16_t*, frames[last_nonb]->i_propagate_cost, frames[0]->i_propagate_cost );
1032 x264_macroblock_tree_finish( h, frames[last_nonb], average_duration, last_nonb );
1033 if( h->param.i_bframe_pyramid && bframes > 1 && !h->param.rc.i_vbv_buffer_size )
1034 x264_macroblock_tree_finish( h, frames[last_nonb+(bframes+1)/2], average_duration, 0 );
1037 static int x264_vbv_frame_cost( x264_t *h, x264_mb_analysis_t *a, x264_frame_t **frames, int p0, int p1, int b )
1039 int cost = x264_slicetype_frame_cost( h, a, frames, p0, p1, b, 0 );
1040 if( h->param.rc.i_aq_mode )
1042 if( h->param.rc.b_mb_tree )
1043 return x264_slicetype_frame_cost_recalculate( h, frames, p0, p1, b );
1045 return frames[b]->i_cost_est_aq[b-p0][p1-b];
1050 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 )
1052 cur_frame->i_cpb_delay = *i_cpb_delay;
1053 cur_frame->i_dpb_output_delay = cur_frame->i_field_cnt - *i_coded_fields;
1055 // add a correction term for frame reordering
1056 cur_frame->i_dpb_output_delay += h->sps->vui.i_num_reorder_frames*2;
1058 // fix possible negative dpb_output_delay because of pulldown changes and reordering
1059 if( cur_frame->i_dpb_output_delay < 0 )
1061 cur_frame->i_cpb_delay += cur_frame->i_dpb_output_delay;
1062 cur_frame->i_dpb_output_delay = 0;
1064 prev_frame->i_cpb_duration += cur_frame->i_dpb_output_delay;
1067 if( cur_frame->b_keyframe )
1070 *i_cpb_delay += cur_frame->i_duration;
1071 *i_coded_fields += cur_frame->i_duration;
1072 cur_frame->i_cpb_duration = cur_frame->i_duration;
1075 static void x264_vbv_lookahead( x264_t *h, x264_mb_analysis_t *a, x264_frame_t **frames, int num_frames, int keyframe )
1077 int last_nonb = 0, cur_nonb = 1, idx = 0;
1078 x264_frame_t *prev_frame = NULL;
1079 int prev_frame_idx = 0;
1080 while( cur_nonb < num_frames && frames[cur_nonb]->i_type == X264_TYPE_B )
1082 int next_nonb = keyframe ? last_nonb : cur_nonb;
1084 if( frames[cur_nonb]->i_coded_fields_lookahead >= 0 )
1086 h->i_coded_fields_lookahead = frames[cur_nonb]->i_coded_fields_lookahead;
1087 h->i_cpb_delay_lookahead = frames[cur_nonb]->i_cpb_delay_lookahead;
1090 while( cur_nonb < num_frames )
1092 /* P/I cost: This shouldn't include the cost of next_nonb */
1093 if( next_nonb != cur_nonb )
1095 int p0 = IS_X264_TYPE_I( frames[cur_nonb]->i_type ) ? cur_nonb : last_nonb;
1096 frames[next_nonb]->i_planned_satd[idx] = x264_vbv_frame_cost( h, a, frames, p0, cur_nonb, cur_nonb );
1097 frames[next_nonb]->i_planned_type[idx] = frames[cur_nonb]->i_type;
1098 frames[cur_nonb]->i_coded_fields_lookahead = h->i_coded_fields_lookahead;
1099 frames[cur_nonb]->i_cpb_delay_lookahead = h->i_cpb_delay_lookahead;
1100 x264_calculate_durations( h, frames[cur_nonb], prev_frame, &h->i_cpb_delay_lookahead, &h->i_coded_fields_lookahead );
1103 frames[next_nonb]->f_planned_cpb_duration[prev_frame_idx] = (double)prev_frame->i_cpb_duration *
1104 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1106 frames[next_nonb]->f_planned_cpb_duration[idx] = (double)frames[cur_nonb]->i_cpb_duration *
1107 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1108 prev_frame = frames[cur_nonb];
1109 prev_frame_idx = idx;
1112 /* Handle the B-frames: coded order */
1113 for( int i = last_nonb+1; i < cur_nonb; i++, idx++ )
1115 frames[next_nonb]->i_planned_satd[idx] = x264_vbv_frame_cost( h, a, frames, last_nonb, cur_nonb, i );
1116 frames[next_nonb]->i_planned_type[idx] = X264_TYPE_B;
1117 frames[i]->i_coded_fields_lookahead = h->i_coded_fields_lookahead;
1118 frames[i]->i_cpb_delay_lookahead = h->i_cpb_delay_lookahead;
1119 x264_calculate_durations( h, frames[i], prev_frame, &h->i_cpb_delay_lookahead, &h->i_coded_fields_lookahead );
1122 frames[next_nonb]->f_planned_cpb_duration[prev_frame_idx] = (double)prev_frame->i_cpb_duration *
1123 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1125 frames[next_nonb]->f_planned_cpb_duration[idx] = (double)frames[i]->i_cpb_duration *
1126 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1127 prev_frame = frames[i];
1128 prev_frame_idx = idx;
1130 last_nonb = cur_nonb;
1132 while( cur_nonb <= num_frames && frames[cur_nonb]->i_type == X264_TYPE_B )
1135 frames[next_nonb]->i_planned_type[idx] = X264_TYPE_AUTO;
1138 static int x264_slicetype_path_cost( x264_t *h, x264_mb_analysis_t *a, x264_frame_t **frames, char *path, int threshold )
1143 path--; /* Since the 1st path element is really the second frame */
1147 /* Find the location of the next P-frame. */
1148 while( path[next_p] != 'P' )
1151 /* Add the cost of the P-frame found above */
1152 cost += x264_slicetype_frame_cost( h, a, frames, cur_p, next_p, next_p, 0 );
1153 /* Early terminate if the cost we have found is larger than the best path cost so far */
1154 if( cost > threshold )
1157 if( h->param.i_bframe_pyramid && next_p - cur_p > 2 )
1159 int middle = cur_p + (next_p - cur_p)/2;
1160 cost += x264_slicetype_frame_cost( h, a, frames, cur_p, next_p, middle, 0 );
1161 for( int next_b = loc; next_b < middle && cost < threshold; next_b++ )
1162 cost += x264_slicetype_frame_cost( h, a, frames, cur_p, middle, next_b, 0 );
1163 for( int next_b = middle+1; next_b < next_p && cost < threshold; next_b++ )
1164 cost += x264_slicetype_frame_cost( h, a, frames, middle, next_p, next_b, 0 );
1167 for( int next_b = loc; next_b < next_p && cost < threshold; next_b++ )
1168 cost += x264_slicetype_frame_cost( h, a, frames, cur_p, next_p, next_b, 0 );
1176 /* Viterbi/trellis slicetype decision algorithm. */
1177 /* Uses strings due to the fact that the speed of the control functions is
1178 negligible compared to the cost of running slicetype_frame_cost, and because
1179 it makes debugging easier. */
1180 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+1] )
1182 char paths[2][X264_LOOKAHEAD_MAX+1];
1183 int num_paths = X264_MIN( h->param.i_bframe+1, length );
1184 int best_cost = COST_MAX;
1187 /* Iterate over all currently possible paths */
1188 for( int path = 0; path < num_paths; path++ )
1190 /* Add suffixes to the current path */
1191 int len = length - (path + 1);
1192 memcpy( paths[idx], best_paths[len % (X264_BFRAME_MAX+1)], len );
1193 memset( paths[idx]+len, 'B', path );
1194 strcpy( paths[idx]+len+path, "P" );
1196 /* Calculate the actual cost of the current path */
1197 int cost = x264_slicetype_path_cost( h, a, frames, paths[idx], best_cost );
1198 if( cost < best_cost )
1205 /* Store the best path. */
1206 memcpy( best_paths[length % (X264_BFRAME_MAX+1)], paths[idx^1], length );
1209 static int scenecut_internal( x264_t *h, x264_mb_analysis_t *a, x264_frame_t **frames, int p0, int p1, int real_scenecut )
1211 x264_frame_t *frame = frames[p1];
1213 /* Don't do scenecuts on the right view of a frame-packed video. */
1214 if( real_scenecut && h->param.i_frame_packing == 5 && (frame->i_frame&1) )
1217 x264_slicetype_frame_cost( h, a, frames, p0, p1, p1, 0 );
1219 int icost = frame->i_cost_est[0][0];
1220 int pcost = frame->i_cost_est[p1-p0][0];
1222 int i_gop_size = frame->i_frame - h->lookahead->i_last_keyframe;
1223 float f_thresh_max = h->param.i_scenecut_threshold / 100.0;
1224 /* magic numbers pulled out of thin air */
1225 float f_thresh_min = f_thresh_max * 0.25;
1228 if( h->param.i_keyint_min == h->param.i_keyint_max )
1229 f_thresh_min = f_thresh_max;
1230 if( i_gop_size <= h->param.i_keyint_min / 4 || h->param.b_intra_refresh )
1231 f_bias = f_thresh_min / 4;
1232 else if( i_gop_size <= h->param.i_keyint_min )
1233 f_bias = f_thresh_min * i_gop_size / h->param.i_keyint_min;
1236 f_bias = f_thresh_min
1237 + ( f_thresh_max - f_thresh_min )
1238 * ( i_gop_size - h->param.i_keyint_min )
1239 / ( h->param.i_keyint_max - h->param.i_keyint_min );
1242 res = pcost >= (1.0 - f_bias) * icost;
1243 if( res && real_scenecut )
1245 int imb = frame->i_intra_mbs[p1-p0];
1246 int pmb = NUM_MBS - imb;
1247 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",
1249 icost, pcost, 1. - (double)pcost / icost,
1250 f_bias, i_gop_size, imb, pmb );
1255 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 )
1257 /* Only do analysis during a normal scenecut check. */
1258 if( real_scenecut && h->param.i_bframe )
1260 int origmaxp1 = p0 + 1;
1261 /* Look ahead to avoid coding short flashes as scenecuts. */
1262 if( h->param.i_bframe_adaptive == X264_B_ADAPT_TRELLIS )
1263 /* Don't analyse any more frames than the trellis would have covered. */
1264 origmaxp1 += h->param.i_bframe;
1267 int maxp1 = X264_MIN( origmaxp1, num_frames );
1269 /* Where A and B are scenes: AAAAAABBBAAAAAA
1270 * If BBB is shorter than (maxp1-p0), it is detected as a flash
1271 * and not considered a scenecut. */
1272 for( int curp1 = p1; curp1 <= maxp1; curp1++ )
1273 if( !scenecut_internal( h, a, frames, p0, curp1, 0 ) )
1274 /* Any frame in between p0 and cur_p1 cannot be a real scenecut. */
1275 for( int i = curp1; i > p0; i-- )
1276 frames[i]->b_scenecut = 0;
1278 /* Where A-F are scenes: AAAAABBCCDDEEFFFFFF
1279 * If each of BB ... EE are shorter than (maxp1-p0), they are
1280 * detected as flashes and not considered scenecuts.
1281 * Instead, the first F frame becomes a scenecut.
1282 * If the video ends before F, no frame becomes a scenecut. */
1283 for( int curp0 = p0; curp0 <= maxp1; curp0++ )
1284 if( origmaxp1 > i_max_search || (curp0 < maxp1 && scenecut_internal( h, a, frames, curp0, maxp1, 0 )) )
1285 /* If cur_p0 is the p0 of a scenecut, it cannot be the p1 of a scenecut. */
1286 frames[curp0]->b_scenecut = 0;
1289 /* Ignore frames that are part of a flash, i.e. cannot be real scenecuts. */
1290 if( !frames[p1]->b_scenecut )
1292 return scenecut_internal( h, a, frames, p0, p1, real_scenecut );
1295 void x264_slicetype_analyse( x264_t *h, int keyframe )
1297 x264_mb_analysis_t a;
1298 x264_frame_t *frames[X264_LOOKAHEAD_MAX+3] = { NULL, };
1299 int num_frames, orig_num_frames, keyint_limit, framecnt;
1300 int i_mb_count = NUM_MBS;
1301 int cost1p0, cost2p0, cost1b1, cost2p1;
1302 int i_max_search = X264_MIN( h->lookahead->next.i_size, X264_LOOKAHEAD_MAX );
1303 int vbv_lookahead = h->param.rc.i_vbv_buffer_size && h->param.rc.i_lookahead;
1304 if( h->param.b_deterministic )
1305 i_max_search = X264_MIN( i_max_search, h->lookahead->i_slicetype_length + !keyframe );
1307 assert( h->frames.b_have_lowres );
1309 if( !h->lookahead->last_nonb )
1311 frames[0] = h->lookahead->last_nonb;
1312 for( framecnt = 0; framecnt < i_max_search && h->lookahead->next.list[framecnt]->i_type == X264_TYPE_AUTO; framecnt++ )
1313 frames[framecnt+1] = h->lookahead->next.list[framecnt];
1315 x264_lowres_context_init( h, &a );
1319 if( h->param.rc.b_mb_tree )
1320 x264_macroblock_tree( h, &a, frames, 0, keyframe );
1324 keyint_limit = h->param.i_keyint_max - frames[0]->i_frame + h->lookahead->i_last_keyframe - 1;
1325 orig_num_frames = num_frames = h->param.b_intra_refresh ? framecnt : X264_MIN( framecnt, keyint_limit );
1327 /* This is important psy-wise: if we have a non-scenecut keyframe,
1328 * there will be significant visual artifacts if the frames just before
1329 * go down in quality due to being referenced less, despite it being
1330 * more RD-optimal. */
1331 if( (h->param.analyse.b_psy && h->param.rc.b_mb_tree) || vbv_lookahead )
1332 num_frames = framecnt;
1333 else if( h->param.b_open_gop && num_frames < framecnt )
1335 else if( num_frames == 0 )
1337 frames[1]->i_type = X264_TYPE_I;
1341 int num_bframes = 0;
1342 int num_analysed_frames = num_frames;
1344 if( h->param.i_scenecut_threshold && scenecut( h, &a, frames, 0, 1, 1, orig_num_frames, i_max_search ) )
1346 frames[1]->i_type = X264_TYPE_I;
1350 if( h->param.i_bframe )
1352 if( h->param.i_bframe_adaptive == X264_B_ADAPT_TRELLIS )
1354 if( num_frames > 1 )
1356 char best_paths[X264_BFRAME_MAX+1][X264_LOOKAHEAD_MAX+1] = {"","P"};
1357 int best_path_index = num_frames % (X264_BFRAME_MAX+1);
1359 /* Perform the frametype analysis. */
1360 for( int j = 2; j <= num_frames; j++ )
1361 x264_slicetype_path( h, &a, frames, j, best_paths );
1363 num_bframes = strspn( best_paths[best_path_index], "B" );
1364 /* Load the results of the analysis into the frame types. */
1365 for( int j = 1; j < num_frames; j++ )
1366 frames[j]->i_type = best_paths[best_path_index][j-1] == 'B' ? X264_TYPE_B : X264_TYPE_P;
1368 frames[num_frames]->i_type = X264_TYPE_P;
1370 else if( h->param.i_bframe_adaptive == X264_B_ADAPT_FAST )
1372 for( int i = 0; i <= num_frames-2; )
1374 cost2p1 = x264_slicetype_frame_cost( h, &a, frames, i+0, i+2, i+2, 1 );
1375 if( frames[i+2]->i_intra_mbs[2] > i_mb_count / 2 )
1377 frames[i+1]->i_type = X264_TYPE_P;
1378 frames[i+2]->i_type = X264_TYPE_P;
1383 cost1b1 = x264_slicetype_frame_cost( h, &a, frames, i+0, i+2, i+1, 0 );
1384 cost1p0 = x264_slicetype_frame_cost( h, &a, frames, i+0, i+1, i+1, 0 );
1385 cost2p0 = x264_slicetype_frame_cost( h, &a, frames, i+1, i+2, i+2, 0 );
1387 if( cost1p0 + cost2p0 < cost1b1 + cost2p1 )
1389 frames[i+1]->i_type = X264_TYPE_P;
1394 // arbitrary and untuned
1395 #define INTER_THRESH 300
1396 #define P_SENS_BIAS (50 - h->param.i_bframe_bias)
1397 frames[i+1]->i_type = X264_TYPE_B;
1400 for( j = i+2; j <= X264_MIN( i+h->param.i_bframe, num_frames-1 ); j++ )
1402 int pthresh = X264_MAX(INTER_THRESH - P_SENS_BIAS * (j-i-1), INTER_THRESH/10);
1403 int pcost = x264_slicetype_frame_cost( h, &a, frames, i+0, j+1, j+1, 1 );
1404 if( pcost > pthresh*i_mb_count || frames[j+1]->i_intra_mbs[j-i+1] > i_mb_count/3 )
1406 frames[j]->i_type = X264_TYPE_B;
1408 frames[j]->i_type = X264_TYPE_P;
1411 frames[num_frames]->i_type = X264_TYPE_P;
1413 while( num_bframes < num_frames && frames[num_bframes+1]->i_type == X264_TYPE_B )
1418 num_bframes = X264_MIN(num_frames-1, h->param.i_bframe);
1419 for( int j = 1; j < num_frames; j++ )
1420 frames[j]->i_type = (j%(num_bframes+1)) ? X264_TYPE_B : X264_TYPE_P;
1421 frames[num_frames]->i_type = X264_TYPE_P;
1424 /* Check scenecut on the first minigop. */
1425 for( int j = 1; j < num_bframes+1; j++ )
1426 if( h->param.i_scenecut_threshold && scenecut( h, &a, frames, j, j+1, 0, orig_num_frames, i_max_search ) )
1428 frames[j]->i_type = X264_TYPE_P;
1429 num_analysed_frames = j;
1433 reset_start = keyframe ? 1 : X264_MIN( num_bframes+2, num_analysed_frames+1 );
1437 for( int j = 1; j <= num_frames; j++ )
1438 frames[j]->i_type = X264_TYPE_P;
1439 reset_start = !keyframe + 1;
1443 /* Perform the actual macroblock tree analysis.
1444 * Don't go farther than the maximum keyframe interval; this helps in short GOPs. */
1445 if( h->param.rc.b_mb_tree )
1446 x264_macroblock_tree( h, &a, frames, X264_MIN(num_frames, h->param.i_keyint_max), keyframe );
1448 /* Enforce keyframe limit. */
1449 if( !h->param.b_intra_refresh )
1450 for( int i = keyint_limit+1; i <= num_frames; i += h->param.i_keyint_max )
1452 frames[i]->i_type = X264_TYPE_I;
1453 reset_start = X264_MIN( reset_start, i+1 );
1454 if( h->param.b_open_gop && h->param.b_bluray_compat )
1455 while( IS_X264_TYPE_B( frames[i-1]->i_type ) )
1460 x264_vbv_lookahead( h, &a, frames, num_frames, keyframe );
1462 /* Restore frametypes for all frames that haven't actually been decided yet. */
1463 for( int j = reset_start; j <= num_frames; j++ )
1464 frames[j]->i_type = X264_TYPE_AUTO;
1467 void x264_slicetype_decide( x264_t *h )
1469 x264_frame_t *frames[X264_BFRAME_MAX+2];
1474 if( !h->lookahead->next.i_size )
1477 int lookahead_size = h->lookahead->next.i_size;
1479 for( int i = 0; i < h->lookahead->next.i_size; i++ )
1481 if( h->param.b_vfr_input )
1483 if( lookahead_size-- > 1 )
1484 h->lookahead->next.list[i]->i_duration = 2 * (h->lookahead->next.list[i+1]->i_pts - h->lookahead->next.list[i]->i_pts);
1486 h->lookahead->next.list[i]->i_duration = h->i_prev_duration;
1489 h->lookahead->next.list[i]->i_duration = delta_tfi_divisor[h->lookahead->next.list[i]->i_pic_struct];
1490 h->i_prev_duration = h->lookahead->next.list[i]->i_duration;
1491 h->lookahead->next.list[i]->f_duration = (double)h->lookahead->next.list[i]->i_duration
1492 * h->sps->vui.i_num_units_in_tick
1493 / h->sps->vui.i_time_scale;
1495 if( h->lookahead->next.list[i]->i_frame > h->i_disp_fields_last_frame && lookahead_size > 0 )
1497 h->lookahead->next.list[i]->i_field_cnt = h->i_disp_fields;
1498 h->i_disp_fields += h->lookahead->next.list[i]->i_duration;
1499 h->i_disp_fields_last_frame = h->lookahead->next.list[i]->i_frame;
1501 else if( lookahead_size == 0 )
1503 h->lookahead->next.list[i]->i_field_cnt = h->i_disp_fields;
1504 h->lookahead->next.list[i]->i_duration = h->i_prev_duration;
1508 if( h->param.rc.b_stat_read )
1510 /* Use the frame types from the first pass */
1511 for( int i = 0; i < h->lookahead->next.i_size; i++ )
1512 h->lookahead->next.list[i]->i_type =
1513 x264_ratecontrol_slice_type( h, h->lookahead->next.list[i]->i_frame );
1515 else if( (h->param.i_bframe && h->param.i_bframe_adaptive)
1516 || h->param.i_scenecut_threshold
1517 || h->param.rc.b_mb_tree
1518 || (h->param.rc.i_vbv_buffer_size && h->param.rc.i_lookahead) )
1519 x264_slicetype_analyse( h, 0 );
1521 for( bframes = 0, brefs = 0;; bframes++ )
1523 frm = h->lookahead->next.list[bframes];
1524 if( frm->i_type == X264_TYPE_BREF && h->param.i_bframe_pyramid < X264_B_PYRAMID_NORMAL &&
1525 brefs == h->param.i_bframe_pyramid )
1527 frm->i_type = X264_TYPE_B;
1528 x264_log( h, X264_LOG_WARNING, "B-ref at frame %d incompatible with B-pyramid %s \n",
1529 frm->i_frame, x264_b_pyramid_names[h->param.i_bframe_pyramid] );
1531 /* pyramid with multiple B-refs needs a big enough dpb that the preceding P-frame stays available.
1532 smaller dpb could be supported by smart enough use of mmco, but it's easier just to forbid it. */
1533 else if( frm->i_type == X264_TYPE_BREF && h->param.i_bframe_pyramid == X264_B_PYRAMID_NORMAL &&
1534 brefs && h->param.i_frame_reference <= (brefs+3) )
1536 frm->i_type = X264_TYPE_B;
1537 x264_log( h, X264_LOG_WARNING, "B-ref at frame %d incompatible with B-pyramid %s and %d reference frames\n",
1538 frm->i_frame, x264_b_pyramid_names[h->param.i_bframe_pyramid], h->param.i_frame_reference );
1541 if( frm->i_type == X264_TYPE_KEYFRAME )
1542 frm->i_type = h->param.b_open_gop ? X264_TYPE_I : X264_TYPE_IDR;
1544 /* Limit GOP size */
1545 if( (!h->param.b_intra_refresh || frm->i_frame == 0) && frm->i_frame - h->lookahead->i_last_keyframe >= h->param.i_keyint_max )
1547 if( frm->i_type == X264_TYPE_AUTO || frm->i_type == X264_TYPE_I )
1548 frm->i_type = h->param.b_open_gop && h->lookahead->i_last_keyframe >= 0 ? X264_TYPE_I : X264_TYPE_IDR;
1549 int warn = frm->i_type != X264_TYPE_IDR;
1550 if( warn && h->param.b_open_gop )
1551 warn &= frm->i_type != X264_TYPE_I;
1553 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 );
1555 if( frm->i_type == X264_TYPE_I && frm->i_frame - h->lookahead->i_last_keyframe >= h->param.i_keyint_min )
1557 if( h->param.b_open_gop )
1559 h->lookahead->i_last_keyframe = frm->i_frame; // Use display order
1560 if( h->param.b_bluray_compat )
1561 h->lookahead->i_last_keyframe -= bframes; // Use bluray order
1562 frm->b_keyframe = 1;
1565 frm->i_type = X264_TYPE_IDR;
1567 if( frm->i_type == X264_TYPE_IDR )
1570 h->lookahead->i_last_keyframe = frm->i_frame;
1571 frm->b_keyframe = 1;
1575 h->lookahead->next.list[bframes]->i_type = X264_TYPE_P;
1579 if( bframes == h->param.i_bframe ||
1580 !h->lookahead->next.list[bframes+1] )
1582 if( IS_X264_TYPE_B( frm->i_type ) )
1583 x264_log( h, X264_LOG_WARNING, "specified frame type is not compatible with max B-frames\n" );
1584 if( frm->i_type == X264_TYPE_AUTO
1585 || IS_X264_TYPE_B( frm->i_type ) )
1586 frm->i_type = X264_TYPE_P;
1589 if( frm->i_type == X264_TYPE_BREF )
1592 if( frm->i_type == X264_TYPE_AUTO )
1593 frm->i_type = X264_TYPE_B;
1595 else if( !IS_X264_TYPE_B( frm->i_type ) ) break;
1599 h->lookahead->next.list[bframes-1]->b_last_minigop_bframe = 1;
1600 h->lookahead->next.list[bframes]->i_bframes = bframes;
1602 /* insert a bref into the sequence */
1603 if( h->param.i_bframe_pyramid && bframes > 1 && !brefs )
1605 h->lookahead->next.list[bframes/2]->i_type = X264_TYPE_BREF;
1609 /* calculate the frame costs ahead of time for x264_rc_analyse_slice while we still have lowres */
1610 if( h->param.rc.i_rc_method != X264_RC_CQP )
1612 x264_mb_analysis_t a;
1614 p1 = b = bframes + 1;
1616 x264_lowres_context_init( h, &a );
1618 frames[0] = h->lookahead->last_nonb;
1619 memcpy( &frames[1], h->lookahead->next.list, (bframes+1) * sizeof(x264_frame_t*) );
1620 if( IS_X264_TYPE_I( h->lookahead->next.list[bframes]->i_type ) )
1625 x264_slicetype_frame_cost( h, &a, frames, p0, p1, b, 0 );
1627 if( (p0 != p1 || bframes) && h->param.rc.i_vbv_buffer_size )
1629 /* We need the intra costs for row SATDs. */
1630 x264_slicetype_frame_cost( h, &a, frames, b, b, b, 0 );
1632 /* We need B-frame costs for row SATDs. */
1634 for( b = 1; b <= bframes; b++ )
1636 if( frames[b]->i_type == X264_TYPE_B )
1637 for( p1 = b; frames[p1]->i_type == X264_TYPE_B; )
1641 x264_slicetype_frame_cost( h, &a, frames, p0, p1, b, 0 );
1642 if( frames[b]->i_type == X264_TYPE_BREF )
1648 /* Analyse for weighted P frames */
1649 if( !h->param.rc.b_stat_read && h->lookahead->next.list[bframes]->i_type == X264_TYPE_P
1650 && h->param.analyse.i_weighted_pred >= X264_WEIGHTP_SIMPLE )
1653 x264_weights_analyse( h, h->lookahead->next.list[bframes], h->lookahead->last_nonb, 0 );
1656 /* shift sequence to coded order.
1657 use a small temporary list to avoid shifting the entire next buffer around */
1658 int i_coded = h->lookahead->next.list[0]->i_frame;
1661 int idx_list[] = { brefs+1, 1 };
1662 for( int i = 0; i < bframes; i++ )
1664 int idx = idx_list[h->lookahead->next.list[i]->i_type == X264_TYPE_BREF]++;
1665 frames[idx] = h->lookahead->next.list[i];
1666 frames[idx]->i_reordered_pts = h->lookahead->next.list[idx]->i_pts;
1668 frames[0] = h->lookahead->next.list[bframes];
1669 frames[0]->i_reordered_pts = h->lookahead->next.list[0]->i_pts;
1670 memcpy( h->lookahead->next.list, frames, (bframes+1) * sizeof(x264_frame_t*) );
1673 for( int i = 0; i <= bframes; i++ )
1675 h->lookahead->next.list[i]->i_coded = i_coded++;
1678 x264_calculate_durations( h, h->lookahead->next.list[i], h->lookahead->next.list[i-1], &h->i_cpb_delay, &h->i_coded_fields );
1679 h->lookahead->next.list[0]->f_planned_cpb_duration[i-1] = (double)h->lookahead->next.list[i-1]->i_cpb_duration *
1680 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1683 x264_calculate_durations( h, h->lookahead->next.list[i], NULL, &h->i_cpb_delay, &h->i_coded_fields );
1685 h->lookahead->next.list[0]->f_planned_cpb_duration[i] = (double)h->lookahead->next.list[i]->i_cpb_duration *
1686 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1690 int x264_rc_analyse_slice( x264_t *h )
1696 if( IS_X264_TYPE_I(h->fenc->i_type) )
1698 else if( h->fenc->i_type == X264_TYPE_P )
1699 p1 = b = h->fenc->i_bframes + 1;
1702 p1 = (h->fref_nearest[1]->i_poc - h->fref_nearest[0]->i_poc)/2;
1703 b = (h->fenc->i_poc - h->fref_nearest[0]->i_poc)/2;
1705 /* We don't need to assign p0/p1 since we are not performing any real analysis here. */
1706 x264_frame_t **frames = &h->fenc - b;
1708 /* cost should have been already calculated by x264_slicetype_decide */
1709 cost = frames[b]->i_cost_est[b-p0][p1-b];
1710 assert( cost >= 0 );
1712 if( h->param.rc.b_mb_tree && !h->param.rc.b_stat_read )
1714 cost = x264_slicetype_frame_cost_recalculate( h, frames, p0, p1, b );
1715 if( b && h->param.rc.i_vbv_buffer_size )
1716 x264_slicetype_frame_cost_recalculate( h, frames, b, b, b );
1718 /* In AQ, use the weighted score instead. */
1719 else if( h->param.rc.i_aq_mode )
1720 cost = frames[b]->i_cost_est_aq[b-p0][p1-b];
1722 h->fenc->i_row_satd = h->fenc->i_row_satds[b-p0][p1-b];
1723 h->fdec->i_row_satd = h->fdec->i_row_satds[b-p0][p1-b];
1724 h->fdec->i_satd = cost;
1725 memcpy( h->fdec->i_row_satd, h->fenc->i_row_satd, h->mb.i_mb_height * sizeof(int) );
1726 if( !IS_X264_TYPE_I(h->fenc->i_type) )
1727 memcpy( h->fdec->i_row_satds[0][0], h->fenc->i_row_satds[0][0], h->mb.i_mb_height * sizeof(int) );
1729 if( h->param.b_intra_refresh && h->param.rc.i_vbv_buffer_size && h->fenc->i_type == X264_TYPE_P )
1731 int ip_factor = 256 * h->param.rc.f_ip_factor; /* fix8 */
1732 for( int y = 0; y < h->mb.i_mb_height; y++ )
1734 int mb_xy = y * h->mb.i_mb_stride + h->fdec->i_pir_start_col;
1735 for( int x = h->fdec->i_pir_start_col; x <= h->fdec->i_pir_end_col; x++, mb_xy++ )
1737 int intra_cost = (h->fenc->i_intra_cost[mb_xy] * ip_factor + 128) >> 8;
1738 int inter_cost = h->fenc->lowres_costs[b-p0][p1-b][mb_xy] & LOWRES_COST_MASK;
1739 int diff = intra_cost - inter_cost;
1740 if( h->param.rc.i_aq_mode )
1741 h->fdec->i_row_satd[y] += (diff * frames[b]->i_inv_qscale_factor[mb_xy] + 128) >> 8;
1743 h->fdec->i_row_satd[y] += diff;
1750 for( int y = 0; y < h->mb.i_mb_height; y++ )
1751 h->fdec->i_row_satd[y] >>= (BIT_DEPTH - 8);
1753 return cost >> (BIT_DEPTH - 8);