1 /*****************************************************************************
2 * slicetype.c: h264 encoder library
3 *****************************************************************************
4 * Copyright (C) 2005-2008 Loren Merritt <lorenm@u.washington.edu>
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License as published by
8 * the Free Software Foundation; either version 2 of the License, or
9 * (at your option) any later version.
11 * This program is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
16 * You should have received a copy of the GNU General Public License
17 * along with this program; if not, write to the Free Software
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02111, USA.
19 *****************************************************************************/
24 #include "common/common.h"
25 #include "common/cpu.h"
26 #include "macroblock.h"
30 static void x264_lowres_context_init( x264_t *h, x264_mb_analysis_t *a )
32 a->i_qp = 12; // arbitrary, but low because SATD scores are 1/4 normal
33 a->i_lambda = x264_lambda_tab[ a->i_qp ];
34 x264_mb_analyse_load_costs( h, a );
35 h->mb.i_me_method = X264_MIN( X264_ME_HEX, h->param.analyse.i_me_method ); // maybe dia?
36 h->mb.i_subpel_refine = 4; // 3 should be enough, but not tweaking for speed now
37 h->mb.b_chroma_me = 0;
40 static int x264_slicetype_mb_cost( x264_t *h, x264_mb_analysis_t *a,
41 x264_frame_t **frames, int p0, int p1, int b,
42 int dist_scale_factor )
44 x264_frame_t *fref0 = frames[p0];
45 x264_frame_t *fref1 = frames[p1];
46 x264_frame_t *fenc = frames[b];
47 const int b_bidir = (b < p1);
48 const int i_mb_x = h->mb.i_mb_x;
49 const int i_mb_y = h->mb.i_mb_y;
50 const int i_mb_stride = h->sps->i_mb_width;
51 const int i_mb_xy = i_mb_x + i_mb_y * i_mb_stride;
52 const int i_stride = fenc->i_stride_lowres;
53 const int i_pel_offset = 8 * ( i_mb_x + i_mb_y * i_stride );
55 DECLARE_ALIGNED_8( uint8_t pix1[9*FDEC_STRIDE] );
56 uint8_t *pix2 = pix1+8;
58 int i_bcost = COST_MAX;
62 h->mb.pic.p_fenc[0] = h->mb.pic.fenc_buf;
63 h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fenc[0], FENC_STRIDE, &fenc->lowres[0][i_pel_offset], i_stride, 8 );
65 if( !p0 && !p1 && !b )
68 // no need for h->mb.mv_min[]
69 h->mb.mv_min_fpel[0] = -8*h->mb.i_mb_x - 4;
70 h->mb.mv_max_fpel[0] = 8*( h->sps->i_mb_width - h->mb.i_mb_x - 1 ) + 4;
71 h->mb.mv_min_spel[0] = 4*( h->mb.mv_min_fpel[0] - 8 );
72 h->mb.mv_max_spel[0] = 4*( h->mb.mv_max_fpel[0] + 8 );
73 if( h->mb.i_mb_x <= 1 )
75 h->mb.mv_min_fpel[1] = -8*h->mb.i_mb_y - 4;
76 h->mb.mv_max_fpel[1] = 8*( h->sps->i_mb_height - h->mb.i_mb_y - 1 ) + 4;
77 h->mb.mv_min_spel[1] = 4*( h->mb.mv_min_fpel[1] - 8 );
78 h->mb.mv_max_spel[1] = 4*( h->mb.mv_max_fpel[1] + 8 );
81 #define LOAD_HPELS_LUMA(dst, src) \
83 (dst)[0] = &(src)[0][i_pel_offset]; \
84 (dst)[1] = &(src)[1][i_pel_offset]; \
85 (dst)[2] = &(src)[2][i_pel_offset]; \
86 (dst)[3] = &(src)[3][i_pel_offset]; \
88 #define SAVE_MVS( mv0, mv1 ) \
90 *(uint32_t*)fenc->mv[0][i_mb_xy] = *(uint32_t*)mv0; \
92 *(uint32_t*)fenc->mv[1][i_mb_xy] = *(uint32_t*)mv1; \
94 #define CLIP_MV( mv ) \
96 mv[0] = x264_clip3( mv[0], h->mb.mv_min_spel[0], h->mb.mv_max_spel[0] ); \
97 mv[1] = x264_clip3( mv[1], h->mb.mv_min_spel[1], h->mb.mv_max_spel[1] ); \
99 #define TRY_BIDIR( mv0, mv1, penalty ) \
104 h->mc.mc_luma( pix1, 16, m[0].p_fref, m[0].i_stride[0], \
105 (mv0)[0], (mv0)[1], 8, 8 ); \
106 src2 = h->mc.get_ref( pix2, &stride2, m[1].p_fref, m[1].i_stride[0], \
107 (mv1)[0], (mv1)[1], 8, 8 ); \
108 h->mc.avg[PIXEL_8x8]( pix1, 16, src2, stride2 ); \
109 i_cost = penalty + h->pixf.mbcmp[PIXEL_8x8]( \
110 m[0].p_fenc[0], FENC_STRIDE, pix1, 16 ); \
111 if( i_bcost > i_cost ) \
114 SAVE_MVS( mv0, mv1 ); \
118 m[0].i_pixel = PIXEL_8x8;
119 m[0].p_cost_mv = a->p_cost_mv;
120 m[0].i_stride[0] = i_stride;
121 m[0].p_fenc[0] = h->mb.pic.p_fenc[0];
122 LOAD_HPELS_LUMA( m[0].p_fref, fref0->lowres );
126 int16_t *mvr = fref1->mv[0][i_mb_xy];
130 h->mc.memcpy_aligned( &m[1], &m[0], sizeof(x264_me_t) );
131 LOAD_HPELS_LUMA( m[1].p_fref, fref1->lowres );
133 dmv[0][0] = ( mvr[0] * dist_scale_factor + 128 ) >> 8;
134 dmv[0][1] = ( mvr[1] * dist_scale_factor + 128 ) >> 8;
135 dmv[1][0] = dmv[0][0] - mvr[0];
136 dmv[1][1] = dmv[0][1] - mvr[1];
140 TRY_BIDIR( dmv[0], dmv[1], 0 );
141 if( dmv[0][0] | dmv[0][1] | dmv[1][0] | dmv[1][1] )
142 TRY_BIDIR( mv0, mv0, 0 );
143 // if( i_bcost < 60 ) // arbitrary threshold
147 i_cost_bak = i_bcost;
148 for( l = 0; l < 1 + b_bidir; l++ )
150 DECLARE_ALIGNED_4(int16_t mvc[4][2]) = {{0}};
152 int16_t (*fenc_mv)[2] = &fenc->mv[l][i_mb_xy];
153 #define MVC(mv) { *(uint32_t*)mvc[i_mvc] = *(uint32_t*)mv; i_mvc++; }
158 MVC(fenc_mv[-i_mb_stride]);
159 if( i_mb_x < h->sps->i_mb_width - 1 )
160 MVC(fenc_mv[-i_mb_stride+1]);
162 MVC(fenc_mv[-i_mb_stride-1]);
165 x264_median_mv( m[l].mvp, mvc[0], mvc[1], mvc[2] );
166 x264_me_search( h, &m[l], mvc, i_mvc );
168 m[l].cost -= 2; // remove mvcost from skip mbs
169 if( *(uint32_t*)m[l].mv )
171 i_bcost = X264_MIN( i_bcost, m[l].cost );
174 if( b_bidir && ( *(uint32_t*)m[0].mv || *(uint32_t*)m[1].mv ) )
175 TRY_BIDIR( m[0].mv, m[1].mv, 5 );
177 if( i_bcost < i_cost_bak )
178 SAVE_MVS( m[0].mv, m[1].mv );
180 //FIXME intra part could be shared across multiple encodings of the frame
182 if( !b_bidir ) // forbid intra-mbs in B-frames, because it's rare and not worth checking
184 uint8_t *pix = &pix1[8+FDEC_STRIDE - 1];
185 uint8_t *src = &fenc->lowres[0][i_pel_offset - 1];
186 const int intra_penalty = 5;
187 int satds[4], i_icost, b_intra;
189 memcpy( pix-FDEC_STRIDE, src-i_stride, 17 );
191 pix[i*FDEC_STRIDE] = src[i*i_stride];
194 if( h->pixf.intra_satd_x3_8x8c && h->pixf.mbcmp[0] == h->pixf.satd[0] )
196 h->pixf.intra_satd_x3_8x8c( h->mb.pic.p_fenc[0], pix, satds );
197 h->predict_8x8c[I_PRED_CHROMA_P]( pix );
198 satds[I_PRED_CHROMA_P] =
199 h->pixf.satd[PIXEL_8x8]( pix, FDEC_STRIDE, h->mb.pic.p_fenc[0], FENC_STRIDE );
205 h->predict_8x8c[i]( pix );
206 satds[i] = h->pixf.mbcmp[PIXEL_8x8]( pix, FDEC_STRIDE, h->mb.pic.p_fenc[0], FENC_STRIDE );
209 i_icost = X264_MIN4( satds[0], satds[1], satds[2], satds[3] );
211 if( i_icost < i_bcost * 2 )
213 DECLARE_ALIGNED_16( uint8_t edge[33] );
214 x264_predict_8x8_filter( pix, edge, ALL_NEIGHBORS, ALL_NEIGHBORS );
218 h->predict_8x8[i]( pix, edge );
219 satd = h->pixf.mbcmp[PIXEL_8x8]( pix, FDEC_STRIDE, h->mb.pic.p_fenc[0], FENC_STRIDE );
220 i_icost = X264_MIN( i_icost, satd );
224 i_icost += intra_penalty;
225 b_intra = i_icost < i_bcost;
228 if( i_mb_x > 0 && i_mb_x < h->sps->i_mb_width - 1
229 && i_mb_y > 0 && i_mb_y < h->sps->i_mb_height - 1 )
231 fenc->i_intra_mbs[b-p0] += b_intra;
232 fenc->i_cost_est[0][0] += i_icost;
242 (h->sps->i_mb_width > 2 && h->sps->i_mb_height > 2 ?\
243 (h->sps->i_mb_width - 2) * (h->sps->i_mb_height - 2) :\
244 h->sps->i_mb_width * h->sps->i_mb_height)
246 static int x264_slicetype_frame_cost( x264_t *h, x264_mb_analysis_t *a,
247 x264_frame_t **frames, int p0, int p1, int b,
248 int b_intra_penalty )
251 /* Don't use the AQ'd scores for slicetype decision. */
254 /* Check whether we already evaluated this frame
255 * If we have tried this frame as P, then we have also tried
256 * the preceding frames as B. (is this still true?) */
257 /* Also check that we already calculated the row SATDs for the current frame. */
258 if( frames[b]->i_cost_est[b-p0][p1-b] >= 0 && (!h->param.rc.i_vbv_buffer_size || frames[b]->i_row_satds[b-p0][p1-b][0] != -1) )
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 if( h->sps->i_mb_width <= 2 || h->sps->i_mb_height <= 2 )
283 for( h->mb.i_mb_y = 0; h->mb.i_mb_y < h->sps->i_mb_height; h->mb.i_mb_y++ )
284 for( h->mb.i_mb_x = 0; h->mb.i_mb_x < h->sps->i_mb_width; h->mb.i_mb_x++ )
285 i_score += x264_slicetype_mb_cost( h, a, frames, p0, p1, b, dist_scale_factor );
287 /* the edge mbs seem to reduce the predictive quality of the
288 * whole frame's score, but are needed for a spatial distribution. */
289 else if( h->param.rc.i_vbv_buffer_size )
291 for( h->mb.i_mb_y = 0; h->mb.i_mb_y < h->sps->i_mb_height; h->mb.i_mb_y++ )
293 row_satd[ h->mb.i_mb_y ] = 0;
294 for( h->mb.i_mb_x = 0; h->mb.i_mb_x < h->sps->i_mb_width; h->mb.i_mb_x++ )
296 int i_mb_cost = x264_slicetype_mb_cost( h, a, frames, p0, p1, b, dist_scale_factor );
297 int i_mb_cost_aq = i_mb_cost;
298 if( h->param.rc.i_aq_mode )
301 i_mb_cost_aq *= pow(2.0,-(frames[b]->f_qp_offset[h->mb.i_mb_x + h->mb.i_mb_y*h->mb.i_mb_stride])/6.0);
303 row_satd[ h->mb.i_mb_y ] += i_mb_cost_aq;
304 if( h->mb.i_mb_y > 0 && h->mb.i_mb_y < h->sps->i_mb_height - 1 &&
305 h->mb.i_mb_x > 0 && h->mb.i_mb_x < h->sps->i_mb_width - 1 )
307 /* Don't use AQ-weighted costs for slicetype decision, only for ratecontrol. */
308 i_score += i_mb_cost;
309 i_score_aq += i_mb_cost_aq;
316 for( h->mb.i_mb_y = 1; h->mb.i_mb_y < h->sps->i_mb_height - 1; h->mb.i_mb_y++ )
317 for( h->mb.i_mb_x = 1; h->mb.i_mb_x < h->sps->i_mb_width - 1; h->mb.i_mb_x++ )
319 int i_mb_cost = x264_slicetype_mb_cost( h, a, frames, p0, p1, b, dist_scale_factor );
320 int i_mb_cost_aq = i_mb_cost;
321 if( h->param.rc.i_aq_mode )
324 i_mb_cost_aq *= pow(2.0,-(frames[b]->f_qp_offset[h->mb.i_mb_x + h->mb.i_mb_y*h->mb.i_mb_stride])/6.0);
326 i_score += i_mb_cost;
327 i_score_aq += i_mb_cost_aq;
332 i_score = i_score * 100 / (120 + h->param.i_bframe_bias);
334 frames[b]->i_cost_est[b-p0][p1-b] = i_score;
335 frames[b]->i_cost_est_aq[b-p0][p1-b] = i_score_aq;
336 // fprintf( stderr, "frm %d %c(%d,%d): %6d %6d imb:%d \n", frames[b]->i_frame,
337 // (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] );
341 if( b_intra_penalty )
343 // arbitrary penalty for I-blocks after B-frames
345 i_score += i_score * frames[b]->i_intra_mbs[b-p0] / (nmb * 8);
350 static int scenecut( x264_t *h, x264_frame_t *frame, int pdist )
352 int icost = frame->i_cost_est[0][0];
353 int pcost = frame->i_cost_est[pdist][0];
355 int i_gop_size = frame->i_frame - h->frames.i_last_idr;
356 float f_thresh_max = h->param.i_scenecut_threshold / 100.0;
357 /* magic numbers pulled out of thin air */
358 float f_thresh_min = f_thresh_max * h->param.i_keyint_min
359 / ( h->param.i_keyint_max * 4 );
362 if( h->param.i_keyint_min == h->param.i_keyint_max )
363 f_thresh_min= f_thresh_max;
364 if( i_gop_size < h->param.i_keyint_min / 4 )
365 f_bias = f_thresh_min / 4;
366 else if( i_gop_size <= h->param.i_keyint_min )
367 f_bias = f_thresh_min * i_gop_size / h->param.i_keyint_min;
370 f_bias = f_thresh_min
371 + ( f_thresh_max - f_thresh_min )
372 * ( i_gop_size - h->param.i_keyint_min )
373 / ( h->param.i_keyint_max - h->param.i_keyint_min );
376 res = pcost >= (1.0 - f_bias) * icost;
379 int imb = frame->i_intra_mbs[pdist];
380 int pmb = NUM_MBS - imb;
381 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",
383 icost, pcost, 1. - (double)pcost / icost,
384 f_bias, i_gop_size, imb, pmb );
389 static void x264_slicetype_analyse( x264_t *h )
391 x264_mb_analysis_t a;
392 x264_frame_t *frames[X264_BFRAME_MAX+3] = { NULL, };
396 int i_mb_count = NUM_MBS;
397 int cost1p0, cost2p0, cost1b1, cost2p1;
400 assert( h->frames.b_have_lowres );
402 if( !h->frames.last_nonb )
404 frames[0] = h->frames.last_nonb;
405 for( j = 0; h->frames.next[j]; j++ )
406 frames[j+1] = h->frames.next[j];
407 keyint_limit = h->param.i_keyint_max - frames[0]->i_frame + h->frames.i_last_idr - 1;
408 num_frames = X264_MIN( j, keyint_limit );
409 if( num_frames == 0 )
412 x264_lowres_context_init( h, &a );
413 idr_frame_type = frames[1]->i_frame - h->frames.i_last_idr >= h->param.i_keyint_min ? X264_TYPE_IDR : X264_TYPE_I;
415 if( num_frames == 1 )
418 frames[1]->i_type = X264_TYPE_P;
419 if( h->param.b_pre_scenecut )
421 x264_slicetype_frame_cost( h, &a, frames, 0, 1, 1, 0 );
422 if( scenecut( h, frames[1], 1 ) )
423 frames[1]->i_type = idr_frame_type;
428 cost2p1 = x264_slicetype_frame_cost( h, &a, frames, 0, 2, 2, 1 );
429 if( frames[2]->i_intra_mbs[2] > i_mb_count / 2 )
432 cost1b1 = x264_slicetype_frame_cost( h, &a, frames, 0, 2, 1, 0 );
433 cost1p0 = x264_slicetype_frame_cost( h, &a, frames, 0, 1, 1, 0 );
434 cost2p0 = x264_slicetype_frame_cost( h, &a, frames, 1, 2, 2, 0 );
435 // fprintf( stderr, "PP: %d + %d <=> BP: %d + %d \n",
436 // cost1p0, cost2p0, cost1b1, cost2p1 );
437 if( cost1p0 + cost2p0 < cost1b1 + cost2p1 )
440 // arbitrary and untuned
441 #define INTER_THRESH 300
442 #define P_SENS_BIAS (50 - h->param.i_bframe_bias)
443 frames[1]->i_type = X264_TYPE_B;
445 for( j = 2; j <= X264_MIN( h->param.i_bframe, num_frames-1 ); j++ )
447 int pthresh = X264_MAX(INTER_THRESH - P_SENS_BIAS * (j-1), INTER_THRESH/10);
448 int pcost = x264_slicetype_frame_cost( h, &a, frames, 0, j+1, j+1, 1 );
450 fprintf( stderr, "frm%d+%d: %d <=> %d, I:%d/%d \n",
451 frames[0]->i_frame, j-1, pthresh, pcost/i_mb_count,
452 frames[j+1]->i_intra_mbs[j+1], i_mb_count );
454 fprintf( stderr, "frm%d+%d: %d <=> %d, I:%d/%d \n",
455 frames[0]->i_frame, j-1, pthresh, pcost,
456 frames[j+1]->i_intra_mbs[j+1], i_mb_count ); */
457 if( pcost > pthresh*i_mb_count || frames[j+1]->i_intra_mbs[j+1] > i_mb_count/3 )
459 frames[j]->i_type = X264_TYPE_P;
463 frames[j]->i_type = X264_TYPE_B;
467 void x264_slicetype_decide( x264_t *h )
473 if( h->frames.next[0] == NULL )
476 if( h->param.rc.b_stat_read )
478 /* Use the frame types from the first pass */
479 for( i = 0; h->frames.next[i] != NULL; i++ )
480 h->frames.next[i]->i_type =
481 x264_ratecontrol_slice_type( h, h->frames.next[i]->i_frame );
483 else if( (h->param.i_bframe && h->param.b_bframe_adaptive)
484 || h->param.b_pre_scenecut )
485 x264_slicetype_analyse( h );
487 for( bframes = 0;; bframes++ )
489 frm = h->frames.next[bframes];
492 if( frm->i_frame - h->frames.i_last_idr >= h->param.i_keyint_max )
494 if( frm->i_type == X264_TYPE_AUTO )
495 frm->i_type = X264_TYPE_IDR;
496 if( frm->i_type != X264_TYPE_IDR )
497 x264_log( h, X264_LOG_WARNING, "specified frame type (%d) is not compatible with keyframe interval\n", frm->i_type );
499 if( frm->i_type == X264_TYPE_IDR )
505 h->frames.next[bframes]->i_type = X264_TYPE_P;
513 if( bframes == h->param.i_bframe
514 || h->frames.next[bframes+1] == NULL )
516 if( IS_X264_TYPE_B( frm->i_type ) )
517 x264_log( h, X264_LOG_WARNING, "specified frame type is not compatible with max B-frames\n" );
518 if( frm->i_type == X264_TYPE_AUTO
519 || IS_X264_TYPE_B( frm->i_type ) )
520 frm->i_type = X264_TYPE_P;
523 if( frm->i_type != X264_TYPE_AUTO && frm->i_type != X264_TYPE_B && frm->i_type != X264_TYPE_BREF )
526 frm->i_type = X264_TYPE_B;
530 int x264_rc_analyse_slice( x264_t *h )
532 x264_mb_analysis_t a;
533 x264_frame_t *frames[X264_BFRAME_MAX+2] = { NULL, };
537 x264_lowres_context_init( h, &a );
539 if( IS_X264_TYPE_I(h->fenc->i_type) )
543 else if( X264_TYPE_P == h->fenc->i_type )
546 while( h->frames.current[p1] && IS_X264_TYPE_B( h->frames.current[p1]->i_type ) )
553 p1 = (h->fref1[0]->i_poc - h->fref0[0]->i_poc)/2;
554 b = (h->fref1[0]->i_poc - h->fenc->i_poc)/2;
555 frames[p1] = h->fref1[0];
557 frames[p0] = h->fref0[0];
560 cost = x264_slicetype_frame_cost( h, &a, frames, p0, p1, b, 0 );
562 /* In AQ, use the weighted score instead. */
563 if( h->param.rc.i_aq_mode )
564 cost = frames[b]->i_cost_est[b-p0][p1-b];
566 h->fenc->i_row_satd = h->fenc->i_row_satds[b-p0][p1-b];
567 h->fdec->i_row_satd = h->fdec->i_row_satds[b-p0][p1-b];
568 h->fdec->i_satd = cost;
569 memcpy( h->fdec->i_row_satd, h->fenc->i_row_satd, h->sps->i_mb_height * sizeof(int) );