1 /*****************************************************************************
2 * slicetype.c: h264 encoder library
3 *****************************************************************************
4 * Copyright (C) 2005 Loren Merritt
6 * Authors: Loren Merritt <lorenm@u.washington.edu>
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License as published by
10 * the Free Software Foundation; either version 2 of the License, or
11 * (at your option) any later version.
13 * This program is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 * GNU General Public License for more details.
18 * You should have received a copy of the GNU General Public License
19 * along with this program; if not, write to the Free Software
20 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111, USA.
21 *****************************************************************************/
26 #include "common/common.h"
27 #include "common/cpu.h"
28 #include "macroblock.h"
32 static void x264_lowres_context_init( x264_t *h, x264_mb_analysis_t *a )
34 a->i_qp = 12; // arbitrary, but low because SATD scores are 1/4 normal
35 a->i_lambda = i_qp0_cost_table[ a->i_qp ];
36 x264_mb_analyse_load_costs( h, a );
37 h->mb.i_me_method = X264_MIN( X264_ME_HEX, h->param.analyse.i_me_method ); // maybe dia?
38 h->mb.i_subpel_refine = 4; // 3 should be enough, but not tweaking for speed now
39 h->mb.b_chroma_me = 0;
42 int x264_slicetype_mb_cost( x264_t *h, x264_mb_analysis_t *a,
43 x264_frame_t **frames, int p0, int p1, int b,
44 int dist_scale_factor )
46 x264_frame_t *fref0 = frames[p0];
47 x264_frame_t *fref1 = frames[p1];
48 x264_frame_t *fenc = frames[b];
49 const int b_bidir = (b < p1);
50 const int i_mb_x = h->mb.i_mb_x;
51 const int i_mb_y = h->mb.i_mb_y;
52 const int i_mb_stride = h->sps->i_mb_width;
53 const int i_mb_xy = i_mb_x + i_mb_y * i_mb_stride;
54 const int i_stride = fenc->i_stride_lowres;
55 const int i_pel_offset = 8 * ( i_mb_x + i_mb_y * i_stride );
57 DECLARE_ALIGNED_8( uint8_t pix1[9*FDEC_STRIDE] );
58 uint8_t *pix2 = pix1+8;
60 int i_bcost = COST_MAX;
64 h->mb.pic.p_fenc[0] = h->mb.pic.fenc_buf;
65 h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fenc[0], FENC_STRIDE, &fenc->lowres[0][i_pel_offset], i_stride, 8 );
67 if( !p0 && !p1 && !b )
70 // no need for h->mb.mv_min[]
71 h->mb.mv_min_fpel[0] = -8*h->mb.i_mb_x - 4;
72 h->mb.mv_max_fpel[0] = 8*( h->sps->i_mb_width - h->mb.i_mb_x - 1 ) + 4;
73 h->mb.mv_min_spel[0] = 4*( h->mb.mv_min_fpel[0] - 8 );
74 h->mb.mv_max_spel[0] = 4*( h->mb.mv_max_fpel[0] + 8 );
75 if( h->mb.i_mb_x <= 1 )
77 h->mb.mv_min_fpel[1] = -8*h->mb.i_mb_y - 4;
78 h->mb.mv_max_fpel[1] = 8*( h->sps->i_mb_height - h->mb.i_mb_y - 1 ) + 4;
79 h->mb.mv_min_spel[1] = 4*( h->mb.mv_min_fpel[1] - 8 );
80 h->mb.mv_max_spel[1] = 4*( h->mb.mv_max_fpel[1] + 8 );
83 #define LOAD_HPELS_LUMA(dst, src) \
85 (dst)[0] = &(src)[0][i_pel_offset]; \
86 (dst)[1] = &(src)[1][i_pel_offset]; \
87 (dst)[2] = &(src)[2][i_pel_offset]; \
88 (dst)[3] = &(src)[3][i_pel_offset]; \
90 #define SAVE_MVS( mv0, mv1 ) \
92 fenc->mv[0][i_mb_xy][0] = mv0[0]; \
93 fenc->mv[0][i_mb_xy][1] = mv0[1]; \
96 fenc->mv[1][i_mb_xy][0] = mv1[0]; \
97 fenc->mv[1][i_mb_xy][1] = mv1[1]; \
100 #define CLIP_MV( mv ) \
102 mv[0] = x264_clip3( mv[0], h->mb.mv_min_spel[0], h->mb.mv_max_spel[0] ); \
103 mv[1] = x264_clip3( mv[1], h->mb.mv_min_spel[1], h->mb.mv_max_spel[1] ); \
105 #define TRY_BIDIR( mv0, mv1, penalty ) \
110 h->mc.mc_luma( pix1, 16, m[0].p_fref, m[0].i_stride[0], \
111 (mv0)[0], (mv0)[1], 8, 8 ); \
112 src2 = h->mc.get_ref( pix2, &stride2, m[1].p_fref, m[1].i_stride[0], \
113 (mv1)[0], (mv1)[1], 8, 8 ); \
114 h->mc.avg[PIXEL_8x8]( pix1, 16, src2, stride2 ); \
115 i_cost = penalty + h->pixf.mbcmp[PIXEL_8x8]( \
116 m[0].p_fenc[0], FENC_STRIDE, pix1, 16 ); \
117 if( i_bcost > i_cost ) \
120 SAVE_MVS( mv0, mv1 ); \
124 m[0].i_pixel = PIXEL_8x8;
125 m[0].p_cost_mv = a->p_cost_mv;
126 m[0].i_stride[0] = i_stride;
127 m[0].p_fenc[0] = h->mb.pic.p_fenc[0];
128 LOAD_HPELS_LUMA( m[0].p_fref, fref0->lowres );
132 int16_t *mvr = fref1->mv[0][i_mb_xy];
137 LOAD_HPELS_LUMA( m[1].p_fref, fref1->lowres );
139 dmv[0][0] = ( mvr[0] * dist_scale_factor + 128 ) >> 8;
140 dmv[0][1] = ( mvr[1] * dist_scale_factor + 128 ) >> 8;
141 dmv[1][0] = dmv[0][0] - mvr[0];
142 dmv[1][1] = dmv[0][1] - mvr[1];
146 TRY_BIDIR( dmv[0], dmv[1], 0 );
147 if( dmv[0][0] || dmv[0][1] || dmv[1][0] || dmv[1][1] )
148 TRY_BIDIR( mv0, mv0, 0 );
149 // if( i_bcost < 60 ) // arbitrary threshold
153 i_cost_bak = i_bcost;
154 for( l = 0; l < 1 + b_bidir; l++ )
156 int mvc[4][2] = {{0}}, i_mvc;
157 int16_t (*fenc_mv)[2] = &fenc->mv[l][i_mb_xy];
159 #define MVC(mv) { mvc[i_mvc][0] = mv[0]; mvc[i_mvc][1] = mv[1]; i_mvc++; }
164 MVC(fenc_mv[-i_mb_stride]);
165 if( i_mb_x < h->sps->i_mb_width - 1 )
166 MVC(fenc_mv[-i_mb_stride+1]);
168 MVC(fenc_mv[-i_mb_stride-1]);
171 m[l].mvp[0] = x264_median( mvc[0][0], mvc[1][0], mvc[2][0] );
172 m[l].mvp[1] = x264_median( mvc[0][1], mvc[1][1], mvc[2][1] );
174 x264_me_search( h, &m[l], mvc, i_mvc );
176 m[l].cost -= 2; // remove mvcost from skip mbs
177 if( m[l].mv[0] || m[l].mv[1] )
179 i_bcost = X264_MIN( i_bcost, m[l].cost );
182 if( b_bidir && (m[0].mv[0] || m[0].mv[1] || m[1].mv[0] || m[1].mv[1]) )
183 TRY_BIDIR( m[0].mv, m[1].mv, 5 );
185 if( i_bcost < i_cost_bak )
186 SAVE_MVS( m[0].mv, m[1].mv );
188 //FIXME intra part could be shared across multiple encodings of the frame
190 if( !b_bidir ) // forbid intra-mbs in B-frames, because it's rare and not worth checking
192 uint8_t *pix = &pix1[8+FDEC_STRIDE - 1];
193 uint8_t *src = &fenc->lowres[0][i_pel_offset - 1];
194 const int intra_penalty = 5;
195 int satds[4], i_icost, b_intra;
197 memcpy( pix-FDEC_STRIDE, src-i_stride, 17 );
199 pix[i*FDEC_STRIDE] = src[i*i_stride];
202 if( h->pixf.intra_satd_x3_8x8c && h->pixf.mbcmp[0] == h->pixf.satd[0] )
204 h->pixf.intra_satd_x3_8x8c( h->mb.pic.p_fenc[0], pix, satds );
205 h->predict_8x8c[I_PRED_CHROMA_P]( pix );
206 satds[I_PRED_CHROMA_P] =
207 h->pixf.satd[PIXEL_8x8]( pix, FDEC_STRIDE, h->mb.pic.p_fenc[0], FENC_STRIDE );
213 h->predict_8x8c[i]( pix );
214 satds[i] = h->pixf.mbcmp[PIXEL_8x8]( pix, FDEC_STRIDE, h->mb.pic.p_fenc[0], FENC_STRIDE );
217 i_icost = X264_MIN4( satds[0], satds[1], satds[2], satds[3] );
219 if( i_icost < i_bcost * 2 )
221 DECLARE_ALIGNED_16( uint8_t edge[33] );
222 x264_predict_8x8_filter( pix, edge, ALL_NEIGHBORS, ALL_NEIGHBORS );
226 h->predict_8x8[i]( pix, edge );
227 satd = h->pixf.mbcmp[PIXEL_8x8]( pix, FDEC_STRIDE, h->mb.pic.p_fenc[0], FENC_STRIDE );
228 i_icost = X264_MIN( i_icost, satd );
232 i_icost += intra_penalty;
233 b_intra = i_icost < i_bcost;
236 if( i_mb_x > 0 && i_mb_x < h->sps->i_mb_width - 1
237 && i_mb_y > 0 && i_mb_y < h->sps->i_mb_height - 1 )
239 fenc->i_intra_mbs[b-p0] += b_intra;
240 fenc->i_cost_est[0][0] += i_icost;
249 int x264_slicetype_frame_cost( x264_t *h, x264_mb_analysis_t *a,
250 x264_frame_t **frames, int p0, int p1, int b,
251 int b_intra_penalty )
255 /* Check whether we already evaluated this frame
256 * If we have tried this frame as P, then we have also tried
257 * the preceding frames as B. (is this still true?) */
258 if( frames[b]->i_cost_est[b-p0][p1-b] >= 0 )
260 i_score = frames[b]->i_cost_est[b-p0][p1-b];
264 int dist_scale_factor = 128;
265 int *row_satd = frames[b]->i_row_satds[b-p0][p1-b];
267 /* Init MVs so that we don't have to check edge conditions when loading predictors. */
268 /* FIXME: not needed every time */
269 memset( frames[b]->mv[0], 0, h->sps->i_mb_height * h->sps->i_mb_width * 2*sizeof(int16_t) );
271 memset( frames[b]->mv[1], 0, h->sps->i_mb_height * h->sps->i_mb_width * 2*sizeof(int16_t) );
275 frames[b]->i_intra_mbs[b-p0] = 0;
276 frames[b]->i_cost_est[0][0] = 0;
279 dist_scale_factor = ( ((b-p0) << 8) + ((p1-p0) >> 1) ) / (p1-p0);
281 /* the edge mbs seem to reduce the predictive quality of the
282 * whole frame's score, but are needed for a spatial distribution. */
283 if( h->param.rc.i_vbv_buffer_size )
285 for( h->mb.i_mb_y = 0; h->mb.i_mb_y < h->sps->i_mb_height; h->mb.i_mb_y++ )
287 row_satd[ h->mb.i_mb_y ] = 0;
288 for( h->mb.i_mb_x = 0; h->mb.i_mb_x < h->sps->i_mb_width; h->mb.i_mb_x++ )
290 int i_mb_cost = x264_slicetype_mb_cost( h, a, frames, p0, p1, b, dist_scale_factor );
291 row_satd[ h->mb.i_mb_y ] += i_mb_cost;
292 if( h->mb.i_mb_y > 0 && h->mb.i_mb_y < h->sps->i_mb_height - 1 &&
293 h->mb.i_mb_x > 0 && h->mb.i_mb_x < h->sps->i_mb_width - 1 )
295 i_score += i_mb_cost;
302 for( h->mb.i_mb_y = 1; h->mb.i_mb_y < h->sps->i_mb_height - 1; h->mb.i_mb_y++ )
303 for( h->mb.i_mb_x = 1; h->mb.i_mb_x < h->sps->i_mb_width - 1; h->mb.i_mb_x++ )
304 i_score += x264_slicetype_mb_cost( h, a, frames, p0, p1, b, dist_scale_factor );
308 i_score = i_score * 100 / (120 + h->param.i_bframe_bias);
310 frames[b]->i_cost_est[b-p0][p1-b] = i_score;
311 // fprintf( stderr, "frm %d %c(%d,%d): %6d %6d imb:%d \n", frames[b]->i_frame,
312 // (p1==0?'I':b<p1?'B':'P'), b-p0, p1-b, i_score, frames[b]->i_cost_est[0][0], frames[b]->i_intra_mbs[b-p0] );
313 x264_cpu_restore( h->param.cpu );
316 if( b_intra_penalty )
318 // arbitrary penalty for I-blocks after B-frames
319 int nmb = (h->sps->i_mb_width - 2) * (h->sps->i_mb_height - 2);
320 i_score += i_score * frames[b]->i_intra_mbs[b-p0] / (nmb * 8);
325 static int scenecut( x264_t *h, x264_frame_t *frame, int pdist )
327 int icost = frame->i_cost_est[0][0];
328 int pcost = frame->i_cost_est[pdist][0];
330 int i_gop_size = frame->i_frame - h->frames.i_last_idr;
331 float f_thresh_max = h->param.i_scenecut_threshold / 100.0;
332 /* magic numbers pulled out of thin air */
333 float f_thresh_min = f_thresh_max * h->param.i_keyint_min
334 / ( h->param.i_keyint_max * 4 );
337 if( h->param.i_keyint_min == h->param.i_keyint_max )
338 f_thresh_min= f_thresh_max;
339 if( i_gop_size < h->param.i_keyint_min / 4 )
340 f_bias = f_thresh_min / 4;
341 else if( i_gop_size <= h->param.i_keyint_min )
342 f_bias = f_thresh_min * i_gop_size / h->param.i_keyint_min;
345 f_bias = f_thresh_min
346 + ( f_thresh_max - f_thresh_min )
347 * ( i_gop_size - h->param.i_keyint_min )
348 / ( h->param.i_keyint_max - h->param.i_keyint_min );
351 res = pcost >= (1.0 - f_bias) * icost;
354 int imb = frame->i_intra_mbs[pdist];
355 int pmb = (h->sps->i_mb_width - 2) * (h->sps->i_mb_height - 2) - imb;
356 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",
358 icost, pcost, 1. - (double)pcost / icost,
359 f_bias, i_gop_size, imb, pmb );
364 void x264_slicetype_analyse( x264_t *h )
366 x264_mb_analysis_t a;
367 x264_frame_t *frames[X264_BFRAME_MAX+3] = { NULL, };
371 int i_mb_count = (h->sps->i_mb_width - 2) * (h->sps->i_mb_height - 2);
372 int cost1p0, cost2p0, cost1b1, cost2p1;
375 assert( h->frames.b_have_lowres );
377 if( !h->frames.last_nonb )
379 frames[0] = h->frames.last_nonb;
380 for( j = 0; h->frames.next[j]; j++ )
381 frames[j+1] = h->frames.next[j];
382 keyint_limit = h->param.i_keyint_max - frames[0]->i_frame + h->frames.i_last_idr - 1;
383 num_frames = X264_MIN( j, keyint_limit );
384 if( num_frames == 0 )
387 x264_lowres_context_init( h, &a );
388 idr_frame_type = frames[1]->i_frame - h->frames.i_last_idr >= h->param.i_keyint_min ? X264_TYPE_IDR : X264_TYPE_I;
390 if( num_frames == 1 )
393 frames[1]->i_type = X264_TYPE_P;
394 if( h->param.b_pre_scenecut )
396 x264_slicetype_frame_cost( h, &a, frames, 0, 1, 1, 0 );
397 if( scenecut( h, frames[1], 1 ) )
398 frames[1]->i_type = idr_frame_type;
403 cost2p1 = x264_slicetype_frame_cost( h, &a, frames, 0, 2, 2, 1 );
404 if( frames[2]->i_intra_mbs[2] > i_mb_count / 2 )
407 cost1b1 = x264_slicetype_frame_cost( h, &a, frames, 0, 2, 1, 0 );
408 cost1p0 = x264_slicetype_frame_cost( h, &a, frames, 0, 1, 1, 0 );
409 cost2p0 = x264_slicetype_frame_cost( h, &a, frames, 1, 2, 2, 0 );
410 // fprintf( stderr, "PP: %d + %d <=> BP: %d + %d \n",
411 // cost1p0, cost2p0, cost1b1, cost2p1 );
412 if( cost1p0 + cost2p0 < cost1b1 + cost2p1 )
415 // arbitrary and untuned
416 #define INTER_THRESH 300
417 #define P_SENS_BIAS (50 - h->param.i_bframe_bias)
418 frames[1]->i_type = X264_TYPE_B;
420 for( j = 2; j <= X264_MIN( h->param.i_bframe, num_frames-1 ); j++ )
422 int pthresh = X264_MAX(INTER_THRESH - P_SENS_BIAS * (j-1), INTER_THRESH/10);
423 int pcost = x264_slicetype_frame_cost( h, &a, frames, 0, j+1, j+1, 1 );
424 // fprintf( stderr, "frm%d+%d: %d <=> %d, I:%d/%d \n",
425 // frames[0]->i_frame, j-1, pthresh, pcost/i_mb_count,
426 // frames[j+1]->i_intra_mbs[j+1], i_mb_count );
427 if( pcost > pthresh*i_mb_count || frames[j+1]->i_intra_mbs[j+1] > i_mb_count/3 )
429 frames[j]->i_type = X264_TYPE_P;
433 frames[j]->i_type = X264_TYPE_B;
437 void x264_slicetype_decide( x264_t *h )
443 if( h->frames.next[0] == NULL )
446 if( h->param.rc.b_stat_read )
448 /* Use the frame types from the first pass */
449 for( i = 0; h->frames.next[i] != NULL; i++ )
450 h->frames.next[i]->i_type =
451 x264_ratecontrol_slice_type( h, h->frames.next[i]->i_frame );
453 else if( (h->param.i_bframe && h->param.b_bframe_adaptive)
454 || h->param.b_pre_scenecut )
455 x264_slicetype_analyse( h );
457 for( bframes = 0;; bframes++ )
459 frm = h->frames.next[bframes];
462 if( frm->i_frame - h->frames.i_last_idr >= h->param.i_keyint_max )
464 if( frm->i_type == X264_TYPE_AUTO )
465 frm->i_type = X264_TYPE_IDR;
466 if( frm->i_type != X264_TYPE_IDR )
467 x264_log( h, X264_LOG_WARNING, "specified frame type (%d) is not compatible with keyframe interval\n", frm->i_type );
469 if( frm->i_type == X264_TYPE_IDR )
475 h->frames.next[bframes]->i_type = X264_TYPE_P;
483 if( bframes == h->param.i_bframe
484 || h->frames.next[bframes+1] == NULL )
486 if( IS_X264_TYPE_B( frm->i_type ) )
487 x264_log( h, X264_LOG_WARNING, "specified frame type is not compatible with max B-frames\n" );
488 if( frm->i_type == X264_TYPE_AUTO
489 || IS_X264_TYPE_B( frm->i_type ) )
490 frm->i_type = X264_TYPE_P;
493 if( frm->i_type != X264_TYPE_AUTO && frm->i_type != X264_TYPE_B && frm->i_type != X264_TYPE_BREF )
496 frm->i_type = X264_TYPE_B;
500 int x264_rc_analyse_slice( x264_t *h )
502 x264_mb_analysis_t a;
503 x264_frame_t *frames[X264_BFRAME_MAX+2] = { NULL, };
507 x264_lowres_context_init( h, &a );
509 if( IS_X264_TYPE_I(h->fenc->i_type) )
513 else if( X264_TYPE_P == h->fenc->i_type )
516 while( h->frames.current[p1] && IS_X264_TYPE_B( h->frames.current[p1]->i_type ) )
523 p1 = (h->fref1[0]->i_poc - h->fref0[0]->i_poc)/2;
524 b = (h->fref1[0]->i_poc - h->fenc->i_poc)/2;
525 frames[p1] = h->fref1[0];
527 frames[p0] = h->fref0[0];
530 cost = x264_slicetype_frame_cost( h, &a, frames, p0, p1, b, 0 );
531 h->fenc->i_row_satd = h->fenc->i_row_satds[b-p0][p1-b];
532 h->fdec->i_row_satd = h->fdec->i_row_satds[b-p0][p1-b];
533 h->fdec->i_satd = cost;
534 memcpy( h->fdec->i_row_satd, h->fenc->i_row_satd, h->sps->i_mb_height * sizeof(int) );