1 /*****************************************************************************
2 * slicetype.c: lookahead analysis
3 *****************************************************************************
4 * Copyright (C) 2005-2021 x264 project
5 *
6 * Authors: Fiona Glaser <fiona@x264.com>
7 * Loren Merritt <lorenm@u.washington.edu>
8 * Dylan Yudaken <dyudaken@gmail.com>
9 *
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.
14 *
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.
19 *
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.
23 *
24 * This program is also available under a commercial proprietary license.
25 * For more information, contact us at licensing@x264.com.
26 *****************************************************************************/
27
28 #include "common/common.h"
29 #include "macroblock.h"
30 #include "me.h"
31
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 };
34
35 static int slicetype_frame_cost( x264_t *h, x264_mb_analysis_t *a,
36 x264_frame_t **frames, int p0, int p1, int b );
37
38 #define x264_weights_analyse x264_template(weights_analyse)
39 void x264_weights_analyse( x264_t *h, x264_frame_t *fenc, x264_frame_t *ref, int b_lookahead );
40
41 #if HAVE_OPENCL
42 #include "slicetype-cl.h"
43 #endif
44
lowres_context_init(x264_t * h,x264_mb_analysis_t * a)45 static void lowres_context_init( x264_t *h, x264_mb_analysis_t *a )
46 {
47 a->i_qp = X264_LOOKAHEAD_QP;
48 a->i_lambda = x264_lambda_tab[ a->i_qp ];
49 mb_analyse_load_costs( h, a );
50 if( h->param.analyse.i_subpel_refine > 1 )
51 {
52 h->mb.i_me_method = X264_MIN( X264_ME_HEX, h->param.analyse.i_me_method );
53 h->mb.i_subpel_refine = 4;
54 }
55 else
56 {
57 h->mb.i_me_method = X264_ME_DIA;
58 h->mb.i_subpel_refine = 2;
59 }
60 h->mb.b_chroma_me = 0;
61 }
62
63 /* makes a non-h264 weight (i.e. fix7), into an h264 weight */
weight_get_h264(int weight_nonh264,int offset,x264_weight_t * w)64 static void weight_get_h264( int weight_nonh264, int offset, x264_weight_t *w )
65 {
66 w->i_offset = offset;
67 w->i_denom = 7;
68 w->i_scale = weight_nonh264;
69 while( w->i_denom > 0 && (w->i_scale > 127) )
70 {
71 w->i_denom--;
72 w->i_scale >>= 1;
73 }
74 w->i_scale = X264_MIN( w->i_scale, 127 );
75 }
76
weight_cost_init_luma(x264_t * h,x264_frame_t * fenc,x264_frame_t * ref,pixel * dest)77 static NOINLINE pixel *weight_cost_init_luma( x264_t *h, x264_frame_t *fenc, x264_frame_t *ref, pixel *dest )
78 {
79 int ref0_distance = fenc->i_frame - ref->i_frame - 1;
80 /* Note: this will never run during lookahead as weights_analyse is only called if no
81 * motion search has been done. */
82 if( fenc->lowres_mvs[0][ref0_distance][0][0] != 0x7FFF )
83 {
84 int i_stride = fenc->i_stride_lowres;
85 int i_lines = fenc->i_lines_lowres;
86 int i_width = fenc->i_width_lowres;
87 int i_mb_xy = 0;
88 pixel *p = dest;
89
90 for( int y = 0; y < i_lines; y += 8, p += i_stride*8 )
91 for( int x = 0; x < i_width; x += 8, i_mb_xy++ )
92 {
93 int mvx = fenc->lowres_mvs[0][ref0_distance][i_mb_xy][0];
94 int mvy = fenc->lowres_mvs[0][ref0_distance][i_mb_xy][1];
95 h->mc.mc_luma( p+x, i_stride, ref->lowres, i_stride,
96 mvx+(x<<2), mvy+(y<<2), 8, 8, x264_weight_none );
97 }
98 x264_emms();
99 return dest;
100 }
101 x264_emms();
102 return ref->lowres[0];
103 }
104
105 /* How data is organized for 4:2:0/4:2:2 chroma weightp:
106 * [U: ref] [U: fenc]
107 * [V: ref] [V: fenc]
108 * fenc = ref + offset
109 * v = u + stride * chroma height */
110
weight_cost_init_chroma(x264_t * h,x264_frame_t * fenc,x264_frame_t * ref,pixel * dstu,pixel * dstv)111 static NOINLINE void weight_cost_init_chroma( x264_t *h, x264_frame_t *fenc, x264_frame_t *ref, pixel *dstu, pixel *dstv )
112 {
113 int ref0_distance = fenc->i_frame - ref->i_frame - 1;
114 int i_stride = fenc->i_stride[1];
115 int i_lines = fenc->i_lines[1];
116 int i_width = fenc->i_width[1];
117 int v_shift = CHROMA_V_SHIFT;
118 int cw = 8*h->mb.i_mb_width;
119 int ch = 16*h->mb.i_mb_height >> v_shift;
120 int height = 16 >> v_shift;
121
122 if( fenc->lowres_mvs[0][ref0_distance][0][0] != 0x7FFF )
123 {
124 x264_frame_expand_border_chroma( h, ref, 1 );
125 for( int y = 0, mb_xy = 0, pel_offset_y = 0; y < i_lines; y += height, pel_offset_y = y*i_stride )
126 for( int x = 0, pel_offset_x = 0; x < i_width; x += 8, mb_xy++, pel_offset_x += 8 )
127 {
128 pixel *pixu = dstu + pel_offset_y + pel_offset_x;
129 pixel *pixv = dstv + pel_offset_y + pel_offset_x;
130 pixel *src1 = ref->plane[1] + pel_offset_y + pel_offset_x*2; /* NV12/NV16 */
131 int mvx = fenc->lowres_mvs[0][ref0_distance][mb_xy][0];
132 int mvy = fenc->lowres_mvs[0][ref0_distance][mb_xy][1];
133 h->mc.mc_chroma( pixu, pixv, i_stride, src1, i_stride, mvx, 2*mvy>>v_shift, 8, height );
134 }
135 }
136 else
137 h->mc.plane_copy_deinterleave( dstu, i_stride, dstv, i_stride, ref->plane[1], i_stride, cw, ch );
138 h->mc.plane_copy_deinterleave( dstu+i_width, i_stride, dstv+i_width, i_stride, fenc->plane[1], i_stride, cw, ch );
139 x264_emms();
140 }
141
weight_cost_init_chroma444(x264_t * h,x264_frame_t * fenc,x264_frame_t * ref,pixel * dst,int p)142 static NOINLINE pixel *weight_cost_init_chroma444( x264_t *h, x264_frame_t *fenc, x264_frame_t *ref, pixel *dst, int p )
143 {
144 int ref0_distance = fenc->i_frame - ref->i_frame - 1;
145 int i_stride = fenc->i_stride[p];
146 int i_lines = fenc->i_lines[p];
147 int i_width = fenc->i_width[p];
148
149 if( fenc->lowres_mvs[0][ref0_distance][0][0] != 0x7FFF )
150 {
151 x264_frame_expand_border_chroma( h, ref, p );
152 for( int y = 0, mb_xy = 0, pel_offset_y = 0; y < i_lines; y += 16, pel_offset_y = y*i_stride )
153 for( int x = 0, pel_offset_x = 0; x < i_width; x += 16, mb_xy++, pel_offset_x += 16 )
154 {
155 pixel *pix = dst + pel_offset_y + pel_offset_x;
156 pixel *src = ref->plane[p] + pel_offset_y + pel_offset_x;
157 int mvx = fenc->lowres_mvs[0][ref0_distance][mb_xy][0] / 2;
158 int mvy = fenc->lowres_mvs[0][ref0_distance][mb_xy][1] / 2;
159 /* We don't want to calculate hpels for fenc frames, so we round the motion
160 * vectors to fullpel here. It's not too bad, I guess? */
161 h->mc.copy_16x16_unaligned( pix, i_stride, src+mvx+mvy*i_stride, i_stride, 16 );
162 }
163 x264_emms();
164 return dst;
165 }
166 x264_emms();
167 return ref->plane[p];
168 }
169
weight_slice_header_cost(x264_t * h,x264_weight_t * w,int b_chroma)170 static int weight_slice_header_cost( x264_t *h, x264_weight_t *w, int b_chroma )
171 {
172 /* Add cost of weights in the slice header. */
173 int lambda = x264_lambda_tab[X264_LOOKAHEAD_QP];
174 /* 4 times higher, because chroma is analyzed at full resolution. */
175 if( b_chroma )
176 lambda *= 4;
177 int numslices;
178 if( h->param.i_slice_count )
179 numslices = h->param.i_slice_count;
180 else if( h->param.i_slice_max_mbs )
181 numslices = (h->mb.i_mb_width * h->mb.i_mb_height + h->param.i_slice_max_mbs-1) / h->param.i_slice_max_mbs;
182 else
183 numslices = 1;
184 /* FIXME: find a way to account for --slice-max-size?
185 * 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.
186 * Cut denom cost in half if chroma, since it's shared between the two chroma planes. */
187 int denom_cost = bs_size_ue( w[0].i_denom ) * (2 - b_chroma);
188 return lambda * numslices * ( 10 + denom_cost + 2 * (bs_size_se( w[0].i_scale ) + bs_size_se( w[0].i_offset )) );
189 }
190
weight_cost_luma(x264_t * h,x264_frame_t * fenc,pixel * src,x264_weight_t * w)191 static NOINLINE unsigned int weight_cost_luma( x264_t *h, x264_frame_t *fenc, pixel *src, x264_weight_t *w )
192 {
193 unsigned int cost = 0;
194 int i_stride = fenc->i_stride_lowres;
195 int i_lines = fenc->i_lines_lowres;
196 int i_width = fenc->i_width_lowres;
197 pixel *fenc_plane = fenc->lowres[0];
198 ALIGNED_ARRAY_16( pixel, buf,[8*8] );
199 int pixoff = 0;
200 int i_mb = 0;
201
202 if( w )
203 {
204 for( int y = 0; y < i_lines; y += 8, pixoff = y*i_stride )
205 for( int x = 0; x < i_width; x += 8, i_mb++, pixoff += 8)
206 {
207 w->weightfn[8>>2]( buf, 8, &src[pixoff], i_stride, w, 8 );
208 int cmp = h->pixf.mbcmp[PIXEL_8x8]( buf, 8, &fenc_plane[pixoff], i_stride );
209 cost += X264_MIN( cmp, fenc->i_intra_cost[i_mb] );
210 }
211 cost += weight_slice_header_cost( h, w, 0 );
212 }
213 else
214 for( int y = 0; y < i_lines; y += 8, pixoff = y*i_stride )
215 for( int x = 0; x < i_width; x += 8, i_mb++, pixoff += 8 )
216 {
217 int cmp = h->pixf.mbcmp[PIXEL_8x8]( &src[pixoff], i_stride, &fenc_plane[pixoff], i_stride );
218 cost += X264_MIN( cmp, fenc->i_intra_cost[i_mb] );
219 }
220 x264_emms();
221 return cost;
222 }
223
weight_cost_chroma(x264_t * h,x264_frame_t * fenc,pixel * ref,x264_weight_t * w)224 static NOINLINE unsigned int weight_cost_chroma( x264_t *h, x264_frame_t *fenc, pixel *ref, x264_weight_t *w )
225 {
226 unsigned int cost = 0;
227 int i_stride = fenc->i_stride[1];
228 int i_lines = fenc->i_lines[1];
229 int i_width = fenc->i_width[1];
230 pixel *src = ref + i_width;
231 ALIGNED_ARRAY_16( pixel, buf, [8*16] );
232 int pixoff = 0;
233 int height = 16 >> CHROMA_V_SHIFT;
234 if( w )
235 {
236 for( int y = 0; y < i_lines; y += height, pixoff = y*i_stride )
237 for( int x = 0; x < i_width; x += 8, pixoff += 8 )
238 {
239 w->weightfn[8>>2]( buf, 8, &ref[pixoff], i_stride, w, height );
240 /* The naive and seemingly sensible algorithm is to use mbcmp as in luma.
241 * But testing shows that for chroma the DC coefficient is by far the most
242 * important part of the coding cost. Thus a more useful chroma weight is
243 * obtained by comparing each block's DC coefficient instead of the actual
244 * pixels. */
245 cost += h->pixf.asd8( buf, 8, &src[pixoff], i_stride, height );
246 }
247 cost += weight_slice_header_cost( h, w, 1 );
248 }
249 else
250 for( int y = 0; y < i_lines; y += height, pixoff = y*i_stride )
251 for( int x = 0; x < i_width; x += 8, pixoff += 8 )
252 cost += h->pixf.asd8( &ref[pixoff], i_stride, &src[pixoff], i_stride, height );
253 x264_emms();
254 return cost;
255 }
256
weight_cost_chroma444(x264_t * h,x264_frame_t * fenc,pixel * ref,x264_weight_t * w,int p)257 static NOINLINE unsigned int weight_cost_chroma444( x264_t *h, x264_frame_t *fenc, pixel *ref, x264_weight_t *w, int p )
258 {
259 unsigned int cost = 0;
260 int i_stride = fenc->i_stride[p];
261 int i_lines = fenc->i_lines[p];
262 int i_width = fenc->i_width[p];
263 pixel *src = fenc->plane[p];
264 ALIGNED_ARRAY_64( pixel, buf, [16*16] );
265 int pixoff = 0;
266 if( w )
267 {
268 for( int y = 0; y < i_lines; y += 16, pixoff = y*i_stride )
269 for( int x = 0; x < i_width; x += 16, pixoff += 16 )
270 {
271 w->weightfn[16>>2]( buf, 16, &ref[pixoff], i_stride, w, 16 );
272 cost += h->pixf.mbcmp[PIXEL_16x16]( buf, 16, &src[pixoff], i_stride );
273 }
274 cost += weight_slice_header_cost( h, w, 1 );
275 }
276 else
277 for( int y = 0; y < i_lines; y += 16, pixoff = y*i_stride )
278 for( int x = 0; x < i_width; x += 16, pixoff += 16 )
279 cost += h->pixf.mbcmp[PIXEL_16x16]( &ref[pixoff], i_stride, &src[pixoff], i_stride );
280 x264_emms();
281 return cost;
282 }
283
x264_weights_analyse(x264_t * h,x264_frame_t * fenc,x264_frame_t * ref,int b_lookahead)284 void x264_weights_analyse( x264_t *h, x264_frame_t *fenc, x264_frame_t *ref, int b_lookahead )
285 {
286 int i_delta_index = fenc->i_frame - ref->i_frame - 1;
287 /* epsilon is chosen to require at least a numerator of 127 (with denominator = 128) */
288 const float epsilon = 1.f/128.f;
289 x264_weight_t *weights = fenc->weight[0];
290 SET_WEIGHT( weights[0], 0, 1, 0, 0 );
291 SET_WEIGHT( weights[1], 0, 1, 0, 0 );
292 SET_WEIGHT( weights[2], 0, 1, 0, 0 );
293 int chroma_initted = 0;
294 float guess_scale[3];
295 float fenc_mean[3];
296 float ref_mean[3];
297 for( int plane = 0; plane <= 2*!b_lookahead; plane++ )
298 {
299 if( !plane || CHROMA_FORMAT )
300 {
301 int zero_bias = !ref->i_pixel_ssd[plane];
302 float fenc_var = fenc->i_pixel_ssd[plane] + zero_bias;
303 float ref_var = ref->i_pixel_ssd[plane] + zero_bias;
304 guess_scale[plane] = sqrtf( fenc_var / ref_var );
305 fenc_mean[plane] = (float)(fenc->i_pixel_sum[plane] + zero_bias) / (fenc->i_lines[!!plane] * fenc->i_width[!!plane]) / (1 << (BIT_DEPTH - 8));
306 ref_mean[plane] = (float)( ref->i_pixel_sum[plane] + zero_bias) / (fenc->i_lines[!!plane] * fenc->i_width[!!plane]) / (1 << (BIT_DEPTH - 8));
307 }
308 else
309 {
310 guess_scale[plane] = 1;
311 fenc_mean[plane] = 0;
312 ref_mean[plane] = 0;
313 }
314 }
315
316 int chroma_denom = 7;
317 if( !b_lookahead )
318 {
319 /* make sure both our scale factors fit */
320 while( chroma_denom > 0 )
321 {
322 float thresh = 127.f / (1<<chroma_denom);
323 if( guess_scale[1] < thresh && guess_scale[2] < thresh )
324 break;
325 chroma_denom--;
326 }
327 }
328
329 /* Don't check chroma in lookahead, or if there wasn't a luma weight. */
330 for( int plane = 0; plane < (CHROMA_FORMAT ? 3 : 1) && !( plane && ( !weights[0].weightfn || b_lookahead ) ); plane++ )
331 {
332 int minoff, minscale, mindenom;
333 unsigned int minscore, origscore;
334 int found;
335
336 //early termination
337 if( fabsf( ref_mean[plane] - fenc_mean[plane] ) < 0.5f && fabsf( 1.f - guess_scale[plane] ) < epsilon )
338 {
339 SET_WEIGHT( weights[plane], 0, 1, 0, 0 );
340 continue;
341 }
342
343 if( plane )
344 {
345 weights[plane].i_denom = chroma_denom;
346 weights[plane].i_scale = x264_clip3( round( guess_scale[plane] * (1<<chroma_denom) ), 0, 255 );
347 if( weights[plane].i_scale > 127 )
348 {
349 weights[1].weightfn = weights[2].weightfn = NULL;
350 break;
351 }
352 }
353 else
354 weight_get_h264( round( guess_scale[plane] * 128 ), 0, &weights[plane] );
355
356 found = 0;
357 mindenom = weights[plane].i_denom;
358 minscale = weights[plane].i_scale;
359 minoff = 0;
360
361 pixel *mcbuf;
362 if( !plane )
363 {
364 if( !fenc->b_intra_calculated )
365 {
366 x264_mb_analysis_t a;
367 lowres_context_init( h, &a );
368 slicetype_frame_cost( h, &a, &fenc, 0, 0, 0 );
369 }
370 mcbuf = weight_cost_init_luma( h, fenc, ref, h->mb.p_weight_buf[0] );
371 origscore = minscore = weight_cost_luma( h, fenc, mcbuf, NULL );
372 }
373 else
374 {
375 if( CHROMA444 )
376 {
377 mcbuf = weight_cost_init_chroma444( h, fenc, ref, h->mb.p_weight_buf[0], plane );
378 origscore = minscore = weight_cost_chroma444( h, fenc, mcbuf, NULL, plane );
379 }
380 else
381 {
382 pixel *dstu = h->mb.p_weight_buf[0];
383 pixel *dstv = h->mb.p_weight_buf[0]+fenc->i_stride[1]*fenc->i_lines[1];
384 if( !chroma_initted++ )
385 weight_cost_init_chroma( h, fenc, ref, dstu, dstv );
386 mcbuf = plane == 1 ? dstu : dstv;
387 origscore = minscore = weight_cost_chroma( h, fenc, mcbuf, NULL );
388 }
389 }
390
391 if( !minscore )
392 continue;
393
394 /* Picked somewhat arbitrarily */
395 static const uint8_t weight_check_distance[][2] =
396 {
397 {0,0},{0,0},{0,1},{0,1},
398 {0,1},{0,1},{0,1},{1,1},
399 {1,1},{2,1},{2,1},{4,2}
400 };
401 int scale_dist = b_lookahead ? 0 : weight_check_distance[h->param.analyse.i_subpel_refine][0];
402 int offset_dist = b_lookahead ? 0 : weight_check_distance[h->param.analyse.i_subpel_refine][1];
403
404 int start_scale = x264_clip3( minscale - scale_dist, 0, 127 );
405 int end_scale = x264_clip3( minscale + scale_dist, 0, 127 );
406 for( int i_scale = start_scale; i_scale <= end_scale; i_scale++ )
407 {
408 int cur_scale = i_scale;
409 int cur_offset = fenc_mean[plane] - ref_mean[plane] * cur_scale / (1 << mindenom) + 0.5f * b_lookahead;
410 if( cur_offset < - 128 || cur_offset > 127 )
411 {
412 /* Rescale considering the constraints on cur_offset. We do it in this order
413 * because scale has a much wider range than offset (because of denom), so
414 * it should almost never need to be clamped. */
415 cur_offset = x264_clip3( cur_offset, -128, 127 );
416 cur_scale = x264_clip3f( (1 << mindenom) * (fenc_mean[plane] - cur_offset) / ref_mean[plane] + 0.5f, 0, 127 );
417 }
418 int start_offset = x264_clip3( cur_offset - offset_dist, -128, 127 );
419 int end_offset = x264_clip3( cur_offset + offset_dist, -128, 127 );
420 for( int i_off = start_offset; i_off <= end_offset; i_off++ )
421 {
422 SET_WEIGHT( weights[plane], 1, cur_scale, mindenom, i_off );
423 unsigned int s;
424 if( plane )
425 {
426 if( CHROMA444 )
427 s = weight_cost_chroma444( h, fenc, mcbuf, &weights[plane], plane );
428 else
429 s = weight_cost_chroma( h, fenc, mcbuf, &weights[plane] );
430 }
431 else
432 s = weight_cost_luma( h, fenc, mcbuf, &weights[plane] );
433 COPY4_IF_LT( minscore, s, minscale, cur_scale, minoff, i_off, found, 1 );
434
435 // Don't check any more offsets if the previous one had a lower cost than the current one
436 if( minoff == start_offset && i_off != start_offset )
437 break;
438 }
439 }
440 x264_emms();
441
442 /* Use a smaller denominator if possible */
443 if( !plane )
444 {
445 while( mindenom > 0 && !(minscale&1) )
446 {
447 mindenom--;
448 minscale >>= 1;
449 }
450 }
451
452 /* FIXME: More analysis can be done here on SAD vs. SATD termination. */
453 /* 0.2% termination derived experimentally to avoid weird weights in frames that are mostly intra. */
454 if( !found || (minscale == 1 << mindenom && minoff == 0) || (float)minscore / origscore > 0.998f )
455 {
456 SET_WEIGHT( weights[plane], 0, 1, 0, 0 );
457 continue;
458 }
459 else
460 SET_WEIGHT( weights[plane], 1, minscale, mindenom, minoff );
461
462 if( h->param.analyse.i_weighted_pred == X264_WEIGHTP_FAKE && weights[0].weightfn && !plane )
463 fenc->f_weighted_cost_delta[i_delta_index] = (float)minscore / origscore;
464 }
465
466 /* Optimize and unify denominator */
467 if( weights[1].weightfn || weights[2].weightfn )
468 {
469 int denom = weights[1].weightfn ? weights[1].i_denom : weights[2].i_denom;
470 int both_weighted = weights[1].weightfn && weights[2].weightfn;
471 /* If only one plane is weighted, the other has an implicit scale of 1<<denom.
472 * With denom==7, this comes out to 128, which is invalid, so don't allow that. */
473 while( (!both_weighted && denom==7) ||
474 (denom > 0 && !(weights[1].weightfn && (weights[1].i_scale&1))
475 && !(weights[2].weightfn && (weights[2].i_scale&1))) )
476 {
477 denom--;
478 for( int i = 1; i <= 2; i++ )
479 if( weights[i].weightfn )
480 {
481 weights[i].i_scale >>= 1;
482 weights[i].i_denom = denom;
483 }
484 }
485 }
486 for( int i = 1; i <= 2; i++ )
487 if( weights[i].weightfn )
488 h->mc.weight_cache( h, &weights[i] );
489
490 if( weights[0].weightfn && b_lookahead )
491 {
492 //scale lowres in lookahead for slicetype_frame_cost
493 pixel *src = ref->buffer_lowres;
494 pixel *dst = h->mb.p_weight_buf[0];
495 int width = ref->i_width_lowres + PADH2;
496 int height = ref->i_lines_lowres + PADV*2;
497 x264_weight_scale_plane( h, dst, ref->i_stride_lowres, src, ref->i_stride_lowres,
498 width, height, &weights[0] );
499 fenc->weighted[0] = h->mb.p_weight_buf[0] + PADH_ALIGN + ref->i_stride_lowres * PADV;
500 }
501 }
502
503 /* Output buffers are separated by 128 bytes to avoid false sharing of cachelines
504 * in multithreaded lookahead. */
505 #define PAD_SIZE 32
506 /* cost_est, cost_est_aq, intra_mbs, num rows */
507 #define NUM_INTS 4
508 #define COST_EST 0
509 #define COST_EST_AQ 1
510 #define INTRA_MBS 2
511 #define NUM_ROWS 3
512 #define ROW_SATD (NUM_INTS + (h->mb.i_mb_y - h->i_threadslice_start))
513
slicetype_mb_cost(x264_t * h,x264_mb_analysis_t * a,x264_frame_t ** frames,int p0,int p1,int b,int dist_scale_factor,int do_search[2],const x264_weight_t * w,int * output_inter,int * output_intra)514 static void slicetype_mb_cost( x264_t *h, x264_mb_analysis_t *a,
515 x264_frame_t **frames, int p0, int p1, int b,
516 int dist_scale_factor, int do_search[2], const x264_weight_t *w,
517 int *output_inter, int *output_intra )
518 {
519 x264_frame_t *fref0 = frames[p0];
520 x264_frame_t *fref1 = frames[p1];
521 x264_frame_t *fenc = frames[b];
522 const int b_bidir = (b < p1);
523 const int i_mb_x = h->mb.i_mb_x;
524 const int i_mb_y = h->mb.i_mb_y;
525 const int i_mb_stride = h->mb.i_mb_width;
526 const int i_mb_xy = i_mb_x + i_mb_y * i_mb_stride;
527 const int i_stride = fenc->i_stride_lowres;
528 const int i_pel_offset = 8 * (i_mb_x + i_mb_y * i_stride);
529 const int i_bipred_weight = h->param.analyse.b_weighted_bipred ? 64 - (dist_scale_factor>>2) : 32;
530 int16_t (*fenc_mvs[2])[2] = { b != p0 ? &fenc->lowres_mvs[0][b-p0-1][i_mb_xy] : NULL, b != p1 ? &fenc->lowres_mvs[1][p1-b-1][i_mb_xy] : NULL };
531 int (*fenc_costs[2]) = { b != p0 ? &fenc->lowres_mv_costs[0][b-p0-1][i_mb_xy] : NULL, b != p1 ? &fenc->lowres_mv_costs[1][p1-b-1][i_mb_xy] : NULL };
532 int b_frame_score_mb = (i_mb_x > 0 && i_mb_x < h->mb.i_mb_width - 1 &&
533 i_mb_y > 0 && i_mb_y < h->mb.i_mb_height - 1) ||
534 h->mb.i_mb_width <= 2 || h->mb.i_mb_height <= 2;
535
536 ALIGNED_ARRAY_16( pixel, pix1,[9*FDEC_STRIDE] );
537 pixel *pix2 = pix1+8;
538 x264_me_t m[2];
539 int i_bcost = COST_MAX;
540 int list_used = 0;
541 /* A small, arbitrary bias to avoid VBV problems caused by zero-residual lookahead blocks. */
542 int lowres_penalty = 4;
543
544 h->mb.pic.p_fenc[0] = h->mb.pic.fenc_buf;
545 h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fenc[0], FENC_STRIDE, &fenc->lowres[0][i_pel_offset], i_stride, 8 );
546
547 if( p0 == p1 )
548 goto lowres_intra_mb;
549
550 int mv_range = 2 * h->param.analyse.i_mv_range;
551 // no need for h->mb.mv_min[]
552 h->mb.mv_min_spel[0] = X264_MAX( 4*(-8*h->mb.i_mb_x - 12), -mv_range );
553 h->mb.mv_max_spel[0] = X264_MIN( 4*(8*(h->mb.i_mb_width - h->mb.i_mb_x - 1) + 12), mv_range-1 );
554 h->mb.mv_limit_fpel[0][0] = h->mb.mv_min_spel[0] >> 2;
555 h->mb.mv_limit_fpel[1][0] = h->mb.mv_max_spel[0] >> 2;
556 if( h->mb.i_mb_x >= h->mb.i_mb_width - 2 )
557 {
558 h->mb.mv_min_spel[1] = X264_MAX( 4*(-8*h->mb.i_mb_y - 12), -mv_range );
559 h->mb.mv_max_spel[1] = X264_MIN( 4*(8*( h->mb.i_mb_height - h->mb.i_mb_y - 1) + 12), mv_range-1 );
560 h->mb.mv_limit_fpel[0][1] = h->mb.mv_min_spel[1] >> 2;
561 h->mb.mv_limit_fpel[1][1] = h->mb.mv_max_spel[1] >> 2;
562 }
563
564 #define LOAD_HPELS_LUMA(dst, src) \
565 { \
566 (dst)[0] = &(src)[0][i_pel_offset]; \
567 (dst)[1] = &(src)[1][i_pel_offset]; \
568 (dst)[2] = &(src)[2][i_pel_offset]; \
569 (dst)[3] = &(src)[3][i_pel_offset]; \
570 }
571 #define LOAD_WPELS_LUMA(dst,src) \
572 (dst) = &(src)[i_pel_offset];
573
574 #define CLIP_MV( mv ) \
575 { \
576 mv[0] = x264_clip3( mv[0], h->mb.mv_min_spel[0], h->mb.mv_max_spel[0] ); \
577 mv[1] = x264_clip3( mv[1], h->mb.mv_min_spel[1], h->mb.mv_max_spel[1] ); \
578 }
579 #define TRY_BIDIR( mv0, mv1, penalty ) \
580 { \
581 int i_cost; \
582 if( h->param.analyse.i_subpel_refine <= 1 ) \
583 { \
584 int hpel_idx1 = (((mv0)[0]&2)>>1) + ((mv0)[1]&2); \
585 int hpel_idx2 = (((mv1)[0]&2)>>1) + ((mv1)[1]&2); \
586 pixel *src1 = m[0].p_fref[hpel_idx1] + ((mv0)[0]>>2) + ((mv0)[1]>>2) * m[0].i_stride[0]; \
587 pixel *src2 = m[1].p_fref[hpel_idx2] + ((mv1)[0]>>2) + ((mv1)[1]>>2) * m[1].i_stride[0]; \
588 h->mc.avg[PIXEL_8x8]( pix1, 16, src1, m[0].i_stride[0], src2, m[1].i_stride[0], i_bipred_weight ); \
589 } \
590 else \
591 { \
592 intptr_t stride1 = 16, stride2 = 16; \
593 pixel *src1, *src2; \
594 src1 = h->mc.get_ref( pix1, &stride1, m[0].p_fref, m[0].i_stride[0], \
595 (mv0)[0], (mv0)[1], 8, 8, w ); \
596 src2 = h->mc.get_ref( pix2, &stride2, m[1].p_fref, m[1].i_stride[0], \
597 (mv1)[0], (mv1)[1], 8, 8, w ); \
598 h->mc.avg[PIXEL_8x8]( pix1, 16, src1, stride1, src2, stride2, i_bipred_weight ); \
599 } \
600 i_cost = penalty * a->i_lambda + h->pixf.mbcmp[PIXEL_8x8]( \
601 m[0].p_fenc[0], FENC_STRIDE, pix1, 16 ); \
602 COPY2_IF_LT( i_bcost, i_cost, list_used, 3 ); \
603 }
604
605 m[0].i_pixel = PIXEL_8x8;
606 m[0].p_cost_mv = a->p_cost_mv;
607 m[0].i_stride[0] = i_stride;
608 m[0].p_fenc[0] = h->mb.pic.p_fenc[0];
609 m[0].weight = w;
610 m[0].i_ref = 0;
611 LOAD_HPELS_LUMA( m[0].p_fref, fref0->lowres );
612 m[0].p_fref_w = m[0].p_fref[0];
613 if( w[0].weightfn )
614 LOAD_WPELS_LUMA( m[0].p_fref_w, fenc->weighted[0] );
615
616 if( b_bidir )
617 {
618 ALIGNED_ARRAY_8( int16_t, dmv,[2],[2] );
619
620 m[1].i_pixel = PIXEL_8x8;
621 m[1].p_cost_mv = a->p_cost_mv;
622 m[1].i_stride[0] = i_stride;
623 m[1].p_fenc[0] = h->mb.pic.p_fenc[0];
624 m[1].i_ref = 0;
625 m[1].weight = x264_weight_none;
626 LOAD_HPELS_LUMA( m[1].p_fref, fref1->lowres );
627 m[1].p_fref_w = m[1].p_fref[0];
628
629 if( fref1->lowres_mvs[0][p1-p0-1][0][0] != 0x7FFF )
630 {
631 int16_t *mvr = fref1->lowres_mvs[0][p1-p0-1][i_mb_xy];
632 dmv[0][0] = ( mvr[0] * dist_scale_factor + 128 ) >> 8;
633 dmv[0][1] = ( mvr[1] * dist_scale_factor + 128 ) >> 8;
634 dmv[1][0] = dmv[0][0] - mvr[0];
635 dmv[1][1] = dmv[0][1] - mvr[1];
636 CLIP_MV( dmv[0] );
637 CLIP_MV( dmv[1] );
638 if( h->param.analyse.i_subpel_refine <= 1 )
639 M64( dmv ) &= ~0x0001000100010001ULL; /* mv & ~1 */
640 }
641 else
642 M64( dmv ) = 0;
643
644 TRY_BIDIR( dmv[0], dmv[1], 0 );
645 if( M64( dmv ) )
646 {
647 int i_cost;
648 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 );
649 i_cost = h->pixf.mbcmp[PIXEL_8x8]( m[0].p_fenc[0], FENC_STRIDE, pix1, 16 );
650 COPY2_IF_LT( i_bcost, i_cost, list_used, 3 );
651 }
652 }
653
654 for( int l = 0; l < 1 + b_bidir; l++ )
655 {
656 if( do_search[l] )
657 {
658 int i_mvc = 0;
659 int16_t (*fenc_mv)[2] = fenc_mvs[l];
660 ALIGNED_ARRAY_8( int16_t, mvc,[4],[2] );
661
662 /* Reverse-order MV prediction. */
663 M32( mvc[0] ) = 0;
664 M32( mvc[2] ) = 0;
665 #define MVC(mv) { CP32( mvc[i_mvc], mv ); i_mvc++; }
666 if( i_mb_x < h->mb.i_mb_width - 1 )
667 MVC( fenc_mv[1] );
668 if( i_mb_y < h->i_threadslice_end - 1 )
669 {
670 MVC( fenc_mv[i_mb_stride] );
671 if( i_mb_x > 0 )
672 MVC( fenc_mv[i_mb_stride-1] );
673 if( i_mb_x < h->mb.i_mb_width - 1 )
674 MVC( fenc_mv[i_mb_stride+1] );
675 }
676 #undef MVC
677 if( i_mvc <= 1 )
678 CP32( m[l].mvp, mvc[0] );
679 else
680 x264_median_mv( m[l].mvp, mvc[0], mvc[1], mvc[2] );
681
682 /* Fast skip for cases of near-zero residual. Shortcut: don't bother except in the mv0 case,
683 * since anything else is likely to have enough residual to not trigger the skip. */
684 if( !M32( m[l].mvp ) )
685 {
686 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] );
687 if( m[l].cost < 64 )
688 {
689 M32( m[l].mv ) = 0;
690 goto skip_motionest;
691 }
692 }
693
694 x264_me_search( h, &m[l], mvc, i_mvc );
695 m[l].cost -= a->p_cost_mv[0]; // remove mvcost from skip mbs
696 if( M32( m[l].mv ) )
697 m[l].cost += 5 * a->i_lambda;
698
699 skip_motionest:
700 CP32( fenc_mvs[l], m[l].mv );
701 *fenc_costs[l] = m[l].cost;
702 }
703 else
704 {
705 CP32( m[l].mv, fenc_mvs[l] );
706 m[l].cost = *fenc_costs[l];
707 }
708 COPY2_IF_LT( i_bcost, m[l].cost, list_used, l+1 );
709 }
710
711 if( b_bidir && ( M32( m[0].mv ) || M32( m[1].mv ) ) )
712 TRY_BIDIR( m[0].mv, m[1].mv, 5 );
713
714 lowres_intra_mb:
715 if( !fenc->b_intra_calculated )
716 {
717 ALIGNED_ARRAY_16( pixel, edge,[36] );
718 pixel *pix = &pix1[8+FDEC_STRIDE];
719 pixel *src = &fenc->lowres[0][i_pel_offset];
720 const int intra_penalty = 5 * a->i_lambda;
721 int satds[3];
722 int pixoff = 4 / SIZEOF_PIXEL;
723
724 /* Avoid store forwarding stalls by writing larger chunks */
725 memcpy( pix-FDEC_STRIDE, src-i_stride, 16 * SIZEOF_PIXEL );
726 for( int i = -1; i < 8; i++ )
727 M32( &pix[i*FDEC_STRIDE-pixoff] ) = M32( &src[i*i_stride-pixoff] );
728
729 h->pixf.intra_mbcmp_x3_8x8c( h->mb.pic.p_fenc[0], pix, satds );
730 int i_icost = X264_MIN3( satds[0], satds[1], satds[2] );
731
732 if( h->param.analyse.i_subpel_refine > 1 )
733 {
734 h->predict_8x8c[I_PRED_CHROMA_P]( pix );
735 int satd = h->pixf.mbcmp[PIXEL_8x8]( h->mb.pic.p_fenc[0], FENC_STRIDE, pix, FDEC_STRIDE );
736 i_icost = X264_MIN( i_icost, satd );
737 h->predict_8x8_filter( pix, edge, ALL_NEIGHBORS, ALL_NEIGHBORS );
738 for( int i = 3; i < 9; i++ )
739 {
740 h->predict_8x8[i]( pix, edge );
741 satd = h->pixf.mbcmp[PIXEL_8x8]( h->mb.pic.p_fenc[0], FENC_STRIDE, pix, FDEC_STRIDE );
742 i_icost = X264_MIN( i_icost, satd );
743 }
744 }
745
746 i_icost = ((i_icost + intra_penalty) >> (BIT_DEPTH - 8)) + lowres_penalty;
747 fenc->i_intra_cost[i_mb_xy] = i_icost;
748 int i_icost_aq = i_icost;
749 if( h->param.rc.i_aq_mode )
750 i_icost_aq = (i_icost_aq * fenc->i_inv_qscale_factor[i_mb_xy] + 128) >> 8;
751 output_intra[ROW_SATD] += i_icost_aq;
752 if( b_frame_score_mb )
753 {
754 output_intra[COST_EST] += i_icost;
755 output_intra[COST_EST_AQ] += i_icost_aq;
756 }
757 }
758 i_bcost = (i_bcost >> (BIT_DEPTH - 8)) + lowres_penalty;
759
760 /* forbid intra-mbs in B-frames, because it's rare and not worth checking */
761 /* FIXME: Should we still forbid them now that we cache intra scores? */
762 if( !b_bidir )
763 {
764 int i_icost = fenc->i_intra_cost[i_mb_xy];
765 int b_intra = i_icost < i_bcost;
766 if( b_intra )
767 {
768 i_bcost = i_icost;
769 list_used = 0;
770 }
771 if( b_frame_score_mb )
772 output_inter[INTRA_MBS] += b_intra;
773 }
774
775 /* In an I-frame, we've already added the results above in the intra section. */
776 if( p0 != p1 )
777 {
778 int i_bcost_aq = i_bcost;
779 if( h->param.rc.i_aq_mode )
780 i_bcost_aq = (i_bcost_aq * fenc->i_inv_qscale_factor[i_mb_xy] + 128) >> 8;
781 output_inter[ROW_SATD] += i_bcost_aq;
782 if( b_frame_score_mb )
783 {
784 /* Don't use AQ-weighted costs for slicetype decision, only for ratecontrol. */
785 output_inter[COST_EST] += i_bcost;
786 output_inter[COST_EST_AQ] += i_bcost_aq;
787 }
788 }
789
790 fenc->lowres_costs[b-p0][p1-b][i_mb_xy] = X264_MIN( i_bcost, LOWRES_COST_MASK ) + (list_used << LOWRES_COST_SHIFT);
791 }
792 #undef TRY_BIDIR
793
794 #define NUM_MBS\
795 (h->mb.i_mb_width > 2 && h->mb.i_mb_height > 2 ?\
796 (h->mb.i_mb_width - 2) * (h->mb.i_mb_height - 2) :\
797 h->mb.i_mb_width * h->mb.i_mb_height)
798
799 typedef struct
800 {
801 x264_t *h;
802 x264_mb_analysis_t *a;
803 x264_frame_t **frames;
804 int p0;
805 int p1;
806 int b;
807 int dist_scale_factor;
808 int *do_search;
809 const x264_weight_t *w;
810 int *output_inter;
811 int *output_intra;
812 } x264_slicetype_slice_t;
813
slicetype_slice_cost(x264_slicetype_slice_t * s)814 static void slicetype_slice_cost( x264_slicetype_slice_t *s )
815 {
816 x264_t *h = s->h;
817
818 /* Lowres lookahead goes backwards because the MVs are used as predictors in the main encode.
819 * This considerably improves MV prediction overall. */
820
821 /* The edge mbs seem to reduce the predictive quality of the
822 * whole frame's score, but are needed for a spatial distribution. */
823 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;
824
825 int start_y = X264_MIN( h->i_threadslice_end - 1, h->mb.i_mb_height - 2 + do_edges );
826 int end_y = X264_MAX( h->i_threadslice_start, 1 - do_edges );
827 int start_x = h->mb.i_mb_width - 2 + do_edges;
828 int end_x = 1 - do_edges;
829
830 for( h->mb.i_mb_y = start_y; h->mb.i_mb_y >= end_y; h->mb.i_mb_y-- )
831 for( h->mb.i_mb_x = start_x; h->mb.i_mb_x >= end_x; h->mb.i_mb_x-- )
832 slicetype_mb_cost( h, s->a, s->frames, s->p0, s->p1, s->b, s->dist_scale_factor,
833 s->do_search, s->w, s->output_inter, s->output_intra );
834 }
835
slicetype_frame_cost(x264_t * h,x264_mb_analysis_t * a,x264_frame_t ** frames,int p0,int p1,int b)836 static int slicetype_frame_cost( x264_t *h, x264_mb_analysis_t *a,
837 x264_frame_t **frames, int p0, int p1, int b )
838 {
839 int i_score = 0;
840 int do_search[2];
841 const x264_weight_t *w = x264_weight_none;
842 x264_frame_t *fenc = frames[b];
843
844 /* Check whether we already evaluated this frame
845 * If we have tried this frame as P, then we have also tried
846 * the preceding frames as B. (is this still true?) */
847 /* Also check that we already calculated the row SATDs for the current frame. */
848 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) )
849 i_score = fenc->i_cost_est[b-p0][p1-b];
850 else
851 {
852 int dist_scale_factor = 128;
853
854 /* For each list, check to see whether we have lowres motion-searched this reference frame before. */
855 do_search[0] = b != p0 && fenc->lowres_mvs[0][b-p0-1][0][0] == 0x7FFF;
856 do_search[1] = b != p1 && fenc->lowres_mvs[1][p1-b-1][0][0] == 0x7FFF;
857 if( do_search[0] )
858 {
859 if( h->param.analyse.i_weighted_pred && b == p1 )
860 {
861 x264_emms();
862 x264_weights_analyse( h, fenc, frames[p0], 1 );
863 w = fenc->weight[0];
864 }
865 fenc->lowres_mvs[0][b-p0-1][0][0] = 0;
866 }
867 if( do_search[1] ) fenc->lowres_mvs[1][p1-b-1][0][0] = 0;
868
869 if( p1 != p0 )
870 dist_scale_factor = ( ((b-p0) << 8) + ((p1-p0) >> 1) ) / (p1-p0);
871
872 int output_buf_size = h->mb.i_mb_height + (NUM_INTS + PAD_SIZE) * h->param.i_lookahead_threads;
873 int *output_inter[X264_LOOKAHEAD_THREAD_MAX+1];
874 int *output_intra[X264_LOOKAHEAD_THREAD_MAX+1];
875 output_inter[0] = h->scratch_buffer2;
876 output_intra[0] = output_inter[0] + output_buf_size;
877
878 #if HAVE_OPENCL
879 if( h->param.b_opencl )
880 {
881 x264_opencl_lowres_init(h, fenc, a->i_lambda );
882 if( do_search[0] )
883 {
884 x264_opencl_lowres_init( h, frames[p0], a->i_lambda );
885 x264_opencl_motionsearch( h, frames, b, p0, 0, a->i_lambda, w );
886 }
887 if( do_search[1] )
888 {
889 x264_opencl_lowres_init( h, frames[p1], a->i_lambda );
890 x264_opencl_motionsearch( h, frames, b, p1, 1, a->i_lambda, NULL );
891 }
892 if( b != p0 )
893 x264_opencl_finalize_cost( h, a->i_lambda, frames, p0, p1, b, dist_scale_factor );
894 x264_opencl_flush( h );
895
896 i_score = fenc->i_cost_est[b-p0][p1-b];
897 }
898 else
899 #endif
900 {
901 if( h->param.i_lookahead_threads > 1 )
902 {
903 x264_slicetype_slice_t s[X264_LOOKAHEAD_THREAD_MAX];
904
905 for( int i = 0; i < h->param.i_lookahead_threads; i++ )
906 {
907 x264_t *t = h->lookahead_thread[i];
908
909 /* FIXME move this somewhere else */
910 t->mb.i_me_method = h->mb.i_me_method;
911 t->mb.i_subpel_refine = h->mb.i_subpel_refine;
912 t->mb.b_chroma_me = h->mb.b_chroma_me;
913
914 s[i] = (x264_slicetype_slice_t){ t, a, frames, p0, p1, b, dist_scale_factor, do_search, w,
915 output_inter[i], output_intra[i] };
916
917 t->i_threadslice_start = ((h->mb.i_mb_height * i + h->param.i_lookahead_threads/2) / h->param.i_lookahead_threads);
918 t->i_threadslice_end = ((h->mb.i_mb_height * (i+1) + h->param.i_lookahead_threads/2) / h->param.i_lookahead_threads);
919
920 int thread_height = t->i_threadslice_end - t->i_threadslice_start;
921 int thread_output_size = thread_height + NUM_INTS;
922 memset( output_inter[i], 0, thread_output_size * sizeof(int) );
923 memset( output_intra[i], 0, thread_output_size * sizeof(int) );
924 output_inter[i][NUM_ROWS] = output_intra[i][NUM_ROWS] = thread_height;
925
926 output_inter[i+1] = output_inter[i] + thread_output_size + PAD_SIZE;
927 output_intra[i+1] = output_intra[i] + thread_output_size + PAD_SIZE;
928
929 x264_threadpool_run( h->lookaheadpool, (void*)slicetype_slice_cost, &s[i] );
930 }
931 for( int i = 0; i < h->param.i_lookahead_threads; i++ )
932 x264_threadpool_wait( h->lookaheadpool, &s[i] );
933 }
934 else
935 {
936 h->i_threadslice_start = 0;
937 h->i_threadslice_end = h->mb.i_mb_height;
938 memset( output_inter[0], 0, (output_buf_size - PAD_SIZE) * sizeof(int) );
939 memset( output_intra[0], 0, (output_buf_size - PAD_SIZE) * sizeof(int) );
940 output_inter[0][NUM_ROWS] = output_intra[0][NUM_ROWS] = h->mb.i_mb_height;
941 x264_slicetype_slice_t s = (x264_slicetype_slice_t){ h, a, frames, p0, p1, b, dist_scale_factor, do_search, w,
942 output_inter[0], output_intra[0] };
943 slicetype_slice_cost( &s );
944 }
945
946 /* Sum up accumulators */
947 if( b == p1 )
948 fenc->i_intra_mbs[b-p0] = 0;
949 if( !fenc->b_intra_calculated )
950 {
951 fenc->i_cost_est[0][0] = 0;
952 fenc->i_cost_est_aq[0][0] = 0;
953 }
954 fenc->i_cost_est[b-p0][p1-b] = 0;
955 fenc->i_cost_est_aq[b-p0][p1-b] = 0;
956
957 int *row_satd_inter = fenc->i_row_satds[b-p0][p1-b];
958 int *row_satd_intra = fenc->i_row_satds[0][0];
959 for( int i = 0; i < h->param.i_lookahead_threads; i++ )
960 {
961 if( b == p1 )
962 fenc->i_intra_mbs[b-p0] += output_inter[i][INTRA_MBS];
963 if( !fenc->b_intra_calculated )
964 {
965 fenc->i_cost_est[0][0] += output_intra[i][COST_EST];
966 fenc->i_cost_est_aq[0][0] += output_intra[i][COST_EST_AQ];
967 }
968
969 fenc->i_cost_est[b-p0][p1-b] += output_inter[i][COST_EST];
970 fenc->i_cost_est_aq[b-p0][p1-b] += output_inter[i][COST_EST_AQ];
971
972 if( h->param.rc.i_vbv_buffer_size )
973 {
974 int row_count = output_inter[i][NUM_ROWS];
975 memcpy( row_satd_inter, output_inter[i] + NUM_INTS, row_count * sizeof(int) );
976 if( !fenc->b_intra_calculated )
977 memcpy( row_satd_intra, output_intra[i] + NUM_INTS, row_count * sizeof(int) );
978 row_satd_inter += row_count;
979 row_satd_intra += row_count;
980 }
981 }
982
983 i_score = fenc->i_cost_est[b-p0][p1-b];
984 if( b != p1 )
985 i_score = (uint64_t)i_score * 100 / (120 + h->param.i_bframe_bias);
986 else
987 fenc->b_intra_calculated = 1;
988
989 fenc->i_cost_est[b-p0][p1-b] = i_score;
990 x264_emms();
991 }
992 }
993
994 return i_score;
995 }
996
997 /* If MB-tree changes the quantizers, we need to recalculate the frame cost without
998 * re-running lookahead. */
slicetype_frame_cost_recalculate(x264_t * h,x264_frame_t ** frames,int p0,int p1,int b)999 static int slicetype_frame_cost_recalculate( x264_t *h, x264_frame_t **frames, int p0, int p1, int b )
1000 {
1001 int i_score = 0;
1002 int *row_satd = frames[b]->i_row_satds[b-p0][p1-b];
1003 float *qp_offset = IS_X264_TYPE_B(frames[b]->i_type) ? frames[b]->f_qp_offset_aq : frames[b]->f_qp_offset;
1004 x264_emms();
1005 for( h->mb.i_mb_y = h->mb.i_mb_height - 1; h->mb.i_mb_y >= 0; h->mb.i_mb_y-- )
1006 {
1007 row_satd[ h->mb.i_mb_y ] = 0;
1008 for( h->mb.i_mb_x = h->mb.i_mb_width - 1; h->mb.i_mb_x >= 0; h->mb.i_mb_x-- )
1009 {
1010 int i_mb_xy = h->mb.i_mb_x + h->mb.i_mb_y*h->mb.i_mb_stride;
1011 int i_mb_cost = frames[b]->lowres_costs[b-p0][p1-b][i_mb_xy] & LOWRES_COST_MASK;
1012 float qp_adj = qp_offset[i_mb_xy];
1013 i_mb_cost = (i_mb_cost * x264_exp2fix8(qp_adj) + 128) >> 8;
1014 row_satd[ h->mb.i_mb_y ] += i_mb_cost;
1015 if( (h->mb.i_mb_y > 0 && h->mb.i_mb_y < h->mb.i_mb_height - 1 &&
1016 h->mb.i_mb_x > 0 && h->mb.i_mb_x < h->mb.i_mb_width - 1) ||
1017 h->mb.i_mb_width <= 2 || h->mb.i_mb_height <= 2 )
1018 {
1019 i_score += i_mb_cost;
1020 }
1021 }
1022 }
1023 return i_score;
1024 }
1025
1026 /* Trade off precision in mbtree for increased range */
1027 #define MBTREE_PRECISION 0.5f
1028
macroblock_tree_finish(x264_t * h,x264_frame_t * frame,float average_duration,int ref0_distance)1029 static void macroblock_tree_finish( x264_t *h, x264_frame_t *frame, float average_duration, int ref0_distance )
1030 {
1031 int fps_factor = round( CLIP_DURATION(average_duration) / CLIP_DURATION(frame->f_duration) * 256 / MBTREE_PRECISION );
1032 float weightdelta = 0.0;
1033 if( ref0_distance && frame->f_weighted_cost_delta[ref0_distance-1] > 0 )
1034 weightdelta = (1.0 - frame->f_weighted_cost_delta[ref0_distance-1]);
1035
1036 /* Allow the strength to be adjusted via qcompress, since the two
1037 * concepts are very similar. */
1038 float strength = 5.0f * (1.0f - h->param.rc.f_qcompress);
1039 for( int mb_index = 0; mb_index < h->mb.i_mb_count; mb_index++ )
1040 {
1041 int intra_cost = (frame->i_intra_cost[mb_index] * frame->i_inv_qscale_factor[mb_index] + 128) >> 8;
1042 if( intra_cost )
1043 {
1044 int propagate_cost = (frame->i_propagate_cost[mb_index] * fps_factor + 128) >> 8;
1045 float log2_ratio = x264_log2(intra_cost + propagate_cost) - x264_log2(intra_cost) + weightdelta;
1046 frame->f_qp_offset[mb_index] = frame->f_qp_offset_aq[mb_index] - strength * log2_ratio;
1047 }
1048 }
1049 }
1050
macroblock_tree_propagate(x264_t * h,x264_frame_t ** frames,float average_duration,int p0,int p1,int b,int referenced)1051 static void macroblock_tree_propagate( x264_t *h, x264_frame_t **frames, float average_duration, int p0, int p1, int b, int referenced )
1052 {
1053 uint16_t *ref_costs[2] = {frames[p0]->i_propagate_cost,frames[p1]->i_propagate_cost};
1054 int dist_scale_factor = ( ((b-p0) << 8) + ((p1-p0) >> 1) ) / (p1-p0);
1055 int i_bipred_weight = h->param.analyse.b_weighted_bipred ? 64 - (dist_scale_factor>>2) : 32;
1056 int16_t (*mvs[2])[2] = { b != p0 ? frames[b]->lowres_mvs[0][b-p0-1] : NULL, b != p1 ? frames[b]->lowres_mvs[1][p1-b-1] : NULL };
1057 int bipred_weights[2] = {i_bipred_weight, 64 - i_bipred_weight};
1058 int16_t *buf = h->scratch_buffer;
1059 uint16_t *propagate_cost = frames[b]->i_propagate_cost;
1060 uint16_t *lowres_costs = frames[b]->lowres_costs[b-p0][p1-b];
1061
1062 x264_emms();
1063 float fps_factor = CLIP_DURATION(frames[b]->f_duration) / (CLIP_DURATION(average_duration) * 256.0f) * MBTREE_PRECISION;
1064
1065 /* For non-reffed frames the source costs are always zero, so just memset one row and re-use it. */
1066 if( !referenced )
1067 memset( frames[b]->i_propagate_cost, 0, h->mb.i_mb_width * sizeof(uint16_t) );
1068
1069 for( h->mb.i_mb_y = 0; h->mb.i_mb_y < h->mb.i_mb_height; h->mb.i_mb_y++ )
1070 {
1071 int mb_index = h->mb.i_mb_y*h->mb.i_mb_stride;
1072 h->mc.mbtree_propagate_cost( buf, propagate_cost,
1073 frames[b]->i_intra_cost+mb_index, lowres_costs+mb_index,
1074 frames[b]->i_inv_qscale_factor+mb_index, &fps_factor, h->mb.i_mb_width );
1075 if( referenced )
1076 propagate_cost += h->mb.i_mb_width;
1077
1078 h->mc.mbtree_propagate_list( h, ref_costs[0], &mvs[0][mb_index], buf, &lowres_costs[mb_index],
1079 bipred_weights[0], h->mb.i_mb_y, h->mb.i_mb_width, 0 );
1080 if( b != p1 )
1081 {
1082 h->mc.mbtree_propagate_list( h, ref_costs[1], &mvs[1][mb_index], buf, &lowres_costs[mb_index],
1083 bipred_weights[1], h->mb.i_mb_y, h->mb.i_mb_width, 1 );
1084 }
1085 }
1086
1087 if( h->param.rc.i_vbv_buffer_size && h->param.rc.i_lookahead && referenced )
1088 macroblock_tree_finish( h, frames[b], average_duration, b == p1 ? b - p0 : 0 );
1089 }
1090
macroblock_tree(x264_t * h,x264_mb_analysis_t * a,x264_frame_t ** frames,int num_frames,int b_intra)1091 static void macroblock_tree( x264_t *h, x264_mb_analysis_t *a, x264_frame_t **frames, int num_frames, int b_intra )
1092 {
1093 int idx = !b_intra;
1094 int last_nonb, cur_nonb = 1;
1095 int bframes = 0;
1096
1097 x264_emms();
1098 float total_duration = 0.0;
1099 for( int j = 0; j <= num_frames; j++ )
1100 total_duration += frames[j]->f_duration;
1101 float average_duration = total_duration / (num_frames + 1);
1102
1103 int i = num_frames;
1104
1105 if( b_intra )
1106 slicetype_frame_cost( h, a, frames, 0, 0, 0 );
1107
1108 while( i > 0 && IS_X264_TYPE_B( frames[i]->i_type ) )
1109 i--;
1110 last_nonb = i;
1111
1112 /* Lookaheadless MB-tree is not a theoretically distinct case; the same extrapolation could
1113 * be applied to the end of a lookahead buffer of any size. However, it's most needed when
1114 * lookahead=0, so that's what's currently implemented. */
1115 if( !h->param.rc.i_lookahead )
1116 {
1117 if( b_intra )
1118 {
1119 memset( frames[0]->i_propagate_cost, 0, h->mb.i_mb_count * sizeof(uint16_t) );
1120 memcpy( frames[0]->f_qp_offset, frames[0]->f_qp_offset_aq, h->mb.i_mb_count * sizeof(float) );
1121 return;
1122 }
1123 XCHG( uint16_t*, frames[last_nonb]->i_propagate_cost, frames[0]->i_propagate_cost );
1124 memset( frames[0]->i_propagate_cost, 0, h->mb.i_mb_count * sizeof(uint16_t) );
1125 }
1126 else
1127 {
1128 if( last_nonb < idx )
1129 return;
1130 memset( frames[last_nonb]->i_propagate_cost, 0, h->mb.i_mb_count * sizeof(uint16_t) );
1131 }
1132
1133 while( i-- > idx )
1134 {
1135 cur_nonb = i;
1136 while( IS_X264_TYPE_B( frames[cur_nonb]->i_type ) && cur_nonb > 0 )
1137 cur_nonb--;
1138 if( cur_nonb < idx )
1139 break;
1140 slicetype_frame_cost( h, a, frames, cur_nonb, last_nonb, last_nonb );
1141 memset( frames[cur_nonb]->i_propagate_cost, 0, h->mb.i_mb_count * sizeof(uint16_t) );
1142 bframes = last_nonb - cur_nonb - 1;
1143 if( h->param.i_bframe_pyramid && bframes > 1 )
1144 {
1145 int middle = (bframes + 1)/2 + cur_nonb;
1146 slicetype_frame_cost( h, a, frames, cur_nonb, last_nonb, middle );
1147 memset( frames[middle]->i_propagate_cost, 0, h->mb.i_mb_count * sizeof(uint16_t) );
1148 while( i > cur_nonb )
1149 {
1150 int p0 = i > middle ? middle : cur_nonb;
1151 int p1 = i < middle ? middle : last_nonb;
1152 if( i != middle )
1153 {
1154 slicetype_frame_cost( h, a, frames, p0, p1, i );
1155 macroblock_tree_propagate( h, frames, average_duration, p0, p1, i, 0 );
1156 }
1157 i--;
1158 }
1159 macroblock_tree_propagate( h, frames, average_duration, cur_nonb, last_nonb, middle, 1 );
1160 }
1161 else
1162 {
1163 while( i > cur_nonb )
1164 {
1165 slicetype_frame_cost( h, a, frames, cur_nonb, last_nonb, i );
1166 macroblock_tree_propagate( h, frames, average_duration, cur_nonb, last_nonb, i, 0 );
1167 i--;
1168 }
1169 }
1170 macroblock_tree_propagate( h, frames, average_duration, cur_nonb, last_nonb, last_nonb, 1 );
1171 last_nonb = cur_nonb;
1172 }
1173
1174 if( !h->param.rc.i_lookahead )
1175 {
1176 slicetype_frame_cost( h, a, frames, 0, last_nonb, last_nonb );
1177 macroblock_tree_propagate( h, frames, average_duration, 0, last_nonb, last_nonb, 1 );
1178 XCHG( uint16_t*, frames[last_nonb]->i_propagate_cost, frames[0]->i_propagate_cost );
1179 }
1180
1181 macroblock_tree_finish( h, frames[last_nonb], average_duration, last_nonb );
1182 if( h->param.i_bframe_pyramid && bframes > 1 && !h->param.rc.i_vbv_buffer_size )
1183 macroblock_tree_finish( h, frames[last_nonb+(bframes+1)/2], average_duration, 0 );
1184 }
1185
vbv_frame_cost(x264_t * h,x264_mb_analysis_t * a,x264_frame_t ** frames,int p0,int p1,int b)1186 static int vbv_frame_cost( x264_t *h, x264_mb_analysis_t *a, x264_frame_t **frames, int p0, int p1, int b )
1187 {
1188 int cost = slicetype_frame_cost( h, a, frames, p0, p1, b );
1189 if( h->param.rc.i_aq_mode )
1190 {
1191 if( h->param.rc.b_mb_tree )
1192 return slicetype_frame_cost_recalculate( h, frames, p0, p1, b );
1193 else
1194 return frames[b]->i_cost_est_aq[b-p0][p1-b];
1195 }
1196 return cost;
1197 }
1198
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)1199 static void 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 {
1201 cur_frame->i_cpb_delay = *i_cpb_delay;
1202 cur_frame->i_dpb_output_delay = cur_frame->i_field_cnt - *i_coded_fields;
1203
1204 // add a correction term for frame reordering
1205 cur_frame->i_dpb_output_delay += h->sps->vui.i_num_reorder_frames*2;
1206
1207 // fix possible negative dpb_output_delay because of pulldown changes and reordering
1208 if( cur_frame->i_dpb_output_delay < 0 )
1209 {
1210 cur_frame->i_cpb_delay += cur_frame->i_dpb_output_delay;
1211 cur_frame->i_dpb_output_delay = 0;
1212 if( prev_frame )
1213 prev_frame->i_cpb_duration += cur_frame->i_dpb_output_delay;
1214 }
1215
1216 // don't reset cpb delay for IDR frames when using intra-refresh
1217 if( cur_frame->b_keyframe && !h->param.b_intra_refresh )
1218 *i_cpb_delay = 0;
1219
1220 *i_cpb_delay += cur_frame->i_duration;
1221 *i_coded_fields += cur_frame->i_duration;
1222 cur_frame->i_cpb_duration = cur_frame->i_duration;
1223 }
1224
vbv_lookahead(x264_t * h,x264_mb_analysis_t * a,x264_frame_t ** frames,int num_frames,int keyframe)1225 static void vbv_lookahead( x264_t *h, x264_mb_analysis_t *a, x264_frame_t **frames, int num_frames, int keyframe )
1226 {
1227 int last_nonb = 0, cur_nonb = 1, idx = 0;
1228 x264_frame_t *prev_frame = NULL;
1229 int prev_frame_idx = 0;
1230 while( cur_nonb < num_frames && IS_X264_TYPE_B( frames[cur_nonb]->i_type ) )
1231 cur_nonb++;
1232 int next_nonb = keyframe ? last_nonb : cur_nonb;
1233
1234 if( frames[cur_nonb]->i_coded_fields_lookahead >= 0 )
1235 {
1236 h->i_coded_fields_lookahead = frames[cur_nonb]->i_coded_fields_lookahead;
1237 h->i_cpb_delay_lookahead = frames[cur_nonb]->i_cpb_delay_lookahead;
1238 }
1239
1240 while( cur_nonb < num_frames )
1241 {
1242 /* P/I cost: This shouldn't include the cost of next_nonb */
1243 if( next_nonb != cur_nonb )
1244 {
1245 int p0 = IS_X264_TYPE_I( frames[cur_nonb]->i_type ) ? cur_nonb : last_nonb;
1246 frames[next_nonb]->i_planned_satd[idx] = vbv_frame_cost( h, a, frames, p0, cur_nonb, cur_nonb );
1247 frames[next_nonb]->i_planned_type[idx] = frames[cur_nonb]->i_type;
1248 frames[cur_nonb]->i_coded_fields_lookahead = h->i_coded_fields_lookahead;
1249 frames[cur_nonb]->i_cpb_delay_lookahead = h->i_cpb_delay_lookahead;
1250 calculate_durations( h, frames[cur_nonb], prev_frame, &h->i_cpb_delay_lookahead, &h->i_coded_fields_lookahead );
1251 if( prev_frame )
1252 {
1253 frames[next_nonb]->f_planned_cpb_duration[prev_frame_idx] = (double)prev_frame->i_cpb_duration *
1254 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1255 }
1256 frames[next_nonb]->f_planned_cpb_duration[idx] = (double)frames[cur_nonb]->i_cpb_duration *
1257 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1258 prev_frame = frames[cur_nonb];
1259 prev_frame_idx = idx;
1260 idx++;
1261 }
1262 /* Handle the B-frames: coded order */
1263 for( int i = last_nonb+1; i < cur_nonb; i++, idx++ )
1264 {
1265 frames[next_nonb]->i_planned_satd[idx] = vbv_frame_cost( h, a, frames, last_nonb, cur_nonb, i );
1266 frames[next_nonb]->i_planned_type[idx] = X264_TYPE_B;
1267 frames[i]->i_coded_fields_lookahead = h->i_coded_fields_lookahead;
1268 frames[i]->i_cpb_delay_lookahead = h->i_cpb_delay_lookahead;
1269 calculate_durations( h, frames[i], prev_frame, &h->i_cpb_delay_lookahead, &h->i_coded_fields_lookahead );
1270 if( prev_frame )
1271 {
1272 frames[next_nonb]->f_planned_cpb_duration[prev_frame_idx] = (double)prev_frame->i_cpb_duration *
1273 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1274 }
1275 frames[next_nonb]->f_planned_cpb_duration[idx] = (double)frames[i]->i_cpb_duration *
1276 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1277 prev_frame = frames[i];
1278 prev_frame_idx = idx;
1279 }
1280 last_nonb = cur_nonb;
1281 cur_nonb++;
1282 while( cur_nonb <= num_frames && IS_X264_TYPE_B( frames[cur_nonb]->i_type ) )
1283 cur_nonb++;
1284 }
1285 frames[next_nonb]->i_planned_type[idx] = X264_TYPE_AUTO;
1286 }
1287
slicetype_path_cost(x264_t * h,x264_mb_analysis_t * a,x264_frame_t ** frames,char * path,uint64_t threshold)1288 static uint64_t slicetype_path_cost( x264_t *h, x264_mb_analysis_t *a, x264_frame_t **frames, char *path, uint64_t threshold )
1289 {
1290 uint64_t cost = 0;
1291 int loc = 1;
1292 int cur_nonb = 0;
1293 path--; /* Since the 1st path element is really the second frame */
1294 while( path[loc] )
1295 {
1296 int next_nonb = loc;
1297 /* Find the location of the next non-B-frame. */
1298 while( path[next_nonb] == 'B' )
1299 next_nonb++;
1300
1301 /* Add the cost of the non-B-frame found above */
1302 if( path[next_nonb] == 'P' )
1303 cost += slicetype_frame_cost( h, a, frames, cur_nonb, next_nonb, next_nonb );
1304 else /* I-frame */
1305 cost += slicetype_frame_cost( h, a, frames, next_nonb, next_nonb, next_nonb );
1306 /* Early terminate if the cost we have found is larger than the best path cost so far */
1307 if( cost > threshold )
1308 break;
1309
1310 if( h->param.i_bframe_pyramid && next_nonb - cur_nonb > 2 )
1311 {
1312 int middle = cur_nonb + (next_nonb - cur_nonb)/2;
1313 cost += slicetype_frame_cost( h, a, frames, cur_nonb, next_nonb, middle );
1314 for( int next_b = loc; next_b < middle && cost < threshold; next_b++ )
1315 cost += slicetype_frame_cost( h, a, frames, cur_nonb, middle, next_b );
1316 for( int next_b = middle+1; next_b < next_nonb && cost < threshold; next_b++ )
1317 cost += slicetype_frame_cost( h, a, frames, middle, next_nonb, next_b );
1318 }
1319 else
1320 for( int next_b = loc; next_b < next_nonb && cost < threshold; next_b++ )
1321 cost += slicetype_frame_cost( h, a, frames, cur_nonb, next_nonb, next_b );
1322
1323 loc = next_nonb + 1;
1324 cur_nonb = next_nonb;
1325 }
1326 return cost;
1327 }
1328
1329 /* Viterbi/trellis slicetype decision algorithm. */
1330 /* Uses strings due to the fact that the speed of the control functions is
1331 negligible compared to the cost of running slicetype_frame_cost, and because
1332 it makes debugging easier. */
slicetype_path(x264_t * h,x264_mb_analysis_t * a,x264_frame_t ** frames,int length,char (* best_paths)[X264_LOOKAHEAD_MAX+1])1333 static void slicetype_path( x264_t *h, x264_mb_analysis_t *a, x264_frame_t **frames, int length, char (*best_paths)[X264_LOOKAHEAD_MAX+1] )
1334 {
1335 char paths[2][X264_LOOKAHEAD_MAX+1];
1336 int num_paths = X264_MIN( h->param.i_bframe+1, length );
1337 uint64_t best_cost = COST_MAX64;
1338 int best_possible = 0;
1339 int idx = 0;
1340
1341 /* Iterate over all currently possible paths */
1342 for( int path = 0; path < num_paths; path++ )
1343 {
1344 /* Add suffixes to the current path */
1345 int len = length - (path + 1);
1346 memcpy( paths[idx], best_paths[len % (X264_BFRAME_MAX+1)], len );
1347 memset( paths[idx]+len, 'B', path );
1348 strcpy( paths[idx]+len+path, "P" );
1349
1350 int possible = 1;
1351 for( int i = 1; i <= length; i++ )
1352 {
1353 int i_type = frames[i]->i_type;
1354 if( i_type == X264_TYPE_AUTO )
1355 continue;
1356 if( IS_X264_TYPE_B( i_type ) )
1357 possible = possible && (i < len || i == length || paths[idx][i-1] == 'B');
1358 else
1359 {
1360 possible = possible && (i < len || paths[idx][i-1] != 'B');
1361 paths[idx][i-1] = IS_X264_TYPE_I( i_type ) ? 'I' : 'P';
1362 }
1363 }
1364
1365 if( possible || !best_possible )
1366 {
1367 if( possible && !best_possible )
1368 best_cost = COST_MAX64;
1369 /* Calculate the actual cost of the current path */
1370 uint64_t cost = slicetype_path_cost( h, a, frames, paths[idx], best_cost );
1371 if( cost < best_cost )
1372 {
1373 best_cost = cost;
1374 best_possible = possible;
1375 idx ^= 1;
1376 }
1377 }
1378 }
1379
1380 /* Store the best path. */
1381 memcpy( best_paths[length % (X264_BFRAME_MAX+1)], paths[idx^1], length );
1382 }
1383
scenecut_internal(x264_t * h,x264_mb_analysis_t * a,x264_frame_t ** frames,int p0,int p1,int real_scenecut)1384 static int scenecut_internal( x264_t *h, x264_mb_analysis_t *a, x264_frame_t **frames, int p0, int p1, int real_scenecut )
1385 {
1386 x264_frame_t *frame = frames[p1];
1387
1388 /* Don't do scenecuts on the right view of a frame-packed video. */
1389 if( real_scenecut && h->param.i_frame_packing == 5 && (frame->i_frame&1) )
1390 return 0;
1391
1392 slicetype_frame_cost( h, a, frames, p0, p1, p1 );
1393
1394 int icost = frame->i_cost_est[0][0];
1395 int pcost = frame->i_cost_est[p1-p0][0];
1396 float f_bias;
1397 int i_gop_size = frame->i_frame - h->lookahead->i_last_keyframe;
1398 float f_thresh_max = h->param.i_scenecut_threshold / 100.0;
1399 /* magic numbers pulled out of thin air */
1400 float f_thresh_min = f_thresh_max * 0.25;
1401 int res;
1402
1403 if( h->param.i_keyint_min == h->param.i_keyint_max )
1404 f_thresh_min = f_thresh_max;
1405 if( i_gop_size <= h->param.i_keyint_min / 4 || h->param.b_intra_refresh )
1406 f_bias = f_thresh_min / 4;
1407 else if( i_gop_size <= h->param.i_keyint_min )
1408 f_bias = f_thresh_min * i_gop_size / h->param.i_keyint_min;
1409 else
1410 {
1411 f_bias = f_thresh_min
1412 + ( f_thresh_max - f_thresh_min )
1413 * ( i_gop_size - h->param.i_keyint_min )
1414 / ( h->param.i_keyint_max - h->param.i_keyint_min );
1415 }
1416
1417 res = pcost >= (1.0 - f_bias) * icost;
1418 if( res && real_scenecut )
1419 {
1420 int imb = frame->i_intra_mbs[p1-p0];
1421 int pmb = NUM_MBS - imb;
1422 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",
1423 frame->i_frame,
1424 icost, pcost, 1. - (double)pcost / icost,
1425 f_bias, i_gop_size, imb, pmb );
1426 }
1427 return res;
1428 }
1429
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)1430 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 )
1431 {
1432 /* Only do analysis during a normal scenecut check. */
1433 if( real_scenecut && h->param.i_bframe )
1434 {
1435 int origmaxp1 = p0 + 1;
1436 /* Look ahead to avoid coding short flashes as scenecuts. */
1437 if( h->param.i_bframe_adaptive == X264_B_ADAPT_TRELLIS )
1438 /* Don't analyse any more frames than the trellis would have covered. */
1439 origmaxp1 += h->param.i_bframe;
1440 else
1441 origmaxp1++;
1442 int maxp1 = X264_MIN( origmaxp1, num_frames );
1443
1444 /* Where A and B are scenes: AAAAAABBBAAAAAA
1445 * If BBB is shorter than (maxp1-p0), it is detected as a flash
1446 * and not considered a scenecut. */
1447 for( int curp1 = p1; curp1 <= maxp1; curp1++ )
1448 if( !scenecut_internal( h, a, frames, p0, curp1, 0 ) )
1449 /* Any frame in between p0 and cur_p1 cannot be a real scenecut. */
1450 for( int i = curp1; i > p0; i-- )
1451 frames[i]->b_scenecut = 0;
1452
1453 /* Where A-F are scenes: AAAAABBCCDDEEFFFFFF
1454 * If each of BB ... EE are shorter than (maxp1-p0), they are
1455 * detected as flashes and not considered scenecuts.
1456 * Instead, the first F frame becomes a scenecut.
1457 * If the video ends before F, no frame becomes a scenecut. */
1458 for( int curp0 = p0; curp0 <= maxp1; curp0++ )
1459 if( origmaxp1 > i_max_search || (curp0 < maxp1 && scenecut_internal( h, a, frames, curp0, maxp1, 0 )) )
1460 /* If cur_p0 is the p0 of a scenecut, it cannot be the p1 of a scenecut. */
1461 frames[curp0]->b_scenecut = 0;
1462 }
1463
1464 /* Ignore frames that are part of a flash, i.e. cannot be real scenecuts. */
1465 if( !frames[p1]->b_scenecut )
1466 return 0;
1467 return scenecut_internal( h, a, frames, p0, p1, real_scenecut );
1468 }
1469
1470 #define IS_X264_TYPE_AUTO_OR_I(x) ((x)==X264_TYPE_AUTO || IS_X264_TYPE_I(x))
1471 #define IS_X264_TYPE_AUTO_OR_B(x) ((x)==X264_TYPE_AUTO || IS_X264_TYPE_B(x))
1472
x264_slicetype_analyse(x264_t * h,int intra_minigop)1473 void x264_slicetype_analyse( x264_t *h, int intra_minigop )
1474 {
1475 x264_mb_analysis_t a;
1476 x264_frame_t *frames[X264_LOOKAHEAD_MAX+3] = { NULL, };
1477 int num_frames, orig_num_frames, keyint_limit, framecnt;
1478 int i_max_search = X264_MIN( h->lookahead->next.i_size, X264_LOOKAHEAD_MAX );
1479 int b_vbv_lookahead = h->param.rc.i_vbv_buffer_size && h->param.rc.i_lookahead;
1480 /* For determinism we should limit the search to the number of frames lookahead has for sure
1481 * in h->lookahead->next.list buffer, except at the end of stream.
1482 * For normal calls with (intra_minigop == 0) that is h->lookahead->i_slicetype_length + 1 frames.
1483 * And for I-frame calls (intra_minigop != 0) we already removed intra_minigop frames from there. */
1484 if( h->param.b_deterministic )
1485 i_max_search = X264_MIN( i_max_search, h->lookahead->i_slicetype_length + 1 - intra_minigop );
1486 int keyframe = !!intra_minigop;
1487
1488 assert( h->frames.b_have_lowres );
1489
1490 if( !h->lookahead->last_nonb )
1491 return;
1492 frames[0] = h->lookahead->last_nonb;
1493 for( framecnt = 0; framecnt < i_max_search; framecnt++ )
1494 frames[framecnt+1] = h->lookahead->next.list[framecnt];
1495
1496 lowres_context_init( h, &a );
1497
1498 if( !framecnt )
1499 {
1500 if( h->param.rc.b_mb_tree )
1501 macroblock_tree( h, &a, frames, 0, keyframe );
1502 return;
1503 }
1504
1505 keyint_limit = h->param.i_keyint_max - frames[0]->i_frame + h->lookahead->i_last_keyframe - 1;
1506 orig_num_frames = num_frames = h->param.b_intra_refresh ? framecnt : X264_MIN( framecnt, keyint_limit );
1507
1508 /* This is important psy-wise: if we have a non-scenecut keyframe,
1509 * there will be significant visual artifacts if the frames just before
1510 * go down in quality due to being referenced less, despite it being
1511 * more RD-optimal. */
1512 if( (h->param.analyse.b_psy && h->param.rc.b_mb_tree) || b_vbv_lookahead )
1513 num_frames = framecnt;
1514 else if( h->param.b_open_gop && num_frames < framecnt )
1515 num_frames++;
1516 else if( num_frames == 0 )
1517 {
1518 frames[1]->i_type = X264_TYPE_I;
1519 return;
1520 }
1521
1522 if( IS_X264_TYPE_AUTO_OR_I( frames[1]->i_type ) &&
1523 h->param.i_scenecut_threshold && scenecut( h, &a, frames, 0, 1, 1, orig_num_frames, i_max_search ) )
1524 {
1525 if( frames[1]->i_type == X264_TYPE_AUTO )
1526 frames[1]->i_type = X264_TYPE_I;
1527 return;
1528 }
1529
1530 #if HAVE_OPENCL
1531 x264_opencl_slicetype_prep( h, frames, num_frames, a.i_lambda );
1532 #endif
1533
1534 /* Replace forced keyframes with I/IDR-frames */
1535 for( int j = 1; j <= num_frames; j++ )
1536 {
1537 if( frames[j]->i_type == X264_TYPE_KEYFRAME )
1538 frames[j]->i_type = h->param.b_open_gop ? X264_TYPE_I : X264_TYPE_IDR;
1539 }
1540
1541 /* Close GOP at IDR-frames */
1542 for( int j = 2; j <= num_frames; j++ )
1543 {
1544 if( frames[j]->i_type == X264_TYPE_IDR && IS_X264_TYPE_AUTO_OR_B( frames[j-1]->i_type ) )
1545 frames[j-1]->i_type = X264_TYPE_P;
1546 }
1547
1548 int num_analysed_frames = num_frames;
1549 int reset_start;
1550
1551 if( h->param.i_bframe )
1552 {
1553 if( h->param.i_bframe_adaptive == X264_B_ADAPT_TRELLIS )
1554 {
1555 if( num_frames > 1 )
1556 {
1557 char best_paths[X264_BFRAME_MAX+1][X264_LOOKAHEAD_MAX+1] = {"","P"};
1558 int best_path_index = num_frames % (X264_BFRAME_MAX+1);
1559
1560 /* Perform the frametype analysis. */
1561 for( int j = 2; j <= num_frames; j++ )
1562 slicetype_path( h, &a, frames, j, best_paths );
1563
1564 /* Load the results of the analysis into the frame types. */
1565 for( int j = 1; j < num_frames; j++ )
1566 {
1567 if( best_paths[best_path_index][j-1] != 'B' )
1568 {
1569 if( IS_X264_TYPE_AUTO_OR_B( frames[j]->i_type ) )
1570 frames[j]->i_type = X264_TYPE_P;
1571 }
1572 else
1573 {
1574 if( frames[j]->i_type == X264_TYPE_AUTO )
1575 frames[j]->i_type = X264_TYPE_B;
1576 }
1577 }
1578 }
1579 }
1580 else if( h->param.i_bframe_adaptive == X264_B_ADAPT_FAST )
1581 {
1582 int last_nonb = 0;
1583 int num_bframes = h->param.i_bframe;
1584 char path[X264_LOOKAHEAD_MAX+1];
1585 for( int j = 1; j < num_frames; j++ )
1586 {
1587 if( j-1 > 0 && IS_X264_TYPE_B( frames[j-1]->i_type ) )
1588 num_bframes--;
1589 else
1590 {
1591 last_nonb = j-1;
1592 num_bframes = h->param.i_bframe;
1593 }
1594 if( !num_bframes )
1595 {
1596 if( IS_X264_TYPE_AUTO_OR_B( frames[j]->i_type ) )
1597 frames[j]->i_type = X264_TYPE_P;
1598 continue;
1599 }
1600
1601 if( frames[j]->i_type != X264_TYPE_AUTO )
1602 continue;
1603
1604 if( IS_X264_TYPE_B( frames[j+1]->i_type ) )
1605 {
1606 frames[j]->i_type = X264_TYPE_P;
1607 continue;
1608 }
1609
1610 int bframes = j - last_nonb - 1;
1611 memset( path, 'B', bframes );
1612 strcpy( path+bframes, "PP" );
1613 uint64_t cost_p = slicetype_path_cost( h, &a, frames+last_nonb, path, COST_MAX64 );
1614 strcpy( path+bframes, "BP" );
1615 uint64_t cost_b = slicetype_path_cost( h, &a, frames+last_nonb, path, cost_p );
1616
1617 if( cost_b < cost_p )
1618 frames[j]->i_type = X264_TYPE_B;
1619 else
1620 frames[j]->i_type = X264_TYPE_P;
1621 }
1622 }
1623 else
1624 {
1625 int num_bframes = h->param.i_bframe;
1626 for( int j = 1; j < num_frames; j++ )
1627 {
1628 if( !num_bframes )
1629 {
1630 if( IS_X264_TYPE_AUTO_OR_B( frames[j]->i_type ) )
1631 frames[j]->i_type = X264_TYPE_P;
1632 }
1633 else if( frames[j]->i_type == X264_TYPE_AUTO )
1634 {
1635 if( IS_X264_TYPE_B( frames[j+1]->i_type ) )
1636 frames[j]->i_type = X264_TYPE_P;
1637 else
1638 frames[j]->i_type = X264_TYPE_B;
1639 }
1640 if( IS_X264_TYPE_B( frames[j]->i_type ) )
1641 num_bframes--;
1642 else
1643 num_bframes = h->param.i_bframe;
1644 }
1645 }
1646 if( IS_X264_TYPE_AUTO_OR_B( frames[num_frames]->i_type ) )
1647 frames[num_frames]->i_type = X264_TYPE_P;
1648
1649 int num_bframes = 0;
1650 while( num_bframes < num_frames && IS_X264_TYPE_B( frames[num_bframes+1]->i_type ) )
1651 num_bframes++;
1652
1653 /* Check scenecut on the first minigop. */
1654 for( int j = 1; j < num_bframes+1; j++ )
1655 {
1656 if( frames[j]->i_forced_type == X264_TYPE_AUTO && IS_X264_TYPE_AUTO_OR_I( frames[j+1]->i_forced_type ) &&
1657 h->param.i_scenecut_threshold && scenecut( h, &a, frames, j, j+1, 0, orig_num_frames, i_max_search ) )
1658 {
1659 frames[j]->i_type = X264_TYPE_P;
1660 num_analysed_frames = j;
1661 break;
1662 }
1663 }
1664
1665 reset_start = keyframe ? 1 : X264_MIN( num_bframes+2, num_analysed_frames+1 );
1666 }
1667 else
1668 {
1669 for( int j = 1; j <= num_frames; j++ )
1670 if( IS_X264_TYPE_AUTO_OR_B( frames[j]->i_type ) )
1671 frames[j]->i_type = X264_TYPE_P;
1672 reset_start = !keyframe + 1;
1673 }
1674
1675 /* Perform the actual macroblock tree analysis.
1676 * Don't go farther than the maximum keyframe interval; this helps in short GOPs. */
1677 if( h->param.rc.b_mb_tree )
1678 macroblock_tree( h, &a, frames, X264_MIN(num_frames, h->param.i_keyint_max), keyframe );
1679
1680 /* Enforce keyframe limit. */
1681 if( !h->param.b_intra_refresh )
1682 {
1683 int last_keyframe = h->lookahead->i_last_keyframe;
1684 int last_possible = 0;
1685 for( int j = 1; j <= num_frames; j++ )
1686 {
1687 x264_frame_t *frm = frames[j];
1688 int keyframe_dist = frm->i_frame - last_keyframe;
1689
1690 if( IS_X264_TYPE_AUTO_OR_I( frm->i_forced_type ) )
1691 {
1692 if( h->param.b_open_gop || !IS_X264_TYPE_B( frames[j-1]->i_forced_type ) )
1693 last_possible = j;
1694 }
1695 if( keyframe_dist >= h->param.i_keyint_max )
1696 {
1697 if( last_possible != 0 && last_possible != j )
1698 {
1699 j = last_possible;
1700 frm = frames[j];
1701 keyframe_dist = frm->i_frame - last_keyframe;
1702 }
1703 last_possible = 0;
1704 if( frm->i_type != X264_TYPE_IDR )
1705 frm->i_type = h->param.b_open_gop ? X264_TYPE_I : X264_TYPE_IDR;
1706 }
1707 if( frm->i_type == X264_TYPE_I && keyframe_dist >= h->param.i_keyint_min )
1708 {
1709 if( h->param.b_open_gop )
1710 {
1711 last_keyframe = frm->i_frame;
1712 if( h->param.b_bluray_compat )
1713 {
1714 // Use bluray order
1715 int bframes = 0;
1716 while( bframes < j-1 && IS_X264_TYPE_B( frames[j-1-bframes]->i_type ) )
1717 bframes++;
1718 last_keyframe -= bframes;
1719 }
1720 }
1721 else if( frm->i_forced_type != X264_TYPE_I )
1722 frm->i_type = X264_TYPE_IDR;
1723 }
1724 if( frm->i_type == X264_TYPE_IDR )
1725 {
1726 last_keyframe = frm->i_frame;
1727 if( j > 1 && IS_X264_TYPE_B( frames[j-1]->i_type ) )
1728 frames[j-1]->i_type = X264_TYPE_P;
1729 }
1730 }
1731 }
1732
1733 if( b_vbv_lookahead )
1734 vbv_lookahead( h, &a, frames, num_frames, keyframe );
1735
1736 /* Restore frametypes for all frames that haven't actually been decided yet. */
1737 for( int j = reset_start; j <= num_frames; j++ )
1738 frames[j]->i_type = frames[j]->i_forced_type;
1739
1740 #if HAVE_OPENCL
1741 x264_opencl_slicetype_end( h );
1742 #endif
1743 }
1744
x264_slicetype_decide(x264_t * h)1745 void x264_slicetype_decide( x264_t *h )
1746 {
1747 x264_frame_t *frames[X264_BFRAME_MAX+2];
1748 x264_frame_t *frm;
1749 int bframes;
1750 int brefs;
1751
1752 if( !h->lookahead->next.i_size )
1753 return;
1754
1755 int lookahead_size = h->lookahead->next.i_size;
1756
1757 for( int i = 0; i < h->lookahead->next.i_size; i++ )
1758 {
1759 if( h->param.b_vfr_input )
1760 {
1761 if( lookahead_size-- > 1 )
1762 h->lookahead->next.list[i]->i_duration = 2 * (h->lookahead->next.list[i+1]->i_pts - h->lookahead->next.list[i]->i_pts);
1763 else
1764 h->lookahead->next.list[i]->i_duration = h->i_prev_duration;
1765 }
1766 else
1767 h->lookahead->next.list[i]->i_duration = delta_tfi_divisor[h->lookahead->next.list[i]->i_pic_struct];
1768 h->i_prev_duration = h->lookahead->next.list[i]->i_duration;
1769 h->lookahead->next.list[i]->f_duration = (double)h->lookahead->next.list[i]->i_duration
1770 * h->sps->vui.i_num_units_in_tick
1771 / h->sps->vui.i_time_scale;
1772
1773 if( h->lookahead->next.list[i]->i_frame > h->i_disp_fields_last_frame && lookahead_size > 0 )
1774 {
1775 h->lookahead->next.list[i]->i_field_cnt = h->i_disp_fields;
1776 h->i_disp_fields += h->lookahead->next.list[i]->i_duration;
1777 h->i_disp_fields_last_frame = h->lookahead->next.list[i]->i_frame;
1778 }
1779 else if( lookahead_size == 0 )
1780 {
1781 h->lookahead->next.list[i]->i_field_cnt = h->i_disp_fields;
1782 h->lookahead->next.list[i]->i_duration = h->i_prev_duration;
1783 }
1784 }
1785
1786 if( h->param.rc.b_stat_read )
1787 {
1788 /* Use the frame types from the first pass */
1789 for( int i = 0; i < h->lookahead->next.i_size; i++ )
1790 h->lookahead->next.list[i]->i_type =
1791 x264_ratecontrol_slice_type( h, h->lookahead->next.list[i]->i_frame );
1792 }
1793 else if( (h->param.i_bframe && h->param.i_bframe_adaptive)
1794 || h->param.i_scenecut_threshold
1795 || h->param.rc.b_mb_tree
1796 || (h->param.rc.i_vbv_buffer_size && h->param.rc.i_lookahead) )
1797 x264_slicetype_analyse( h, 0 );
1798
1799 for( bframes = 0, brefs = 0;; bframes++ )
1800 {
1801 frm = h->lookahead->next.list[bframes];
1802
1803 if( frm->i_forced_type != X264_TYPE_AUTO && frm->i_type != frm->i_forced_type &&
1804 !(frm->i_forced_type == X264_TYPE_KEYFRAME && IS_X264_TYPE_I( frm->i_type )) )
1805 {
1806 x264_log( h, X264_LOG_WARNING, "forced frame type (%d) at %d was changed to frame type (%d)\n",
1807 frm->i_forced_type, frm->i_frame, frm->i_type );
1808 }
1809
1810 if( frm->i_type == X264_TYPE_BREF && h->param.i_bframe_pyramid < X264_B_PYRAMID_NORMAL &&
1811 brefs == h->param.i_bframe_pyramid )
1812 {
1813 frm->i_type = X264_TYPE_B;
1814 x264_log( h, X264_LOG_WARNING, "B-ref at frame %d incompatible with B-pyramid %s \n",
1815 frm->i_frame, x264_b_pyramid_names[h->param.i_bframe_pyramid] );
1816 }
1817 /* pyramid with multiple B-refs needs a big enough dpb that the preceding P-frame stays available.
1818 smaller dpb could be supported by smart enough use of mmco, but it's easier just to forbid it. */
1819 else if( frm->i_type == X264_TYPE_BREF && h->param.i_bframe_pyramid == X264_B_PYRAMID_NORMAL &&
1820 brefs && h->param.i_frame_reference <= (brefs+3) )
1821 {
1822 frm->i_type = X264_TYPE_B;
1823 x264_log( h, X264_LOG_WARNING, "B-ref at frame %d incompatible with B-pyramid %s and %d reference frames\n",
1824 frm->i_frame, x264_b_pyramid_names[h->param.i_bframe_pyramid], h->param.i_frame_reference );
1825 }
1826
1827 if( frm->i_type == X264_TYPE_KEYFRAME )
1828 frm->i_type = h->param.b_open_gop ? X264_TYPE_I : X264_TYPE_IDR;
1829
1830 /* Limit GOP size */
1831 if( (!h->param.b_intra_refresh || frm->i_frame == 0) && frm->i_frame - h->lookahead->i_last_keyframe >= h->param.i_keyint_max )
1832 {
1833 if( frm->i_type == X264_TYPE_AUTO || frm->i_type == X264_TYPE_I )
1834 frm->i_type = h->param.b_open_gop && h->lookahead->i_last_keyframe >= 0 ? X264_TYPE_I : X264_TYPE_IDR;
1835 int warn = frm->i_type != X264_TYPE_IDR;
1836 if( warn && h->param.b_open_gop )
1837 warn &= frm->i_type != X264_TYPE_I;
1838 if( warn )
1839 {
1840 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 );
1841 frm->i_type = h->param.b_open_gop && h->lookahead->i_last_keyframe >= 0 ? X264_TYPE_I : X264_TYPE_IDR;
1842 }
1843 }
1844 if( frm->i_type == X264_TYPE_I && frm->i_frame - h->lookahead->i_last_keyframe >= h->param.i_keyint_min )
1845 {
1846 if( h->param.b_open_gop )
1847 {
1848 h->lookahead->i_last_keyframe = frm->i_frame; // Use display order
1849 if( h->param.b_bluray_compat )
1850 h->lookahead->i_last_keyframe -= bframes; // Use bluray order
1851 frm->b_keyframe = 1;
1852 }
1853 else
1854 frm->i_type = X264_TYPE_IDR;
1855 }
1856 if( frm->i_type == X264_TYPE_IDR )
1857 {
1858 /* Close GOP */
1859 h->lookahead->i_last_keyframe = frm->i_frame;
1860 frm->b_keyframe = 1;
1861 if( bframes > 0 )
1862 {
1863 bframes--;
1864 h->lookahead->next.list[bframes]->i_type = X264_TYPE_P;
1865 }
1866 }
1867
1868 if( bframes == h->param.i_bframe ||
1869 !h->lookahead->next.list[bframes+1] )
1870 {
1871 if( IS_X264_TYPE_B( frm->i_type ) )
1872 x264_log( h, X264_LOG_WARNING, "specified frame type is not compatible with max B-frames\n" );
1873 if( frm->i_type == X264_TYPE_AUTO
1874 || IS_X264_TYPE_B( frm->i_type ) )
1875 frm->i_type = X264_TYPE_P;
1876 }
1877
1878 if( frm->i_type == X264_TYPE_BREF )
1879 brefs++;
1880
1881 if( frm->i_type == X264_TYPE_AUTO )
1882 frm->i_type = X264_TYPE_B;
1883
1884 else if( !IS_X264_TYPE_B( frm->i_type ) ) break;
1885 }
1886
1887 if( bframes )
1888 h->lookahead->next.list[bframes-1]->b_last_minigop_bframe = 1;
1889 h->lookahead->next.list[bframes]->i_bframes = bframes;
1890
1891 /* insert a bref into the sequence */
1892 if( h->param.i_bframe_pyramid && bframes > 1 && !brefs )
1893 {
1894 h->lookahead->next.list[(bframes-1)/2]->i_type = X264_TYPE_BREF;
1895 brefs++;
1896 }
1897
1898 /* calculate the frame costs ahead of time for x264_rc_analyse_slice while we still have lowres */
1899 if( h->param.rc.i_rc_method != X264_RC_CQP )
1900 {
1901 x264_mb_analysis_t a;
1902 int p0, p1, b;
1903 p1 = b = bframes + 1;
1904
1905 lowres_context_init( h, &a );
1906
1907 frames[0] = h->lookahead->last_nonb;
1908 memcpy( &frames[1], h->lookahead->next.list, (bframes+1) * sizeof(x264_frame_t*) );
1909 if( IS_X264_TYPE_I( h->lookahead->next.list[bframes]->i_type ) )
1910 p0 = bframes + 1;
1911 else // P
1912 p0 = 0;
1913
1914 slicetype_frame_cost( h, &a, frames, p0, p1, b );
1915
1916 if( (p0 != p1 || bframes) && h->param.rc.i_vbv_buffer_size )
1917 {
1918 /* We need the intra costs for row SATDs. */
1919 slicetype_frame_cost( h, &a, frames, b, b, b );
1920
1921 /* We need B-frame costs for row SATDs. */
1922 p0 = 0;
1923 for( b = 1; b <= bframes; b++ )
1924 {
1925 if( frames[b]->i_type == X264_TYPE_B )
1926 for( p1 = b; frames[p1]->i_type == X264_TYPE_B; )
1927 p1++;
1928 else
1929 p1 = bframes + 1;
1930 slicetype_frame_cost( h, &a, frames, p0, p1, b );
1931 if( frames[b]->i_type == X264_TYPE_BREF )
1932 p0 = b;
1933 }
1934 }
1935 }
1936
1937 /* Analyse for weighted P frames */
1938 if( !h->param.rc.b_stat_read && h->lookahead->next.list[bframes]->i_type == X264_TYPE_P
1939 && h->param.analyse.i_weighted_pred >= X264_WEIGHTP_SIMPLE )
1940 {
1941 x264_emms();
1942 x264_weights_analyse( h, h->lookahead->next.list[bframes], h->lookahead->last_nonb, 0 );
1943 }
1944
1945 /* shift sequence to coded order.
1946 use a small temporary list to avoid shifting the entire next buffer around */
1947 int i_coded = h->lookahead->next.list[0]->i_frame;
1948 if( bframes )
1949 {
1950 int idx_list[] = { brefs+1, 1 };
1951 for( int i = 0; i < bframes; i++ )
1952 {
1953 int idx = idx_list[h->lookahead->next.list[i]->i_type == X264_TYPE_BREF]++;
1954 frames[idx] = h->lookahead->next.list[i];
1955 frames[idx]->i_reordered_pts = h->lookahead->next.list[idx]->i_pts;
1956 }
1957 frames[0] = h->lookahead->next.list[bframes];
1958 frames[0]->i_reordered_pts = h->lookahead->next.list[0]->i_pts;
1959 memcpy( h->lookahead->next.list, frames, (bframes+1) * sizeof(x264_frame_t*) );
1960 }
1961
1962 for( int i = 0; i <= bframes; i++ )
1963 {
1964 h->lookahead->next.list[i]->i_coded = i_coded++;
1965 if( i )
1966 {
1967 calculate_durations( h, h->lookahead->next.list[i], h->lookahead->next.list[i-1], &h->i_cpb_delay, &h->i_coded_fields );
1968 h->lookahead->next.list[0]->f_planned_cpb_duration[i-1] = (double)h->lookahead->next.list[i]->i_cpb_duration *
1969 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1970 }
1971 else
1972 calculate_durations( h, h->lookahead->next.list[i], NULL, &h->i_cpb_delay, &h->i_coded_fields );
1973 }
1974 }
1975
x264_rc_analyse_slice(x264_t * h)1976 int x264_rc_analyse_slice( x264_t *h )
1977 {
1978 int p0 = 0, p1, b;
1979 int cost;
1980 x264_emms();
1981
1982 if( IS_X264_TYPE_I(h->fenc->i_type) )
1983 p1 = b = 0;
1984 else if( h->fenc->i_type == X264_TYPE_P )
1985 p1 = b = h->fenc->i_bframes + 1;
1986 else //B
1987 {
1988 p1 = (h->fref_nearest[1]->i_poc - h->fref_nearest[0]->i_poc)/2;
1989 b = (h->fenc->i_poc - h->fref_nearest[0]->i_poc)/2;
1990 }
1991 /* We don't need to assign p0/p1 since we are not performing any real analysis here. */
1992 x264_frame_t **frames = &h->fenc - b;
1993
1994 /* cost should have been already calculated by x264_slicetype_decide */
1995 cost = frames[b]->i_cost_est[b-p0][p1-b];
1996 assert( cost >= 0 );
1997
1998 if( h->param.rc.b_mb_tree && !h->param.rc.b_stat_read )
1999 {
2000 cost = slicetype_frame_cost_recalculate( h, frames, p0, p1, b );
2001 if( b && h->param.rc.i_vbv_buffer_size )
2002 slicetype_frame_cost_recalculate( h, frames, b, b, b );
2003 }
2004 /* In AQ, use the weighted score instead. */
2005 else if( h->param.rc.i_aq_mode )
2006 cost = frames[b]->i_cost_est_aq[b-p0][p1-b];
2007
2008 h->fenc->i_row_satd = h->fenc->i_row_satds[b-p0][p1-b];
2009 h->fdec->i_row_satd = h->fdec->i_row_satds[b-p0][p1-b];
2010 h->fdec->i_satd = cost;
2011 memcpy( h->fdec->i_row_satd, h->fenc->i_row_satd, h->mb.i_mb_height * sizeof(int) );
2012 if( !IS_X264_TYPE_I(h->fenc->i_type) )
2013 memcpy( h->fdec->i_row_satds[0][0], h->fenc->i_row_satds[0][0], h->mb.i_mb_height * sizeof(int) );
2014
2015 if( h->param.b_intra_refresh && h->param.rc.i_vbv_buffer_size && h->fenc->i_type == X264_TYPE_P )
2016 {
2017 int ip_factor = 256 * h->param.rc.f_ip_factor; /* fix8 */
2018 for( int y = 0; y < h->mb.i_mb_height; y++ )
2019 {
2020 int mb_xy = y * h->mb.i_mb_stride + h->fdec->i_pir_start_col;
2021 for( int x = h->fdec->i_pir_start_col; x <= h->fdec->i_pir_end_col; x++, mb_xy++ )
2022 {
2023 int intra_cost = (h->fenc->i_intra_cost[mb_xy] * ip_factor + 128) >> 8;
2024 int inter_cost = h->fenc->lowres_costs[b-p0][p1-b][mb_xy] & LOWRES_COST_MASK;
2025 int diff = intra_cost - inter_cost;
2026 if( h->param.rc.i_aq_mode )
2027 h->fdec->i_row_satd[y] += (diff * frames[b]->i_inv_qscale_factor[mb_xy] + 128) >> 8;
2028 else
2029 h->fdec->i_row_satd[y] += diff;
2030 cost += diff;
2031 }
2032 }
2033 }
2034
2035 return cost;
2036 }
2037