1 /*****************************************************************************
2 * slicetype.c: lookahead analysis
3 *****************************************************************************
4 * Copyright (C) 2005-2014 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 void x264_weights_analyse( x264_t *h, x264_frame_t *fenc, x264_frame_t *ref, int b_lookahead );
42 int x264_opencl_lowres_init( x264_t *h, x264_frame_t *fenc, int lambda );
43 int x264_opencl_motionsearch( x264_t *h, x264_frame_t **frames, int b, int ref, int b_islist1, int lambda, const x264_weight_t *w );
44 int x264_opencl_finalize_cost( x264_t *h, int lambda, x264_frame_t **frames, int p0, int p1, int b, int dist_scale_factor );
45 int x264_opencl_precalculate_frame_cost( x264_t *h, x264_frame_t **frames, int lambda, int p0, int p1, int b );
46 void x264_opencl_flush( x264_t *h );
47 void x264_opencl_slicetype_prep( x264_t *h, x264_frame_t **frames, int num_frames, int lambda );
48 void x264_opencl_slicetype_end( x264_t *h );
51 static void x264_lowres_context_init( x264_t *h, x264_mb_analysis_t *a )
53 a->i_qp = X264_LOOKAHEAD_QP;
54 a->i_lambda = x264_lambda_tab[ a->i_qp ];
55 x264_mb_analyse_load_costs( h, a );
56 if( h->param.analyse.i_subpel_refine > 1 )
58 h->mb.i_me_method = X264_MIN( X264_ME_HEX, h->param.analyse.i_me_method );
59 h->mb.i_subpel_refine = 4;
63 h->mb.i_me_method = X264_ME_DIA;
64 h->mb.i_subpel_refine = 2;
66 h->mb.b_chroma_me = 0;
69 /* makes a non-h264 weight (i.e. fix7), into an h264 weight */
70 static void x264_weight_get_h264( int weight_nonh264, int offset, x264_weight_t *w )
74 w->i_scale = weight_nonh264;
75 while( w->i_denom > 0 && (w->i_scale > 127) )
80 w->i_scale = X264_MIN( w->i_scale, 127 );
83 static NOINLINE pixel *x264_weight_cost_init_luma( x264_t *h, x264_frame_t *fenc, x264_frame_t *ref, pixel *dest )
85 int ref0_distance = fenc->i_frame - ref->i_frame - 1;
86 /* Note: this will never run during lookahead as weights_analyse is only called if no
87 * motion search has been done. */
88 if( fenc->lowres_mvs[0][ref0_distance][0][0] != 0x7FFF )
90 int i_stride = fenc->i_stride_lowres;
91 int i_lines = fenc->i_lines_lowres;
92 int i_width = fenc->i_width_lowres;
96 for( int y = 0; y < i_lines; y += 8, p += i_stride*8 )
97 for( int x = 0; x < i_width; x += 8, i_mb_xy++ )
99 int mvx = fenc->lowres_mvs[0][ref0_distance][i_mb_xy][0];
100 int mvy = fenc->lowres_mvs[0][ref0_distance][i_mb_xy][1];
101 h->mc.mc_luma( p+x, i_stride, ref->lowres, i_stride,
102 mvx+(x<<2), mvy+(y<<2), 8, 8, x264_weight_none );
108 return ref->lowres[0];
111 /* How data is organized for 4:2:0/4:2:2 chroma weightp:
114 * fenc = ref + offset
115 * v = u + stride * chroma height */
117 static NOINLINE void x264_weight_cost_init_chroma( x264_t *h, x264_frame_t *fenc, x264_frame_t *ref, pixel *dstu, pixel *dstv )
119 int ref0_distance = fenc->i_frame - ref->i_frame - 1;
120 int i_stride = fenc->i_stride[1];
121 int i_offset = i_stride / 2;
122 int i_lines = fenc->i_lines[1];
123 int i_width = fenc->i_width[1];
124 int v_shift = CHROMA_V_SHIFT;
125 int cw = 8*h->mb.i_mb_width;
126 int ch = 16*h->mb.i_mb_height >> v_shift;
127 int height = 16 >> v_shift;
129 if( fenc->lowres_mvs[0][ref0_distance][0][0] != 0x7FFF )
131 x264_frame_expand_border_chroma( h, ref, 1 );
132 for( int y = 0, mb_xy = 0, pel_offset_y = 0; y < i_lines; y += height, pel_offset_y = y*i_stride )
133 for( int x = 0, pel_offset_x = 0; x < i_width; x += 8, mb_xy++, pel_offset_x += 8 )
135 pixel *pixu = dstu + pel_offset_y + pel_offset_x;
136 pixel *pixv = dstv + pel_offset_y + pel_offset_x;
137 pixel *src1 = ref->plane[1] + pel_offset_y + pel_offset_x*2; /* NV12/NV16 */
138 int mvx = fenc->lowres_mvs[0][ref0_distance][mb_xy][0];
139 int mvy = fenc->lowres_mvs[0][ref0_distance][mb_xy][1];
140 h->mc.mc_chroma( pixu, pixv, i_stride, src1, i_stride, mvx, 2*mvy>>v_shift, 8, height );
144 h->mc.plane_copy_deinterleave( dstu, i_stride, dstv, i_stride, ref->plane[1], i_stride, cw, ch );
145 h->mc.plane_copy_deinterleave( dstu+i_offset, i_stride, dstv+i_offset, i_stride, fenc->plane[1], i_stride, cw, ch );
149 static NOINLINE pixel *x264_weight_cost_init_chroma444( x264_t *h, x264_frame_t *fenc, x264_frame_t *ref, pixel *dst, int p )
151 int ref0_distance = fenc->i_frame - ref->i_frame - 1;
152 int i_stride = fenc->i_stride[p];
153 int i_lines = fenc->i_lines[p];
154 int i_width = fenc->i_width[p];
156 if( fenc->lowres_mvs[0][ref0_distance][0][0] != 0x7FFF )
158 x264_frame_expand_border_chroma( h, ref, p );
159 for( int y = 0, mb_xy = 0, pel_offset_y = 0; y < i_lines; y += 16, pel_offset_y = y*i_stride )
160 for( int x = 0, pel_offset_x = 0; x < i_width; x += 16, mb_xy++, pel_offset_x += 16 )
162 pixel *pix = dst + pel_offset_y + pel_offset_x;
163 pixel *src = ref->plane[p] + pel_offset_y + pel_offset_x;
164 int mvx = fenc->lowres_mvs[0][ref0_distance][mb_xy][0] / 2;
165 int mvy = fenc->lowres_mvs[0][ref0_distance][mb_xy][1] / 2;
166 /* We don't want to calculate hpels for fenc frames, so we round the motion
167 * vectors to fullpel here. It's not too bad, I guess? */
168 h->mc.copy_16x16_unaligned( pix, i_stride, src+mvx+mvy*i_stride, i_stride, 16 );
174 return ref->plane[p];
177 static int x264_weight_slice_header_cost( x264_t *h, x264_weight_t *w, int b_chroma )
179 /* Add cost of weights in the slice header. */
180 int lambda = x264_lambda_tab[X264_LOOKAHEAD_QP];
181 /* 4 times higher, because chroma is analyzed at full resolution. */
185 if( h->param.i_slice_count )
186 numslices = h->param.i_slice_count;
187 else if( h->param.i_slice_max_mbs )
188 numslices = (h->mb.i_mb_width * h->mb.i_mb_height + h->param.i_slice_max_mbs-1) / h->param.i_slice_max_mbs;
191 /* FIXME: find a way to account for --slice-max-size?
192 * 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.
193 * Cut denom cost in half if chroma, since it's shared between the two chroma planes. */
194 int denom_cost = bs_size_ue( w[0].i_denom ) * (2 - b_chroma);
195 return lambda * numslices * ( 10 + denom_cost + 2 * (bs_size_se( w[0].i_scale ) + bs_size_se( w[0].i_offset )) );
198 static NOINLINE unsigned int x264_weight_cost_luma( x264_t *h, x264_frame_t *fenc, pixel *src, x264_weight_t *w )
200 unsigned int cost = 0;
201 int i_stride = fenc->i_stride_lowres;
202 int i_lines = fenc->i_lines_lowres;
203 int i_width = fenc->i_width_lowres;
204 pixel *fenc_plane = fenc->lowres[0];
205 ALIGNED_ARRAY_16( pixel, buf,[8*8] );
211 for( int y = 0; y < i_lines; y += 8, pixoff = y*i_stride )
212 for( int x = 0; x < i_width; x += 8, i_mb++, pixoff += 8)
214 w->weightfn[8>>2]( buf, 8, &src[pixoff], i_stride, w, 8 );
215 int cmp = h->pixf.mbcmp[PIXEL_8x8]( buf, 8, &fenc_plane[pixoff], i_stride );
216 cost += X264_MIN( cmp, fenc->i_intra_cost[i_mb] );
218 cost += x264_weight_slice_header_cost( h, w, 0 );
221 for( int y = 0; y < i_lines; y += 8, pixoff = y*i_stride )
222 for( int x = 0; x < i_width; x += 8, i_mb++, pixoff += 8 )
224 int cmp = h->pixf.mbcmp[PIXEL_8x8]( &src[pixoff], i_stride, &fenc_plane[pixoff], i_stride );
225 cost += X264_MIN( cmp, fenc->i_intra_cost[i_mb] );
231 static NOINLINE unsigned int x264_weight_cost_chroma( x264_t *h, x264_frame_t *fenc, pixel *ref, x264_weight_t *w )
233 unsigned int cost = 0;
234 int i_stride = fenc->i_stride[1];
235 int i_lines = fenc->i_lines[1];
236 int i_width = fenc->i_width[1];
237 pixel *src = ref + (i_stride >> 1);
238 ALIGNED_ARRAY_16( pixel, buf, [8*16] );
240 int height = 16 >> CHROMA_V_SHIFT;
243 for( int y = 0; y < i_lines; y += height, pixoff = y*i_stride )
244 for( int x = 0; x < i_width; x += 8, pixoff += 8 )
246 w->weightfn[8>>2]( buf, 8, &ref[pixoff], i_stride, w, height );
247 /* The naive and seemingly sensible algorithm is to use mbcmp as in luma.
248 * But testing shows that for chroma the DC coefficient is by far the most
249 * important part of the coding cost. Thus a more useful chroma weight is
250 * obtained by comparing each block's DC coefficient instead of the actual
252 cost += h->pixf.asd8( buf, 8, &src[pixoff], i_stride, height );
254 cost += x264_weight_slice_header_cost( h, w, 1 );
257 for( int y = 0; y < i_lines; y += height, pixoff = y*i_stride )
258 for( int x = 0; x < i_width; x += 8, pixoff += 8 )
259 cost += h->pixf.asd8( &ref[pixoff], i_stride, &src[pixoff], i_stride, height );
264 static NOINLINE unsigned int x264_weight_cost_chroma444( x264_t *h, x264_frame_t *fenc, pixel *ref, x264_weight_t *w, int p )
266 unsigned int cost = 0;
267 int i_stride = fenc->i_stride[p];
268 int i_lines = fenc->i_lines[p];
269 int i_width = fenc->i_width[p];
270 pixel *src = fenc->plane[p];
271 ALIGNED_ARRAY_16( pixel, buf, [16*16] );
275 for( int y = 0; y < i_lines; y += 16, pixoff = y*i_stride )
276 for( int x = 0; x < i_width; x += 16, pixoff += 16 )
278 w->weightfn[16>>2]( buf, 16, &ref[pixoff], i_stride, w, 16 );
279 cost += h->pixf.mbcmp[PIXEL_16x16]( buf, 16, &src[pixoff], i_stride );
281 cost += x264_weight_slice_header_cost( h, w, 1 );
284 for( int y = 0; y < i_lines; y += 16, pixoff = y*i_stride )
285 for( int x = 0; x < i_width; x += 16, pixoff += 16 )
286 cost += h->pixf.mbcmp[PIXEL_16x16]( &ref[pixoff], i_stride, &src[pixoff], i_stride );
291 void x264_weights_analyse( x264_t *h, x264_frame_t *fenc, x264_frame_t *ref, int b_lookahead )
293 int i_delta_index = fenc->i_frame - ref->i_frame - 1;
294 /* epsilon is chosen to require at least a numerator of 127 (with denominator = 128) */
295 const float epsilon = 1.f/128.f;
296 x264_weight_t *weights = fenc->weight[0];
297 SET_WEIGHT( weights[0], 0, 1, 0, 0 );
298 SET_WEIGHT( weights[1], 0, 1, 0, 0 );
299 SET_WEIGHT( weights[2], 0, 1, 0, 0 );
300 int chroma_initted = 0;
301 float guess_scale[3];
304 for( int plane = 0; plane <= 2*!b_lookahead; plane++ )
306 float fenc_var = fenc->i_pixel_ssd[plane] + !ref->i_pixel_ssd[plane];
307 float ref_var = ref->i_pixel_ssd[plane] + !ref->i_pixel_ssd[plane];
308 guess_scale[plane] = sqrtf( fenc_var / ref_var );
309 fenc_mean[plane] = (float)fenc->i_pixel_sum[plane] / (fenc->i_lines[!!plane] * fenc->i_width[!!plane]) / (1 << (BIT_DEPTH - 8));
310 ref_mean[plane] = (float) ref->i_pixel_sum[plane] / (fenc->i_lines[!!plane] * fenc->i_width[!!plane]) / (1 << (BIT_DEPTH - 8));
313 int chroma_denom = 7;
316 /* make sure both our scale factors fit */
317 while( chroma_denom > 0 )
319 float thresh = 127.f / (1<<chroma_denom);
320 if( guess_scale[1] < thresh && guess_scale[2] < thresh )
326 /* Don't check chroma in lookahead, or if there wasn't a luma weight. */
327 for( int plane = 0; plane <= 2 && !( plane && ( !weights[0].weightfn || b_lookahead ) ); plane++ )
329 int minoff, minscale, mindenom;
330 unsigned int minscore, origscore;
334 if( fabsf( ref_mean[plane] - fenc_mean[plane] ) < 0.5f && fabsf( 1.f - guess_scale[plane] ) < epsilon )
336 SET_WEIGHT( weights[plane], 0, 1, 0, 0 );
342 weights[plane].i_denom = chroma_denom;
343 weights[plane].i_scale = x264_clip3( round( guess_scale[plane] * (1<<chroma_denom) ), 0, 255 );
344 if( weights[plane].i_scale > 127 )
346 weights[1].weightfn = weights[2].weightfn = NULL;
351 x264_weight_get_h264( round( guess_scale[plane] * 128 ), 0, &weights[plane] );
354 mindenom = weights[plane].i_denom;
355 minscale = weights[plane].i_scale;
361 if( !fenc->b_intra_calculated )
363 x264_mb_analysis_t a;
364 x264_lowres_context_init( h, &a );
365 x264_slicetype_frame_cost( h, &a, &fenc, 0, 0, 0, 0 );
367 mcbuf = x264_weight_cost_init_luma( h, fenc, ref, h->mb.p_weight_buf[0] );
368 origscore = minscore = x264_weight_cost_luma( h, fenc, mcbuf, NULL );
374 mcbuf = x264_weight_cost_init_chroma444( h, fenc, ref, h->mb.p_weight_buf[0], plane );
375 origscore = minscore = x264_weight_cost_chroma444( h, fenc, mcbuf, NULL, plane );
379 pixel *dstu = h->mb.p_weight_buf[0];
380 pixel *dstv = h->mb.p_weight_buf[0]+fenc->i_stride[1]*fenc->i_lines[1];
381 if( !chroma_initted++ )
382 x264_weight_cost_init_chroma( h, fenc, ref, dstu, dstv );
383 mcbuf = plane == 1 ? dstu : dstv;
384 origscore = minscore = x264_weight_cost_chroma( h, fenc, mcbuf, NULL );
391 /* Picked somewhat arbitrarily */
392 static const uint8_t weight_check_distance[][2] =
394 {0,0},{0,0},{0,1},{0,1},
395 {0,1},{0,1},{0,1},{1,1},
396 {1,1},{2,1},{2,1},{4,2}
398 int scale_dist = b_lookahead ? 0 : weight_check_distance[h->param.analyse.i_subpel_refine][0];
399 int offset_dist = b_lookahead ? 0 : weight_check_distance[h->param.analyse.i_subpel_refine][1];
401 int start_scale = x264_clip3( minscale - scale_dist, 0, 127 );
402 int end_scale = x264_clip3( minscale + scale_dist, 0, 127 );
403 for( int i_scale = start_scale; i_scale <= end_scale; i_scale++ )
405 int cur_scale = i_scale;
406 int cur_offset = fenc_mean[plane] - ref_mean[plane] * cur_scale / (1 << mindenom) + 0.5f * b_lookahead;
407 if( cur_offset < - 128 || cur_offset > 127 )
409 /* Rescale considering the constraints on cur_offset. We do it in this order
410 * because scale has a much wider range than offset (because of denom), so
411 * it should almost never need to be clamped. */
412 cur_offset = x264_clip3( cur_offset, -128, 127 );
413 cur_scale = (1 << mindenom) * (fenc_mean[plane] - cur_offset) / ref_mean[plane] + 0.5f;
414 cur_scale = x264_clip3( cur_scale, 0, 127 );
416 int start_offset = x264_clip3( cur_offset - offset_dist, -128, 127 );
417 int end_offset = x264_clip3( cur_offset + offset_dist, -128, 127 );
418 for( int i_off = start_offset; i_off <= end_offset; i_off++ )
420 SET_WEIGHT( weights[plane], 1, cur_scale, mindenom, i_off );
425 s = x264_weight_cost_chroma444( h, fenc, mcbuf, &weights[plane], plane );
427 s = x264_weight_cost_chroma( h, fenc, mcbuf, &weights[plane] );
430 s = x264_weight_cost_luma( h, fenc, mcbuf, &weights[plane] );
431 COPY4_IF_LT( minscore, s, minscale, cur_scale, minoff, i_off, found, 1 );
433 // Don't check any more offsets if the previous one had a lower cost than the current one
434 if( minoff == start_offset && i_off != start_offset )
440 /* Use a smaller denominator if possible */
443 while( mindenom > 0 && !(minscale&1) )
450 /* FIXME: More analysis can be done here on SAD vs. SATD termination. */
451 /* 0.2% termination derived experimentally to avoid weird weights in frames that are mostly intra. */
452 if( !found || (minscale == 1 << mindenom && minoff == 0) || (float)minscore / origscore > 0.998f )
454 SET_WEIGHT( weights[plane], 0, 1, 0, 0 );
458 SET_WEIGHT( weights[plane], 1, minscale, mindenom, minoff );
460 if( h->param.analyse.i_weighted_pred == X264_WEIGHTP_FAKE && weights[0].weightfn && !plane )
461 fenc->f_weighted_cost_delta[i_delta_index] = (float)minscore / origscore;
464 /* Optimize and unify denominator */
465 if( weights[1].weightfn || weights[2].weightfn )
467 int denom = weights[1].weightfn ? weights[1].i_denom : weights[2].i_denom;
468 int both_weighted = weights[1].weightfn && weights[2].weightfn;
469 /* If only one plane is weighted, the other has an implicit scale of 1<<denom.
470 * With denom==7, this comes out to 128, which is invalid, so don't allow that. */
471 while( (!both_weighted && denom==7) ||
472 (denom > 0 && !(weights[1].weightfn && (weights[1].i_scale&1))
473 && !(weights[2].weightfn && (weights[2].i_scale&1))) )
476 for( int i = 1; i <= 2; i++ )
477 if( weights[i].weightfn )
479 weights[i].i_scale >>= 1;
480 weights[i].i_denom = denom;
484 for( int i = 1; i <= 2; i++ )
485 if( weights[i].weightfn )
486 h->mc.weight_cache( h, &weights[i] );
488 if( weights[0].weightfn && b_lookahead )
490 //scale lowres in lookahead for slicetype_frame_cost
491 pixel *src = ref->buffer_lowres[0];
492 pixel *dst = h->mb.p_weight_buf[0];
493 int width = ref->i_width_lowres + PADH*2;
494 int height = ref->i_lines_lowres + PADV*2;
495 x264_weight_scale_plane( h, dst, ref->i_stride_lowres, src, ref->i_stride_lowres,
496 width, height, &weights[0] );
497 fenc->weighted[0] = h->mb.p_weight_buf[0] + PADH + ref->i_stride_lowres * PADV;
501 /* Output buffers are separated by 128 bytes to avoid false sharing of cachelines
502 * in multithreaded lookahead. */
504 /* cost_est, cost_est_aq, intra_mbs, num rows */
507 #define COST_EST_AQ 1
510 #define ROW_SATD (NUM_INTS + (h->mb.i_mb_y - h->i_threadslice_start))
512 static void x264_slicetype_mb_cost( x264_t *h, x264_mb_analysis_t *a,
513 x264_frame_t **frames, int p0, int p1, int b,
514 int dist_scale_factor, int do_search[2], const x264_weight_t *w,
515 int *output_inter, int *output_intra )
517 x264_frame_t *fref0 = frames[p0];
518 x264_frame_t *fref1 = frames[p1];
519 x264_frame_t *fenc = frames[b];
520 const int b_bidir = (b < p1);
521 const int i_mb_x = h->mb.i_mb_x;
522 const int i_mb_y = h->mb.i_mb_y;
523 const int i_mb_stride = h->mb.i_mb_width;
524 const int i_mb_xy = i_mb_x + i_mb_y * i_mb_stride;
525 const int i_stride = fenc->i_stride_lowres;
526 const int i_pel_offset = 8 * (i_mb_x + i_mb_y * i_stride);
527 const int i_bipred_weight = h->param.analyse.b_weighted_bipred ? 64 - (dist_scale_factor>>2) : 32;
528 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] };
529 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] };
530 int b_frame_score_mb = (i_mb_x > 0 && i_mb_x < h->mb.i_mb_width - 1 &&
531 i_mb_y > 0 && i_mb_y < h->mb.i_mb_height - 1) ||
532 h->mb.i_mb_width <= 2 || h->mb.i_mb_height <= 2;
534 ALIGNED_ARRAY_16( pixel, pix1,[9*FDEC_STRIDE] );
535 pixel *pix2 = pix1+8;
537 int i_bcost = COST_MAX;
539 /* A small, arbitrary bias to avoid VBV problems caused by zero-residual lookahead blocks. */
540 int lowres_penalty = 4;
542 h->mb.pic.p_fenc[0] = h->mb.pic.fenc_buf;
543 h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fenc[0], FENC_STRIDE, &fenc->lowres[0][i_pel_offset], i_stride, 8 );
546 goto lowres_intra_mb;
548 // no need for h->mb.mv_min[]
549 h->mb.mv_limit_fpel[0][0] = -8*h->mb.i_mb_x - 4;
550 h->mb.mv_limit_fpel[1][0] = 8*( h->mb.i_mb_width - h->mb.i_mb_x - 1 ) + 4;
551 h->mb.mv_min_spel[0] = 4*( h->mb.mv_limit_fpel[0][0] - 8 );
552 h->mb.mv_max_spel[0] = 4*( h->mb.mv_limit_fpel[1][0] + 8 );
553 if( h->mb.i_mb_x >= h->mb.i_mb_width - 2 )
555 h->mb.mv_limit_fpel[0][1] = -8*h->mb.i_mb_y - 4;
556 h->mb.mv_limit_fpel[1][1] = 8*( h->mb.i_mb_height - h->mb.i_mb_y - 1 ) + 4;
557 h->mb.mv_min_spel[1] = 4*( h->mb.mv_limit_fpel[0][1] - 8 );
558 h->mb.mv_max_spel[1] = 4*( h->mb.mv_limit_fpel[1][1] + 8 );
561 #define LOAD_HPELS_LUMA(dst, src) \
563 (dst)[0] = &(src)[0][i_pel_offset]; \
564 (dst)[1] = &(src)[1][i_pel_offset]; \
565 (dst)[2] = &(src)[2][i_pel_offset]; \
566 (dst)[3] = &(src)[3][i_pel_offset]; \
568 #define LOAD_WPELS_LUMA(dst,src) \
569 (dst) = &(src)[i_pel_offset];
571 #define CLIP_MV( mv ) \
573 mv[0] = x264_clip3( mv[0], h->mb.mv_min_spel[0], h->mb.mv_max_spel[0] ); \
574 mv[1] = x264_clip3( mv[1], h->mb.mv_min_spel[1], h->mb.mv_max_spel[1] ); \
576 #define TRY_BIDIR( mv0, mv1, penalty ) \
579 if( h->param.analyse.i_subpel_refine <= 1 ) \
581 int hpel_idx1 = (((mv0)[0]&2)>>1) + ((mv0)[1]&2); \
582 int hpel_idx2 = (((mv1)[0]&2)>>1) + ((mv1)[1]&2); \
583 pixel *src1 = m[0].p_fref[hpel_idx1] + ((mv0)[0]>>2) + ((mv0)[1]>>2) * m[0].i_stride[0]; \
584 pixel *src2 = m[1].p_fref[hpel_idx2] + ((mv1)[0]>>2) + ((mv1)[1]>>2) * m[1].i_stride[0]; \
585 h->mc.avg[PIXEL_8x8]( pix1, 16, src1, m[0].i_stride[0], src2, m[1].i_stride[0], i_bipred_weight ); \
589 intptr_t stride1 = 16, stride2 = 16; \
590 pixel *src1, *src2; \
591 src1 = h->mc.get_ref( pix1, &stride1, m[0].p_fref, m[0].i_stride[0], \
592 (mv0)[0], (mv0)[1], 8, 8, w ); \
593 src2 = h->mc.get_ref( pix2, &stride2, m[1].p_fref, m[1].i_stride[0], \
594 (mv1)[0], (mv1)[1], 8, 8, w ); \
595 h->mc.avg[PIXEL_8x8]( pix1, 16, src1, stride1, src2, stride2, i_bipred_weight ); \
597 i_cost = penalty * a->i_lambda + h->pixf.mbcmp[PIXEL_8x8]( \
598 m[0].p_fenc[0], FENC_STRIDE, pix1, 16 ); \
599 COPY2_IF_LT( i_bcost, i_cost, list_used, 3 ); \
602 m[0].i_pixel = PIXEL_8x8;
603 m[0].p_cost_mv = a->p_cost_mv;
604 m[0].i_stride[0] = i_stride;
605 m[0].p_fenc[0] = h->mb.pic.p_fenc[0];
608 LOAD_HPELS_LUMA( m[0].p_fref, fref0->lowres );
609 m[0].p_fref_w = m[0].p_fref[0];
611 LOAD_WPELS_LUMA( m[0].p_fref_w, fenc->weighted[0] );
615 int16_t *mvr = fref1->lowres_mvs[0][p1-p0-1][i_mb_xy];
616 ALIGNED_ARRAY_8( int16_t, dmv,[2],[2] );
618 m[1].i_pixel = PIXEL_8x8;
619 m[1].p_cost_mv = a->p_cost_mv;
620 m[1].i_stride[0] = i_stride;
621 m[1].p_fenc[0] = h->mb.pic.p_fenc[0];
623 m[1].weight = x264_weight_none;
624 LOAD_HPELS_LUMA( m[1].p_fref, fref1->lowres );
625 m[1].p_fref_w = m[1].p_fref[0];
627 dmv[0][0] = ( mvr[0] * dist_scale_factor + 128 ) >> 8;
628 dmv[0][1] = ( mvr[1] * dist_scale_factor + 128 ) >> 8;
629 dmv[1][0] = dmv[0][0] - mvr[0];
630 dmv[1][1] = dmv[0][1] - mvr[1];
633 if( h->param.analyse.i_subpel_refine <= 1 )
634 M64( dmv ) &= ~0x0001000100010001ULL; /* mv & ~1 */
636 TRY_BIDIR( dmv[0], dmv[1], 0 );
640 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 );
641 i_cost = h->pixf.mbcmp[PIXEL_8x8]( m[0].p_fenc[0], FENC_STRIDE, pix1, 16 );
642 COPY2_IF_LT( i_bcost, i_cost, list_used, 3 );
646 for( int l = 0; l < 1 + b_bidir; l++ )
651 int16_t (*fenc_mv)[2] = fenc_mvs[l];
652 ALIGNED_4( int16_t mvc[4][2] );
654 /* Reverse-order MV prediction. */
657 #define MVC(mv) { CP32( mvc[i_mvc], mv ); i_mvc++; }
658 if( i_mb_x < h->mb.i_mb_width - 1 )
660 if( i_mb_y < h->i_threadslice_end - 1 )
662 MVC( fenc_mv[i_mb_stride] );
664 MVC( fenc_mv[i_mb_stride-1] );
665 if( i_mb_x < h->mb.i_mb_width - 1 )
666 MVC( fenc_mv[i_mb_stride+1] );
670 CP32( m[l].mvp, mvc[0] );
672 x264_median_mv( m[l].mvp, mvc[0], mvc[1], mvc[2] );
674 /* Fast skip for cases of near-zero residual. Shortcut: don't bother except in the mv0 case,
675 * since anything else is likely to have enough residual to not trigger the skip. */
676 if( !M32( m[l].mvp ) )
678 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] );
686 x264_me_search( h, &m[l], mvc, i_mvc );
687 m[l].cost -= a->p_cost_mv[0]; // remove mvcost from skip mbs
689 m[l].cost += 5 * a->i_lambda;
692 CP32( fenc_mvs[l], m[l].mv );
693 *fenc_costs[l] = m[l].cost;
697 CP32( m[l].mv, fenc_mvs[l] );
698 m[l].cost = *fenc_costs[l];
700 COPY2_IF_LT( i_bcost, m[l].cost, list_used, l+1 );
703 if( b_bidir && ( M32( m[0].mv ) || M32( m[1].mv ) ) )
704 TRY_BIDIR( m[0].mv, m[1].mv, 5 );
707 if( !fenc->b_intra_calculated )
709 ALIGNED_ARRAY_16( pixel, edge,[36] );
710 pixel *pix = &pix1[8+FDEC_STRIDE];
711 pixel *src = &fenc->lowres[0][i_pel_offset];
712 const int intra_penalty = 5 * a->i_lambda;
714 int pixoff = 4 / sizeof(pixel);
716 /* Avoid store forwarding stalls by writing larger chunks */
717 memcpy( pix-FDEC_STRIDE, src-i_stride, 16 * sizeof(pixel) );
718 for( int i = -1; i < 8; i++ )
719 M32( &pix[i*FDEC_STRIDE-pixoff] ) = M32( &src[i*i_stride-pixoff] );
721 h->pixf.intra_mbcmp_x3_8x8c( h->mb.pic.p_fenc[0], pix, satds );
722 int i_icost = X264_MIN3( satds[0], satds[1], satds[2] );
724 if( h->param.analyse.i_subpel_refine > 1 )
726 h->predict_8x8c[I_PRED_CHROMA_P]( pix );
727 int satd = h->pixf.mbcmp[PIXEL_8x8]( pix, FDEC_STRIDE, h->mb.pic.p_fenc[0], FENC_STRIDE );
728 i_icost = X264_MIN( i_icost, satd );
729 h->predict_8x8_filter( pix, edge, ALL_NEIGHBORS, ALL_NEIGHBORS );
730 for( int i = 3; i < 9; i++ )
732 h->predict_8x8[i]( pix, edge );
733 satd = h->pixf.mbcmp[PIXEL_8x8]( pix, FDEC_STRIDE, h->mb.pic.p_fenc[0], FENC_STRIDE );
734 i_icost = X264_MIN( i_icost, satd );
738 i_icost += intra_penalty + lowres_penalty;
739 fenc->i_intra_cost[i_mb_xy] = i_icost;
740 int i_icost_aq = i_icost;
741 if( h->param.rc.i_aq_mode )
742 i_icost_aq = (i_icost_aq * fenc->i_inv_qscale_factor[i_mb_xy] + 128) >> 8;
743 output_intra[ROW_SATD] += i_icost_aq;
744 if( b_frame_score_mb )
746 output_intra[COST_EST] += i_icost;
747 output_intra[COST_EST_AQ] += i_icost_aq;
750 i_bcost += lowres_penalty;
752 /* forbid intra-mbs in B-frames, because it's rare and not worth checking */
753 /* FIXME: Should we still forbid them now that we cache intra scores? */
756 int i_icost = fenc->i_intra_cost[i_mb_xy];
757 int b_intra = i_icost < i_bcost;
763 if( b_frame_score_mb )
764 output_inter[INTRA_MBS] += b_intra;
767 /* In an I-frame, we've already added the results above in the intra section. */
770 int i_bcost_aq = i_bcost;
771 if( h->param.rc.i_aq_mode )
772 i_bcost_aq = (i_bcost_aq * fenc->i_inv_qscale_factor[i_mb_xy] + 128) >> 8;
773 output_inter[ROW_SATD] += i_bcost_aq;
774 if( b_frame_score_mb )
776 /* Don't use AQ-weighted costs for slicetype decision, only for ratecontrol. */
777 output_inter[COST_EST] += i_bcost;
778 output_inter[COST_EST_AQ] += i_bcost_aq;
782 fenc->lowres_costs[b-p0][p1-b][i_mb_xy] = X264_MIN( i_bcost, LOWRES_COST_MASK ) + (list_used << LOWRES_COST_SHIFT);
787 (h->mb.i_mb_width > 2 && h->mb.i_mb_height > 2 ?\
788 (h->mb.i_mb_width - 2) * (h->mb.i_mb_height - 2) :\
789 h->mb.i_mb_width * h->mb.i_mb_height)
794 x264_mb_analysis_t *a;
795 x264_frame_t **frames;
799 int dist_scale_factor;
801 const x264_weight_t *w;
804 } x264_slicetype_slice_t;
806 static void x264_slicetype_slice_cost( x264_slicetype_slice_t *s )
810 /* Lowres lookahead goes backwards because the MVs are used as predictors in the main encode.
811 * This considerably improves MV prediction overall. */
813 /* The edge mbs seem to reduce the predictive quality of the
814 * whole frame's score, but are needed for a spatial distribution. */
815 int do_edges = h->param.rc.b_mb_tree || h->param.rc.i_vbv_buffer_size || h->mb.i_mb_width <= 2 || h->mb.i_mb_height <= 2;
817 int start_y = X264_MIN( h->i_threadslice_end - 1, h->mb.i_mb_height - 2 + do_edges );
818 int end_y = X264_MAX( h->i_threadslice_start, 1 - do_edges );
819 int start_x = h->mb.i_mb_width - 2 + do_edges;
820 int end_x = 1 - do_edges;
822 for( h->mb.i_mb_y = start_y; h->mb.i_mb_y >= end_y; h->mb.i_mb_y-- )
823 for( h->mb.i_mb_x = start_x; h->mb.i_mb_x >= end_x; h->mb.i_mb_x-- )
824 x264_slicetype_mb_cost( h, s->a, s->frames, s->p0, s->p1, s->b, s->dist_scale_factor,
825 s->do_search, s->w, s->output_inter, s->output_intra );
828 static int x264_slicetype_frame_cost( x264_t *h, x264_mb_analysis_t *a,
829 x264_frame_t **frames, int p0, int p1, int b,
830 int b_intra_penalty )
834 const x264_weight_t *w = x264_weight_none;
835 x264_frame_t *fenc = frames[b];
837 /* Check whether we already evaluated this frame
838 * If we have tried this frame as P, then we have also tried
839 * the preceding frames as B. (is this still true?) */
840 /* Also check that we already calculated the row SATDs for the current frame. */
841 if( fenc->i_cost_est[b-p0][p1-b] >= 0 && (!h->param.rc.i_vbv_buffer_size || fenc->i_row_satds[b-p0][p1-b][0] != -1) )
842 i_score = fenc->i_cost_est[b-p0][p1-b];
845 int dist_scale_factor = 128;
847 /* For each list, check to see whether we have lowres motion-searched this reference frame before. */
848 do_search[0] = b != p0 && fenc->lowres_mvs[0][b-p0-1][0][0] == 0x7FFF;
849 do_search[1] = b != p1 && fenc->lowres_mvs[1][p1-b-1][0][0] == 0x7FFF;
852 if( h->param.analyse.i_weighted_pred && b == p1 )
855 x264_weights_analyse( h, fenc, frames[p0], 1 );
858 fenc->lowres_mvs[0][b-p0-1][0][0] = 0;
860 if( do_search[1] ) fenc->lowres_mvs[1][p1-b-1][0][0] = 0;
863 dist_scale_factor = ( ((b-p0) << 8) + ((p1-p0) >> 1) ) / (p1-p0);
865 int output_buf_size = h->mb.i_mb_height + (NUM_INTS + PAD_SIZE) * h->param.i_lookahead_threads;
866 int *output_inter[X264_LOOKAHEAD_THREAD_MAX+1];
867 int *output_intra[X264_LOOKAHEAD_THREAD_MAX+1];
868 output_inter[0] = h->scratch_buffer2;
869 output_intra[0] = output_inter[0] + output_buf_size;
872 if( h->param.b_opencl )
874 x264_opencl_lowres_init(h, fenc, a->i_lambda );
877 x264_opencl_lowres_init( h, frames[p0], a->i_lambda );
878 x264_opencl_motionsearch( h, frames, b, p0, 0, a->i_lambda, w );
882 x264_opencl_lowres_init( h, frames[p1], a->i_lambda );
883 x264_opencl_motionsearch( h, frames, b, p1, 1, a->i_lambda, NULL );
886 x264_opencl_finalize_cost( h, a->i_lambda, frames, p0, p1, b, dist_scale_factor );
887 x264_opencl_flush( h );
889 i_score = fenc->i_cost_est[b-p0][p1-b];
894 if( h->param.i_lookahead_threads > 1 )
896 x264_slicetype_slice_t s[X264_LOOKAHEAD_THREAD_MAX];
898 for( int i = 0; i < h->param.i_lookahead_threads; i++ )
900 x264_t *t = h->lookahead_thread[i];
902 /* FIXME move this somewhere else */
903 t->mb.i_me_method = h->mb.i_me_method;
904 t->mb.i_subpel_refine = h->mb.i_subpel_refine;
905 t->mb.b_chroma_me = h->mb.b_chroma_me;
907 s[i] = (x264_slicetype_slice_t){ t, a, frames, p0, p1, b, dist_scale_factor, do_search, w,
908 output_inter[i], output_intra[i] };
910 t->i_threadslice_start = ((h->mb.i_mb_height * i + h->param.i_lookahead_threads/2) / h->param.i_lookahead_threads);
911 t->i_threadslice_end = ((h->mb.i_mb_height * (i+1) + h->param.i_lookahead_threads/2) / h->param.i_lookahead_threads);
913 int thread_height = t->i_threadslice_end - t->i_threadslice_start;
914 int thread_output_size = thread_height + NUM_INTS;
915 memset( output_inter[i], 0, thread_output_size * sizeof(int) );
916 memset( output_intra[i], 0, thread_output_size * sizeof(int) );
917 output_inter[i][NUM_ROWS] = output_intra[i][NUM_ROWS] = thread_height;
919 output_inter[i+1] = output_inter[i] + thread_output_size + PAD_SIZE;
920 output_intra[i+1] = output_intra[i] + thread_output_size + PAD_SIZE;
922 x264_threadpool_run( h->lookaheadpool, (void*)x264_slicetype_slice_cost, &s[i] );
924 for( int i = 0; i < h->param.i_lookahead_threads; i++ )
925 x264_threadpool_wait( h->lookaheadpool, &s[i] );
929 h->i_threadslice_start = 0;
930 h->i_threadslice_end = h->mb.i_mb_height;
931 memset( output_inter[0], 0, (output_buf_size - PAD_SIZE) * sizeof(int) );
932 memset( output_intra[0], 0, (output_buf_size - PAD_SIZE) * sizeof(int) );
933 output_inter[0][NUM_ROWS] = output_intra[0][NUM_ROWS] = h->mb.i_mb_height;
934 x264_slicetype_slice_t s = (x264_slicetype_slice_t){ h, a, frames, p0, p1, b, dist_scale_factor, do_search, w,
935 output_inter[0], output_intra[0] };
936 x264_slicetype_slice_cost( &s );
939 /* Sum up accumulators */
941 fenc->i_intra_mbs[b-p0] = 0;
942 if( !fenc->b_intra_calculated )
944 fenc->i_cost_est[0][0] = 0;
945 fenc->i_cost_est_aq[0][0] = 0;
947 fenc->i_cost_est[b-p0][p1-b] = 0;
948 fenc->i_cost_est_aq[b-p0][p1-b] = 0;
950 int *row_satd_inter = fenc->i_row_satds[b-p0][p1-b];
951 int *row_satd_intra = fenc->i_row_satds[0][0];
952 for( int i = 0; i < h->param.i_lookahead_threads; i++ )
955 fenc->i_intra_mbs[b-p0] += output_inter[i][INTRA_MBS];
956 if( !fenc->b_intra_calculated )
958 fenc->i_cost_est[0][0] += output_intra[i][COST_EST];
959 fenc->i_cost_est_aq[0][0] += output_intra[i][COST_EST_AQ];
962 fenc->i_cost_est[b-p0][p1-b] += output_inter[i][COST_EST];
963 fenc->i_cost_est_aq[b-p0][p1-b] += output_inter[i][COST_EST_AQ];
965 if( h->param.rc.i_vbv_buffer_size )
967 int row_count = output_inter[i][NUM_ROWS];
968 memcpy( row_satd_inter, output_inter[i] + NUM_INTS, row_count * sizeof(int) );
969 if( !fenc->b_intra_calculated )
970 memcpy( row_satd_intra, output_intra[i] + NUM_INTS, row_count * sizeof(int) );
971 row_satd_inter += row_count;
972 row_satd_intra += row_count;
976 i_score = fenc->i_cost_est[b-p0][p1-b];
978 i_score = (uint64_t)i_score * 100 / (120 + h->param.i_bframe_bias);
980 fenc->b_intra_calculated = 1;
982 fenc->i_cost_est[b-p0][p1-b] = i_score;
987 if( b_intra_penalty )
989 // arbitrary penalty for I-blocks after B-frames
991 i_score += (uint64_t)i_score * fenc->i_intra_mbs[b-p0] / (nmb * 8);
996 /* If MB-tree changes the quantizers, we need to recalculate the frame cost without
997 * re-running lookahead. */
998 static int x264_slicetype_frame_cost_recalculate( x264_t *h, x264_frame_t **frames, int p0, int p1, int b )
1001 int *row_satd = frames[b]->i_row_satds[b-p0][p1-b];
1002 float *qp_offset = IS_X264_TYPE_B(frames[b]->i_type) ? frames[b]->f_qp_offset_aq : frames[b]->f_qp_offset;
1004 for( h->mb.i_mb_y = h->mb.i_mb_height - 1; h->mb.i_mb_y >= 0; h->mb.i_mb_y-- )
1006 row_satd[ h->mb.i_mb_y ] = 0;
1007 for( h->mb.i_mb_x = h->mb.i_mb_width - 1; h->mb.i_mb_x >= 0; h->mb.i_mb_x-- )
1009 int i_mb_xy = h->mb.i_mb_x + h->mb.i_mb_y*h->mb.i_mb_stride;
1010 int i_mb_cost = frames[b]->lowres_costs[b-p0][p1-b][i_mb_xy] & LOWRES_COST_MASK;
1011 float qp_adj = qp_offset[i_mb_xy];
1012 i_mb_cost = (i_mb_cost * x264_exp2fix8(qp_adj) + 128) >> 8;
1013 row_satd[ h->mb.i_mb_y ] += i_mb_cost;
1014 if( (h->mb.i_mb_y > 0 && h->mb.i_mb_y < h->mb.i_mb_height - 1 &&
1015 h->mb.i_mb_x > 0 && h->mb.i_mb_x < h->mb.i_mb_width - 1) ||
1016 h->mb.i_mb_width <= 2 || h->mb.i_mb_height <= 2 )
1018 i_score += i_mb_cost;
1025 /* Trade off precision in mbtree for increased range */
1026 #define MBTREE_PRECISION 0.5f
1028 static void x264_macroblock_tree_finish( x264_t *h, x264_frame_t *frame, float average_duration, int ref0_distance )
1030 int fps_factor = round( CLIP_DURATION(average_duration) / CLIP_DURATION(frame->f_duration) * 256 / MBTREE_PRECISION );
1031 float weightdelta = 0.0;
1032 if( ref0_distance && frame->f_weighted_cost_delta[ref0_distance-1] > 0 )
1033 weightdelta = (1.0 - frame->f_weighted_cost_delta[ref0_distance-1]);
1035 /* Allow the strength to be adjusted via qcompress, since the two
1036 * concepts are very similar. */
1037 float strength = 5.0f * (1.0f - h->param.rc.f_qcompress);
1038 for( int mb_index = 0; mb_index < h->mb.i_mb_count; mb_index++ )
1040 int intra_cost = (frame->i_intra_cost[mb_index] * frame->i_inv_qscale_factor[mb_index] + 128) >> 8;
1043 int propagate_cost = (frame->i_propagate_cost[mb_index] * fps_factor + 128) >> 8;
1044 float log2_ratio = x264_log2(intra_cost + propagate_cost) - x264_log2(intra_cost) + weightdelta;
1045 frame->f_qp_offset[mb_index] = frame->f_qp_offset_aq[mb_index] - strength * log2_ratio;
1050 static void x264_macroblock_tree_propagate( x264_t *h, x264_frame_t **frames, float average_duration, int p0, int p1, int b, int referenced )
1052 uint16_t *ref_costs[2] = {frames[p0]->i_propagate_cost,frames[p1]->i_propagate_cost};
1053 int dist_scale_factor = ( ((b-p0) << 8) + ((p1-p0) >> 1) ) / (p1-p0);
1054 int i_bipred_weight = h->param.analyse.b_weighted_bipred ? 64 - (dist_scale_factor>>2) : 32;
1055 int16_t (*mvs[2])[2] = { frames[b]->lowres_mvs[0][b-p0-1], frames[b]->lowres_mvs[1][p1-b-1] };
1056 int bipred_weights[2] = {i_bipred_weight, 64 - i_bipred_weight};
1057 int16_t *buf = h->scratch_buffer;
1058 uint16_t *propagate_cost = frames[b]->i_propagate_cost;
1059 uint16_t *lowres_costs = frames[b]->lowres_costs[b-p0][p1-b];
1062 float fps_factor = CLIP_DURATION(frames[b]->f_duration) / (CLIP_DURATION(average_duration) * 256.0f) * MBTREE_PRECISION;
1064 /* For non-reffed frames the source costs are always zero, so just memset one row and re-use it. */
1066 memset( frames[b]->i_propagate_cost, 0, h->mb.i_mb_width * sizeof(uint16_t) );
1068 for( h->mb.i_mb_y = 0; h->mb.i_mb_y < h->mb.i_mb_height; h->mb.i_mb_y++ )
1070 int mb_index = h->mb.i_mb_y*h->mb.i_mb_stride;
1071 h->mc.mbtree_propagate_cost( buf, propagate_cost,
1072 frames[b]->i_intra_cost+mb_index, lowres_costs+mb_index,
1073 frames[b]->i_inv_qscale_factor+mb_index, &fps_factor, h->mb.i_mb_width );
1075 propagate_cost += h->mb.i_mb_width;
1077 h->mc.mbtree_propagate_list( h, ref_costs[0], &mvs[0][mb_index], buf, &lowres_costs[mb_index],
1078 bipred_weights[0], h->mb.i_mb_y, h->mb.i_mb_width, 0 );
1081 h->mc.mbtree_propagate_list( h, ref_costs[1], &mvs[1][mb_index], buf, &lowres_costs[mb_index],
1082 bipred_weights[1], h->mb.i_mb_y, h->mb.i_mb_width, 1 );
1086 if( h->param.rc.i_vbv_buffer_size && h->param.rc.i_lookahead && referenced )
1087 x264_macroblock_tree_finish( h, frames[b], average_duration, b == p1 ? b - p0 : 0 );
1090 static void x264_macroblock_tree( x264_t *h, x264_mb_analysis_t *a, x264_frame_t **frames, int num_frames, int b_intra )
1093 int last_nonb, cur_nonb = 1;
1097 float total_duration = 0.0;
1098 for( int j = 0; j <= num_frames; j++ )
1099 total_duration += frames[j]->f_duration;
1100 float average_duration = total_duration / (num_frames + 1);
1105 x264_slicetype_frame_cost( h, a, frames, 0, 0, 0, 0 );
1107 while( i > 0 && frames[i]->i_type == X264_TYPE_B )
1111 /* Lookaheadless MB-tree is not a theoretically distinct case; the same extrapolation could
1112 * be applied to the end of a lookahead buffer of any size. However, it's most needed when
1113 * lookahead=0, so that's what's currently implemented. */
1114 if( !h->param.rc.i_lookahead )
1118 memset( frames[0]->i_propagate_cost, 0, h->mb.i_mb_count * sizeof(uint16_t) );
1119 memcpy( frames[0]->f_qp_offset, frames[0]->f_qp_offset_aq, h->mb.i_mb_count * sizeof(float) );
1122 XCHG( uint16_t*, frames[last_nonb]->i_propagate_cost, frames[0]->i_propagate_cost );
1123 memset( frames[0]->i_propagate_cost, 0, h->mb.i_mb_count * sizeof(uint16_t) );
1127 if( last_nonb < idx )
1129 memset( frames[last_nonb]->i_propagate_cost, 0, h->mb.i_mb_count * sizeof(uint16_t) );
1135 while( frames[cur_nonb]->i_type == X264_TYPE_B && cur_nonb > 0 )
1137 if( cur_nonb < idx )
1139 x264_slicetype_frame_cost( h, a, frames, cur_nonb, last_nonb, last_nonb, 0 );
1140 memset( frames[cur_nonb]->i_propagate_cost, 0, h->mb.i_mb_count * sizeof(uint16_t) );
1141 bframes = last_nonb - cur_nonb - 1;
1142 if( h->param.i_bframe_pyramid && bframes > 1 )
1144 int middle = (bframes + 1)/2 + cur_nonb;
1145 x264_slicetype_frame_cost( h, a, frames, cur_nonb, last_nonb, middle, 0 );
1146 memset( frames[middle]->i_propagate_cost, 0, h->mb.i_mb_count * sizeof(uint16_t) );
1147 while( i > cur_nonb )
1149 int p0 = i > middle ? middle : cur_nonb;
1150 int p1 = i < middle ? middle : last_nonb;
1153 x264_slicetype_frame_cost( h, a, frames, p0, p1, i, 0 );
1154 x264_macroblock_tree_propagate( h, frames, average_duration, p0, p1, i, 0 );
1158 x264_macroblock_tree_propagate( h, frames, average_duration, cur_nonb, last_nonb, middle, 1 );
1162 while( i > cur_nonb )
1164 x264_slicetype_frame_cost( h, a, frames, cur_nonb, last_nonb, i, 0 );
1165 x264_macroblock_tree_propagate( h, frames, average_duration, cur_nonb, last_nonb, i, 0 );
1169 x264_macroblock_tree_propagate( h, frames, average_duration, cur_nonb, last_nonb, last_nonb, 1 );
1170 last_nonb = cur_nonb;
1173 if( !h->param.rc.i_lookahead )
1175 x264_slicetype_frame_cost( h, a, frames, 0, last_nonb, last_nonb, 0 );
1176 x264_macroblock_tree_propagate( h, frames, average_duration, 0, last_nonb, last_nonb, 1 );
1177 XCHG( uint16_t*, frames[last_nonb]->i_propagate_cost, frames[0]->i_propagate_cost );
1180 x264_macroblock_tree_finish( h, frames[last_nonb], average_duration, last_nonb );
1181 if( h->param.i_bframe_pyramid && bframes > 1 && !h->param.rc.i_vbv_buffer_size )
1182 x264_macroblock_tree_finish( h, frames[last_nonb+(bframes+1)/2], average_duration, 0 );
1185 static int x264_vbv_frame_cost( x264_t *h, x264_mb_analysis_t *a, x264_frame_t **frames, int p0, int p1, int b )
1187 int cost = x264_slicetype_frame_cost( h, a, frames, p0, p1, b, 0 );
1188 if( h->param.rc.i_aq_mode )
1190 if( h->param.rc.b_mb_tree )
1191 return x264_slicetype_frame_cost_recalculate( h, frames, p0, p1, b );
1193 return frames[b]->i_cost_est_aq[b-p0][p1-b];
1198 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 )
1200 cur_frame->i_cpb_delay = *i_cpb_delay;
1201 cur_frame->i_dpb_output_delay = cur_frame->i_field_cnt - *i_coded_fields;
1203 // add a correction term for frame reordering
1204 cur_frame->i_dpb_output_delay += h->sps->vui.i_num_reorder_frames*2;
1206 // fix possible negative dpb_output_delay because of pulldown changes and reordering
1207 if( cur_frame->i_dpb_output_delay < 0 )
1209 cur_frame->i_cpb_delay += cur_frame->i_dpb_output_delay;
1210 cur_frame->i_dpb_output_delay = 0;
1212 prev_frame->i_cpb_duration += cur_frame->i_dpb_output_delay;
1215 // don't reset cpb delay for IDR frames when using intra-refresh
1216 if( cur_frame->b_keyframe && !h->param.b_intra_refresh )
1219 *i_cpb_delay += cur_frame->i_duration;
1220 *i_coded_fields += cur_frame->i_duration;
1221 cur_frame->i_cpb_duration = cur_frame->i_duration;
1224 static void x264_vbv_lookahead( x264_t *h, x264_mb_analysis_t *a, x264_frame_t **frames, int num_frames, int keyframe )
1226 int last_nonb = 0, cur_nonb = 1, idx = 0;
1227 x264_frame_t *prev_frame = NULL;
1228 int prev_frame_idx = 0;
1229 while( cur_nonb < num_frames && frames[cur_nonb]->i_type == X264_TYPE_B )
1231 int next_nonb = keyframe ? last_nonb : cur_nonb;
1233 if( frames[cur_nonb]->i_coded_fields_lookahead >= 0 )
1235 h->i_coded_fields_lookahead = frames[cur_nonb]->i_coded_fields_lookahead;
1236 h->i_cpb_delay_lookahead = frames[cur_nonb]->i_cpb_delay_lookahead;
1239 while( cur_nonb < num_frames )
1241 /* P/I cost: This shouldn't include the cost of next_nonb */
1242 if( next_nonb != cur_nonb )
1244 int p0 = IS_X264_TYPE_I( frames[cur_nonb]->i_type ) ? cur_nonb : last_nonb;
1245 frames[next_nonb]->i_planned_satd[idx] = x264_vbv_frame_cost( h, a, frames, p0, cur_nonb, cur_nonb );
1246 frames[next_nonb]->i_planned_type[idx] = frames[cur_nonb]->i_type;
1247 frames[cur_nonb]->i_coded_fields_lookahead = h->i_coded_fields_lookahead;
1248 frames[cur_nonb]->i_cpb_delay_lookahead = h->i_cpb_delay_lookahead;
1249 x264_calculate_durations( h, frames[cur_nonb], prev_frame, &h->i_cpb_delay_lookahead, &h->i_coded_fields_lookahead );
1252 frames[next_nonb]->f_planned_cpb_duration[prev_frame_idx] = (double)prev_frame->i_cpb_duration *
1253 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1255 frames[next_nonb]->f_planned_cpb_duration[idx] = (double)frames[cur_nonb]->i_cpb_duration *
1256 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1257 prev_frame = frames[cur_nonb];
1258 prev_frame_idx = idx;
1261 /* Handle the B-frames: coded order */
1262 for( int i = last_nonb+1; i < cur_nonb; i++, idx++ )
1264 frames[next_nonb]->i_planned_satd[idx] = x264_vbv_frame_cost( h, a, frames, last_nonb, cur_nonb, i );
1265 frames[next_nonb]->i_planned_type[idx] = X264_TYPE_B;
1266 frames[i]->i_coded_fields_lookahead = h->i_coded_fields_lookahead;
1267 frames[i]->i_cpb_delay_lookahead = h->i_cpb_delay_lookahead;
1268 x264_calculate_durations( h, frames[i], prev_frame, &h->i_cpb_delay_lookahead, &h->i_coded_fields_lookahead );
1271 frames[next_nonb]->f_planned_cpb_duration[prev_frame_idx] = (double)prev_frame->i_cpb_duration *
1272 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1274 frames[next_nonb]->f_planned_cpb_duration[idx] = (double)frames[i]->i_cpb_duration *
1275 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1276 prev_frame = frames[i];
1277 prev_frame_idx = idx;
1279 last_nonb = cur_nonb;
1281 while( cur_nonb <= num_frames && frames[cur_nonb]->i_type == X264_TYPE_B )
1284 frames[next_nonb]->i_planned_type[idx] = X264_TYPE_AUTO;
1287 static int x264_slicetype_path_cost( x264_t *h, x264_mb_analysis_t *a, x264_frame_t **frames, char *path, int threshold )
1292 path--; /* Since the 1st path element is really the second frame */
1296 /* Find the location of the next P-frame. */
1297 while( path[next_p] != 'P' )
1300 /* Add the cost of the P-frame found above */
1301 cost += x264_slicetype_frame_cost( h, a, frames, cur_p, next_p, next_p, 0 );
1302 /* Early terminate if the cost we have found is larger than the best path cost so far */
1303 if( cost > threshold )
1306 if( h->param.i_bframe_pyramid && next_p - cur_p > 2 )
1308 int middle = cur_p + (next_p - cur_p)/2;
1309 cost += x264_slicetype_frame_cost( h, a, frames, cur_p, next_p, middle, 0 );
1310 for( int next_b = loc; next_b < middle && cost < threshold; next_b++ )
1311 cost += x264_slicetype_frame_cost( h, a, frames, cur_p, middle, next_b, 0 );
1312 for( int next_b = middle+1; next_b < next_p && cost < threshold; next_b++ )
1313 cost += x264_slicetype_frame_cost( h, a, frames, middle, next_p, next_b, 0 );
1316 for( int next_b = loc; next_b < next_p && cost < threshold; next_b++ )
1317 cost += x264_slicetype_frame_cost( h, a, frames, cur_p, next_p, next_b, 0 );
1325 /* Viterbi/trellis slicetype decision algorithm. */
1326 /* Uses strings due to the fact that the speed of the control functions is
1327 negligible compared to the cost of running slicetype_frame_cost, and because
1328 it makes debugging easier. */
1329 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] )
1331 char paths[2][X264_LOOKAHEAD_MAX+1];
1332 int num_paths = X264_MIN( h->param.i_bframe+1, length );
1333 int best_cost = COST_MAX;
1336 /* Iterate over all currently possible paths */
1337 for( int path = 0; path < num_paths; path++ )
1339 /* Add suffixes to the current path */
1340 int len = length - (path + 1);
1341 memcpy( paths[idx], best_paths[len % (X264_BFRAME_MAX+1)], len );
1342 memset( paths[idx]+len, 'B', path );
1343 strcpy( paths[idx]+len+path, "P" );
1345 /* Calculate the actual cost of the current path */
1346 int cost = x264_slicetype_path_cost( h, a, frames, paths[idx], best_cost );
1347 if( cost < best_cost )
1354 /* Store the best path. */
1355 memcpy( best_paths[length % (X264_BFRAME_MAX+1)], paths[idx^1], length );
1358 static int scenecut_internal( x264_t *h, x264_mb_analysis_t *a, x264_frame_t **frames, int p0, int p1, int real_scenecut )
1360 x264_frame_t *frame = frames[p1];
1362 /* Don't do scenecuts on the right view of a frame-packed video. */
1363 if( real_scenecut && h->param.i_frame_packing == 5 && (frame->i_frame&1) )
1366 x264_slicetype_frame_cost( h, a, frames, p0, p1, p1, 0 );
1368 int icost = frame->i_cost_est[0][0];
1369 int pcost = frame->i_cost_est[p1-p0][0];
1371 int i_gop_size = frame->i_frame - h->lookahead->i_last_keyframe;
1372 float f_thresh_max = h->param.i_scenecut_threshold / 100.0;
1373 /* magic numbers pulled out of thin air */
1374 float f_thresh_min = f_thresh_max * 0.25;
1377 if( h->param.i_keyint_min == h->param.i_keyint_max )
1378 f_thresh_min = f_thresh_max;
1379 if( i_gop_size <= h->param.i_keyint_min / 4 || h->param.b_intra_refresh )
1380 f_bias = f_thresh_min / 4;
1381 else if( i_gop_size <= h->param.i_keyint_min )
1382 f_bias = f_thresh_min * i_gop_size / h->param.i_keyint_min;
1385 f_bias = f_thresh_min
1386 + ( f_thresh_max - f_thresh_min )
1387 * ( i_gop_size - h->param.i_keyint_min )
1388 / ( h->param.i_keyint_max - h->param.i_keyint_min );
1391 res = pcost >= (1.0 - f_bias) * icost;
1392 if( res && real_scenecut )
1394 int imb = frame->i_intra_mbs[p1-p0];
1395 int pmb = NUM_MBS - imb;
1396 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",
1398 icost, pcost, 1. - (double)pcost / icost,
1399 f_bias, i_gop_size, imb, pmb );
1404 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 )
1406 /* Only do analysis during a normal scenecut check. */
1407 if( real_scenecut && h->param.i_bframe )
1409 int origmaxp1 = p0 + 1;
1410 /* Look ahead to avoid coding short flashes as scenecuts. */
1411 if( h->param.i_bframe_adaptive == X264_B_ADAPT_TRELLIS )
1412 /* Don't analyse any more frames than the trellis would have covered. */
1413 origmaxp1 += h->param.i_bframe;
1416 int maxp1 = X264_MIN( origmaxp1, num_frames );
1418 /* Where A and B are scenes: AAAAAABBBAAAAAA
1419 * If BBB is shorter than (maxp1-p0), it is detected as a flash
1420 * and not considered a scenecut. */
1421 for( int curp1 = p1; curp1 <= maxp1; curp1++ )
1422 if( !scenecut_internal( h, a, frames, p0, curp1, 0 ) )
1423 /* Any frame in between p0 and cur_p1 cannot be a real scenecut. */
1424 for( int i = curp1; i > p0; i-- )
1425 frames[i]->b_scenecut = 0;
1427 /* Where A-F are scenes: AAAAABBCCDDEEFFFFFF
1428 * If each of BB ... EE are shorter than (maxp1-p0), they are
1429 * detected as flashes and not considered scenecuts.
1430 * Instead, the first F frame becomes a scenecut.
1431 * If the video ends before F, no frame becomes a scenecut. */
1432 for( int curp0 = p0; curp0 <= maxp1; curp0++ )
1433 if( origmaxp1 > i_max_search || (curp0 < maxp1 && scenecut_internal( h, a, frames, curp0, maxp1, 0 )) )
1434 /* If cur_p0 is the p0 of a scenecut, it cannot be the p1 of a scenecut. */
1435 frames[curp0]->b_scenecut = 0;
1438 /* Ignore frames that are part of a flash, i.e. cannot be real scenecuts. */
1439 if( !frames[p1]->b_scenecut )
1441 return scenecut_internal( h, a, frames, p0, p1, real_scenecut );
1444 void x264_slicetype_analyse( x264_t *h, int intra_minigop )
1446 x264_mb_analysis_t a;
1447 x264_frame_t *frames[X264_LOOKAHEAD_MAX+3] = { NULL, };
1448 int num_frames, orig_num_frames, keyint_limit, framecnt;
1449 int i_mb_count = NUM_MBS;
1450 int cost1p0, cost2p0, cost1b1, cost2p1;
1451 int i_max_search = X264_MIN( h->lookahead->next.i_size, X264_LOOKAHEAD_MAX );
1452 int vbv_lookahead = h->param.rc.i_vbv_buffer_size && h->param.rc.i_lookahead;
1453 /* For determinism we should limit the search to the number of frames lookahead has for sure
1454 * in h->lookahead->next.list buffer, except at the end of stream.
1455 * For normal calls with (intra_minigop == 0) that is h->lookahead->i_slicetype_length + 1 frames.
1456 * And for I-frame calls (intra_minigop != 0) we already removed intra_minigop frames from there. */
1457 if( h->param.b_deterministic )
1458 i_max_search = X264_MIN( i_max_search, h->lookahead->i_slicetype_length + 1 - intra_minigop );
1459 int keyframe = !!intra_minigop;
1461 assert( h->frames.b_have_lowres );
1463 if( !h->lookahead->last_nonb )
1465 frames[0] = h->lookahead->last_nonb;
1466 for( framecnt = 0; framecnt < i_max_search && h->lookahead->next.list[framecnt]->i_type == X264_TYPE_AUTO; framecnt++ )
1467 frames[framecnt+1] = h->lookahead->next.list[framecnt];
1469 x264_lowres_context_init( h, &a );
1473 if( h->param.rc.b_mb_tree )
1474 x264_macroblock_tree( h, &a, frames, 0, keyframe );
1478 keyint_limit = h->param.i_keyint_max - frames[0]->i_frame + h->lookahead->i_last_keyframe - 1;
1479 orig_num_frames = num_frames = h->param.b_intra_refresh ? framecnt : X264_MIN( framecnt, keyint_limit );
1481 /* This is important psy-wise: if we have a non-scenecut keyframe,
1482 * there will be significant visual artifacts if the frames just before
1483 * go down in quality due to being referenced less, despite it being
1484 * more RD-optimal. */
1485 if( (h->param.analyse.b_psy && h->param.rc.b_mb_tree) || vbv_lookahead )
1486 num_frames = framecnt;
1487 else if( h->param.b_open_gop && num_frames < framecnt )
1489 else if( num_frames == 0 )
1491 frames[1]->i_type = X264_TYPE_I;
1495 int num_bframes = 0;
1496 int num_analysed_frames = num_frames;
1498 if( h->param.i_scenecut_threshold && scenecut( h, &a, frames, 0, 1, 1, orig_num_frames, i_max_search ) )
1500 frames[1]->i_type = X264_TYPE_I;
1505 x264_opencl_slicetype_prep( h, frames, num_frames, a.i_lambda );
1508 if( h->param.i_bframe )
1510 if( h->param.i_bframe_adaptive == X264_B_ADAPT_TRELLIS )
1512 if( num_frames > 1 )
1514 char best_paths[X264_BFRAME_MAX+1][X264_LOOKAHEAD_MAX+1] = {"","P"};
1515 int best_path_index = num_frames % (X264_BFRAME_MAX+1);
1517 /* Perform the frametype analysis. */
1518 for( int j = 2; j <= num_frames; j++ )
1519 x264_slicetype_path( h, &a, frames, j, best_paths );
1521 num_bframes = strspn( best_paths[best_path_index], "B" );
1522 /* Load the results of the analysis into the frame types. */
1523 for( int j = 1; j < num_frames; j++ )
1524 frames[j]->i_type = best_paths[best_path_index][j-1] == 'B' ? X264_TYPE_B : X264_TYPE_P;
1526 frames[num_frames]->i_type = X264_TYPE_P;
1528 else if( h->param.i_bframe_adaptive == X264_B_ADAPT_FAST )
1530 for( int i = 0; i <= num_frames-2; )
1532 cost2p1 = x264_slicetype_frame_cost( h, &a, frames, i+0, i+2, i+2, 1 );
1533 if( frames[i+2]->i_intra_mbs[2] > i_mb_count / 2 )
1535 frames[i+1]->i_type = X264_TYPE_P;
1536 frames[i+2]->i_type = X264_TYPE_P;
1542 if( h->param.b_opencl )
1544 int b_work_done = 0;
1545 b_work_done |= x264_opencl_precalculate_frame_cost(h, frames, a.i_lambda, i+0, i+2, i+1 );
1546 b_work_done |= x264_opencl_precalculate_frame_cost(h, frames, a.i_lambda, i+0, i+1, i+1 );
1547 b_work_done |= x264_opencl_precalculate_frame_cost(h, frames, a.i_lambda, i+1, i+2, i+2 );
1549 x264_opencl_flush( h );
1553 cost1b1 = x264_slicetype_frame_cost( h, &a, frames, i+0, i+2, i+1, 0 );
1554 cost1p0 = x264_slicetype_frame_cost( h, &a, frames, i+0, i+1, i+1, 0 );
1555 cost2p0 = x264_slicetype_frame_cost( h, &a, frames, i+1, i+2, i+2, 0 );
1557 if( cost1p0 + cost2p0 < cost1b1 + cost2p1 )
1559 frames[i+1]->i_type = X264_TYPE_P;
1564 // arbitrary and untuned
1565 #define INTER_THRESH 300
1566 #define P_SENS_BIAS (50 - h->param.i_bframe_bias)
1567 frames[i+1]->i_type = X264_TYPE_B;
1570 for( j = i+2; j <= X264_MIN( i+h->param.i_bframe, num_frames-1 ); j++ )
1572 int pthresh = X264_MAX(INTER_THRESH - P_SENS_BIAS * (j-i-1), INTER_THRESH/10);
1573 int pcost = x264_slicetype_frame_cost( h, &a, frames, i+0, j+1, j+1, 1 );
1574 if( pcost > pthresh*i_mb_count || frames[j+1]->i_intra_mbs[j-i+1] > i_mb_count/3 )
1576 frames[j]->i_type = X264_TYPE_B;
1578 frames[j]->i_type = X264_TYPE_P;
1581 frames[num_frames]->i_type = X264_TYPE_P;
1583 while( num_bframes < num_frames && frames[num_bframes+1]->i_type == X264_TYPE_B )
1588 num_bframes = X264_MIN(num_frames-1, h->param.i_bframe);
1589 for( int j = 1; j < num_frames; j++ )
1590 frames[j]->i_type = (j%(num_bframes+1)) ? X264_TYPE_B : X264_TYPE_P;
1591 frames[num_frames]->i_type = X264_TYPE_P;
1594 /* Check scenecut on the first minigop. */
1595 for( int j = 1; j < num_bframes+1; j++ )
1596 if( h->param.i_scenecut_threshold && scenecut( h, &a, frames, j, j+1, 0, orig_num_frames, i_max_search ) )
1598 frames[j]->i_type = X264_TYPE_P;
1599 num_analysed_frames = j;
1603 reset_start = keyframe ? 1 : X264_MIN( num_bframes+2, num_analysed_frames+1 );
1607 for( int j = 1; j <= num_frames; j++ )
1608 frames[j]->i_type = X264_TYPE_P;
1609 reset_start = !keyframe + 1;
1613 /* Perform the actual macroblock tree analysis.
1614 * Don't go farther than the maximum keyframe interval; this helps in short GOPs. */
1615 if( h->param.rc.b_mb_tree )
1616 x264_macroblock_tree( h, &a, frames, X264_MIN(num_frames, h->param.i_keyint_max), keyframe );
1618 /* Enforce keyframe limit. */
1619 if( !h->param.b_intra_refresh )
1620 for( int i = keyint_limit+1; i <= num_frames; i += h->param.i_keyint_max )
1622 frames[i]->i_type = X264_TYPE_I;
1623 reset_start = X264_MIN( reset_start, i+1 );
1624 if( h->param.b_open_gop && h->param.b_bluray_compat )
1625 while( IS_X264_TYPE_B( frames[i-1]->i_type ) )
1630 x264_vbv_lookahead( h, &a, frames, num_frames, keyframe );
1632 /* Restore frametypes for all frames that haven't actually been decided yet. */
1633 for( int j = reset_start; j <= num_frames; j++ )
1634 frames[j]->i_type = X264_TYPE_AUTO;
1637 x264_opencl_slicetype_end( h );
1641 void x264_slicetype_decide( x264_t *h )
1643 x264_frame_t *frames[X264_BFRAME_MAX+2];
1648 if( !h->lookahead->next.i_size )
1651 int lookahead_size = h->lookahead->next.i_size;
1653 for( int i = 0; i < h->lookahead->next.i_size; i++ )
1655 if( h->param.b_vfr_input )
1657 if( lookahead_size-- > 1 )
1658 h->lookahead->next.list[i]->i_duration = 2 * (h->lookahead->next.list[i+1]->i_pts - h->lookahead->next.list[i]->i_pts);
1660 h->lookahead->next.list[i]->i_duration = h->i_prev_duration;
1663 h->lookahead->next.list[i]->i_duration = delta_tfi_divisor[h->lookahead->next.list[i]->i_pic_struct];
1664 h->i_prev_duration = h->lookahead->next.list[i]->i_duration;
1665 h->lookahead->next.list[i]->f_duration = (double)h->lookahead->next.list[i]->i_duration
1666 * h->sps->vui.i_num_units_in_tick
1667 / h->sps->vui.i_time_scale;
1669 if( h->lookahead->next.list[i]->i_frame > h->i_disp_fields_last_frame && lookahead_size > 0 )
1671 h->lookahead->next.list[i]->i_field_cnt = h->i_disp_fields;
1672 h->i_disp_fields += h->lookahead->next.list[i]->i_duration;
1673 h->i_disp_fields_last_frame = h->lookahead->next.list[i]->i_frame;
1675 else if( lookahead_size == 0 )
1677 h->lookahead->next.list[i]->i_field_cnt = h->i_disp_fields;
1678 h->lookahead->next.list[i]->i_duration = h->i_prev_duration;
1682 if( h->param.rc.b_stat_read )
1684 /* Use the frame types from the first pass */
1685 for( int i = 0; i < h->lookahead->next.i_size; i++ )
1686 h->lookahead->next.list[i]->i_type =
1687 x264_ratecontrol_slice_type( h, h->lookahead->next.list[i]->i_frame );
1689 else if( (h->param.i_bframe && h->param.i_bframe_adaptive)
1690 || h->param.i_scenecut_threshold
1691 || h->param.rc.b_mb_tree
1692 || (h->param.rc.i_vbv_buffer_size && h->param.rc.i_lookahead) )
1693 x264_slicetype_analyse( h, 0 );
1695 for( bframes = 0, brefs = 0;; bframes++ )
1697 frm = h->lookahead->next.list[bframes];
1698 if( frm->i_type == X264_TYPE_BREF && h->param.i_bframe_pyramid < X264_B_PYRAMID_NORMAL &&
1699 brefs == h->param.i_bframe_pyramid )
1701 frm->i_type = X264_TYPE_B;
1702 x264_log( h, X264_LOG_WARNING, "B-ref at frame %d incompatible with B-pyramid %s \n",
1703 frm->i_frame, x264_b_pyramid_names[h->param.i_bframe_pyramid] );
1705 /* pyramid with multiple B-refs needs a big enough dpb that the preceding P-frame stays available.
1706 smaller dpb could be supported by smart enough use of mmco, but it's easier just to forbid it. */
1707 else if( frm->i_type == X264_TYPE_BREF && h->param.i_bframe_pyramid == X264_B_PYRAMID_NORMAL &&
1708 brefs && h->param.i_frame_reference <= (brefs+3) )
1710 frm->i_type = X264_TYPE_B;
1711 x264_log( h, X264_LOG_WARNING, "B-ref at frame %d incompatible with B-pyramid %s and %d reference frames\n",
1712 frm->i_frame, x264_b_pyramid_names[h->param.i_bframe_pyramid], h->param.i_frame_reference );
1715 if( frm->i_type == X264_TYPE_KEYFRAME )
1716 frm->i_type = h->param.b_open_gop ? X264_TYPE_I : X264_TYPE_IDR;
1718 /* Limit GOP size */
1719 if( (!h->param.b_intra_refresh || frm->i_frame == 0) && frm->i_frame - h->lookahead->i_last_keyframe >= h->param.i_keyint_max )
1721 if( frm->i_type == X264_TYPE_AUTO || frm->i_type == X264_TYPE_I )
1722 frm->i_type = h->param.b_open_gop && h->lookahead->i_last_keyframe >= 0 ? X264_TYPE_I : X264_TYPE_IDR;
1723 int warn = frm->i_type != X264_TYPE_IDR;
1724 if( warn && h->param.b_open_gop )
1725 warn &= frm->i_type != X264_TYPE_I;
1728 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 );
1729 frm->i_type = h->param.b_open_gop && h->lookahead->i_last_keyframe >= 0 ? X264_TYPE_I : X264_TYPE_IDR;
1732 if( frm->i_type == X264_TYPE_I && frm->i_frame - h->lookahead->i_last_keyframe >= h->param.i_keyint_min )
1734 if( h->param.b_open_gop )
1736 h->lookahead->i_last_keyframe = frm->i_frame; // Use display order
1737 if( h->param.b_bluray_compat )
1738 h->lookahead->i_last_keyframe -= bframes; // Use bluray order
1739 frm->b_keyframe = 1;
1742 frm->i_type = X264_TYPE_IDR;
1744 if( frm->i_type == X264_TYPE_IDR )
1747 h->lookahead->i_last_keyframe = frm->i_frame;
1748 frm->b_keyframe = 1;
1752 h->lookahead->next.list[bframes]->i_type = X264_TYPE_P;
1756 if( bframes == h->param.i_bframe ||
1757 !h->lookahead->next.list[bframes+1] )
1759 if( IS_X264_TYPE_B( frm->i_type ) )
1760 x264_log( h, X264_LOG_WARNING, "specified frame type is not compatible with max B-frames\n" );
1761 if( frm->i_type == X264_TYPE_AUTO
1762 || IS_X264_TYPE_B( frm->i_type ) )
1763 frm->i_type = X264_TYPE_P;
1766 if( frm->i_type == X264_TYPE_BREF )
1769 if( frm->i_type == X264_TYPE_AUTO )
1770 frm->i_type = X264_TYPE_B;
1772 else if( !IS_X264_TYPE_B( frm->i_type ) ) break;
1776 h->lookahead->next.list[bframes-1]->b_last_minigop_bframe = 1;
1777 h->lookahead->next.list[bframes]->i_bframes = bframes;
1779 /* insert a bref into the sequence */
1780 if( h->param.i_bframe_pyramid && bframes > 1 && !brefs )
1782 h->lookahead->next.list[bframes/2]->i_type = X264_TYPE_BREF;
1786 /* calculate the frame costs ahead of time for x264_rc_analyse_slice while we still have lowres */
1787 if( h->param.rc.i_rc_method != X264_RC_CQP )
1789 x264_mb_analysis_t a;
1791 p1 = b = bframes + 1;
1793 x264_lowres_context_init( h, &a );
1795 frames[0] = h->lookahead->last_nonb;
1796 memcpy( &frames[1], h->lookahead->next.list, (bframes+1) * sizeof(x264_frame_t*) );
1797 if( IS_X264_TYPE_I( h->lookahead->next.list[bframes]->i_type ) )
1802 x264_slicetype_frame_cost( h, &a, frames, p0, p1, b, 0 );
1804 if( (p0 != p1 || bframes) && h->param.rc.i_vbv_buffer_size )
1806 /* We need the intra costs for row SATDs. */
1807 x264_slicetype_frame_cost( h, &a, frames, b, b, b, 0 );
1809 /* We need B-frame costs for row SATDs. */
1811 for( b = 1; b <= bframes; b++ )
1813 if( frames[b]->i_type == X264_TYPE_B )
1814 for( p1 = b; frames[p1]->i_type == X264_TYPE_B; )
1818 x264_slicetype_frame_cost( h, &a, frames, p0, p1, b, 0 );
1819 if( frames[b]->i_type == X264_TYPE_BREF )
1825 /* Analyse for weighted P frames */
1826 if( !h->param.rc.b_stat_read && h->lookahead->next.list[bframes]->i_type == X264_TYPE_P
1827 && h->param.analyse.i_weighted_pred >= X264_WEIGHTP_SIMPLE )
1830 x264_weights_analyse( h, h->lookahead->next.list[bframes], h->lookahead->last_nonb, 0 );
1833 /* shift sequence to coded order.
1834 use a small temporary list to avoid shifting the entire next buffer around */
1835 int i_coded = h->lookahead->next.list[0]->i_frame;
1838 int idx_list[] = { brefs+1, 1 };
1839 for( int i = 0; i < bframes; i++ )
1841 int idx = idx_list[h->lookahead->next.list[i]->i_type == X264_TYPE_BREF]++;
1842 frames[idx] = h->lookahead->next.list[i];
1843 frames[idx]->i_reordered_pts = h->lookahead->next.list[idx]->i_pts;
1845 frames[0] = h->lookahead->next.list[bframes];
1846 frames[0]->i_reordered_pts = h->lookahead->next.list[0]->i_pts;
1847 memcpy( h->lookahead->next.list, frames, (bframes+1) * sizeof(x264_frame_t*) );
1850 for( int i = 0; i <= bframes; i++ )
1852 h->lookahead->next.list[i]->i_coded = i_coded++;
1855 x264_calculate_durations( h, h->lookahead->next.list[i], h->lookahead->next.list[i-1], &h->i_cpb_delay, &h->i_coded_fields );
1856 h->lookahead->next.list[0]->f_planned_cpb_duration[i-1] = (double)h->lookahead->next.list[i]->i_cpb_duration *
1857 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1860 x264_calculate_durations( h, h->lookahead->next.list[i], NULL, &h->i_cpb_delay, &h->i_coded_fields );
1864 int x264_rc_analyse_slice( x264_t *h )
1870 if( IS_X264_TYPE_I(h->fenc->i_type) )
1872 else if( h->fenc->i_type == X264_TYPE_P )
1873 p1 = b = h->fenc->i_bframes + 1;
1876 p1 = (h->fref_nearest[1]->i_poc - h->fref_nearest[0]->i_poc)/2;
1877 b = (h->fenc->i_poc - h->fref_nearest[0]->i_poc)/2;
1879 /* We don't need to assign p0/p1 since we are not performing any real analysis here. */
1880 x264_frame_t **frames = &h->fenc - b;
1882 /* cost should have been already calculated by x264_slicetype_decide */
1883 cost = frames[b]->i_cost_est[b-p0][p1-b];
1884 assert( cost >= 0 );
1886 if( h->param.rc.b_mb_tree && !h->param.rc.b_stat_read )
1888 cost = x264_slicetype_frame_cost_recalculate( h, frames, p0, p1, b );
1889 if( b && h->param.rc.i_vbv_buffer_size )
1890 x264_slicetype_frame_cost_recalculate( h, frames, b, b, b );
1892 /* In AQ, use the weighted score instead. */
1893 else if( h->param.rc.i_aq_mode )
1894 cost = frames[b]->i_cost_est_aq[b-p0][p1-b];
1896 h->fenc->i_row_satd = h->fenc->i_row_satds[b-p0][p1-b];
1897 h->fdec->i_row_satd = h->fdec->i_row_satds[b-p0][p1-b];
1898 h->fdec->i_satd = cost;
1899 memcpy( h->fdec->i_row_satd, h->fenc->i_row_satd, h->mb.i_mb_height * sizeof(int) );
1900 if( !IS_X264_TYPE_I(h->fenc->i_type) )
1901 memcpy( h->fdec->i_row_satds[0][0], h->fenc->i_row_satds[0][0], h->mb.i_mb_height * sizeof(int) );
1903 if( h->param.b_intra_refresh && h->param.rc.i_vbv_buffer_size && h->fenc->i_type == X264_TYPE_P )
1905 int ip_factor = 256 * h->param.rc.f_ip_factor; /* fix8 */
1906 for( int y = 0; y < h->mb.i_mb_height; y++ )
1908 int mb_xy = y * h->mb.i_mb_stride + h->fdec->i_pir_start_col;
1909 for( int x = h->fdec->i_pir_start_col; x <= h->fdec->i_pir_end_col; x++, mb_xy++ )
1911 int intra_cost = (h->fenc->i_intra_cost[mb_xy] * ip_factor + 128) >> 8;
1912 int inter_cost = h->fenc->lowres_costs[b-p0][p1-b][mb_xy] & LOWRES_COST_MASK;
1913 int diff = intra_cost - inter_cost;
1914 if( h->param.rc.i_aq_mode )
1915 h->fdec->i_row_satd[y] += (diff * frames[b]->i_inv_qscale_factor[mb_xy] + 128) >> 8;
1917 h->fdec->i_row_satd[y] += diff;
1924 for( int y = 0; y < h->mb.i_mb_height; y++ )
1925 h->fdec->i_row_satd[y] >>= (BIT_DEPTH - 8);
1927 return cost >> (BIT_DEPTH - 8);