1 /*****************************************************************************
2 * slicetype_decision.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 *****************************************************************************/
27 #include "common/common.h"
28 #include "common/cpu.h"
29 #include "macroblock.h"
33 static void x264_lowres_context_init( x264_t *h, x264_mb_analysis_t *a )
35 a->i_qp = 12; // arbitrary, but low because SATD scores are 1/4 normal
36 a->i_lambda = i_qp0_cost_table[ a->i_qp ];
37 x264_mb_analyse_load_costs( h, a );
38 h->mb.i_me_method = X264_MIN( X264_ME_HEX, h->param.analyse.i_me_method ); // maybe dia?
39 h->mb.i_subpel_refine = 4; // 3 should be enough, but not tweaking for speed now
40 h->mb.b_chroma_me = 0;
43 int x264_slicetype_mb_cost( x264_t *h, x264_mb_analysis_t *a,
44 x264_frame_t **frames, int p0, int p1, int b,
45 int dist_scale_factor )
47 x264_frame_t *fref0 = frames[p0];
48 x264_frame_t *fref1 = frames[p1];
49 x264_frame_t *fenc = frames[b];
50 const int b_bidir = (b < p1);
51 const int i_mb_x = h->mb.i_mb_x;
52 const int i_mb_y = h->mb.i_mb_y;
53 const int i_mb_stride = h->sps->i_mb_width;
54 const int i_mb_xy = i_mb_x + i_mb_y * i_mb_stride;
55 const int i_stride = fenc->i_stride_lowres;
56 const int i_pel_offset = 8 * ( i_mb_x + i_mb_y * i_stride );
58 DECLARE_ALIGNED( uint8_t, pix1[9*FDEC_STRIDE], 8 );
59 uint8_t *pix2 = pix1+8;
61 int i_bcost = COST_MAX;
65 h->mb.pic.p_fenc[0] = h->mb.pic.fenc_buf;
66 h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fenc[0], FENC_STRIDE, &fenc->lowres[0][i_pel_offset], i_stride, 8 );
68 if( !p0 && !p1 && !b )
71 // no need for h->mb.mv_min[]
72 h->mb.mv_min_fpel[0] = -8*h->mb.i_mb_x - 4;
73 h->mb.mv_max_fpel[0] = 8*( h->sps->i_mb_width - h->mb.i_mb_x - 1 ) + 4;
74 h->mb.mv_min_spel[0] = 4*( h->mb.mv_min_fpel[0] - 8 );
75 h->mb.mv_max_spel[0] = 4*( h->mb.mv_max_fpel[0] + 8 );
76 if( h->mb.i_mb_x <= 1 )
78 h->mb.mv_min_fpel[1] = -8*h->mb.i_mb_y - 4;
79 h->mb.mv_max_fpel[1] = 8*( h->sps->i_mb_height - h->mb.i_mb_y - 1 ) + 4;
80 h->mb.mv_min_spel[1] = 4*( h->mb.mv_min_fpel[1] - 8 );
81 h->mb.mv_max_spel[1] = 4*( h->mb.mv_max_fpel[1] + 8 );
84 #define LOAD_HPELS_LUMA(dst, src) \
86 (dst)[0] = &(src)[0][i_pel_offset]; \
87 (dst)[1] = &(src)[1][i_pel_offset]; \
88 (dst)[2] = &(src)[2][i_pel_offset]; \
89 (dst)[3] = &(src)[3][i_pel_offset]; \
91 #define SAVE_MVS( mv0, mv1 ) \
93 fenc->mv[0][i_mb_xy][0] = mv0[0]; \
94 fenc->mv[0][i_mb_xy][1] = mv0[1]; \
97 fenc->mv[1][i_mb_xy][0] = mv1[0]; \
98 fenc->mv[1][i_mb_xy][1] = mv1[1]; \
101 #define CLIP_MV( mv ) \
103 mv[0] = x264_clip3( mv[0], h->mb.mv_min_spel[0], h->mb.mv_max_spel[0] ); \
104 mv[1] = x264_clip3( mv[1], h->mb.mv_min_spel[1], h->mb.mv_max_spel[1] ); \
106 #define TRY_BIDIR( mv0, mv1, penalty ) \
111 h->mc.mc_luma( m[0].p_fref, m[0].i_stride[0], pix1, 16, \
112 (mv0)[0], (mv0)[1], 8, 8 ); \
113 src2 = h->mc.get_ref( m[1].p_fref, m[1].i_stride[0], pix2, &stride2, \
114 (mv1)[0], (mv1)[1], 8, 8 ); \
115 h->mc.avg[PIXEL_8x8]( pix1, 16, src2, stride2 ); \
116 i_cost = penalty + h->pixf.mbcmp[PIXEL_8x8]( \
117 m[0].p_fenc[0], FENC_STRIDE, pix1, 16 ); \
118 if( i_bcost > i_cost ) \
121 SAVE_MVS( mv0, mv1 ); \
125 m[0].i_pixel = PIXEL_8x8;
126 m[0].p_cost_mv = a->p_cost_mv;
127 m[0].i_stride[0] = i_stride;
128 m[0].p_fenc[0] = h->mb.pic.p_fenc[0];
129 LOAD_HPELS_LUMA( m[0].p_fref, fref0->lowres );
133 int16_t *mvr = fref1->mv[0][i_mb_xy];
138 LOAD_HPELS_LUMA( m[1].p_fref, fref1->lowres );
140 dmv[0][0] = ( mvr[0] * dist_scale_factor + 128 ) >> 8;
141 dmv[0][1] = ( mvr[1] * dist_scale_factor + 128 ) >> 8;
142 dmv[1][0] = dmv[0][0] - mvr[0];
143 dmv[1][1] = dmv[0][1] - mvr[1];
147 TRY_BIDIR( dmv[0], dmv[1], 0 );
148 if( dmv[0][0] || dmv[0][1] || dmv[1][0] || dmv[1][1] )
149 TRY_BIDIR( mv0, mv0, 0 );
150 // if( i_bcost < 60 ) // arbitrary threshold
154 i_cost_bak = i_bcost;
155 for( l = 0; l < 1 + b_bidir; l++ )
157 int mvc[4][2] = {{0}}, i_mvc;
158 int16_t (*fenc_mv)[2] = &fenc->mv[l][i_mb_xy];
162 mvc[i_mvc][0] = fenc_mv[-1][0];
163 mvc[i_mvc][1] = fenc_mv[-1][1];
168 mvc[i_mvc][0] = fenc_mv[-i_mb_stride][0];
169 mvc[i_mvc][1] = fenc_mv[-i_mb_stride][1];
171 if( i_mb_x < h->sps->i_mb_width - 1 )
173 mvc[i_mvc][0] = fenc_mv[-i_mb_stride+1][0];
174 mvc[i_mvc][1] = fenc_mv[-i_mb_stride+1][1];
179 mvc[i_mvc][0] = fenc_mv[-i_mb_stride-1][0];
180 mvc[i_mvc][1] = fenc_mv[-i_mb_stride-1][1];
184 m[l].mvp[0] = x264_median( mvc[0][0], mvc[1][0], mvc[2][0] );
185 m[l].mvp[1] = x264_median( mvc[0][1], mvc[1][1], mvc[2][1] );
187 x264_me_search( h, &m[l], mvc, i_mvc );
189 i_bcost = X264_MIN( i_bcost, m[l].cost + 3 );
192 if( b_bidir && (m[0].mv[0] || m[0].mv[1] || m[1].mv[0] || m[1].mv[1]) )
193 TRY_BIDIR( m[0].mv, m[1].mv, 5 );
195 if( i_bcost < i_cost_bak )
196 SAVE_MVS( m[0].mv, m[1].mv );
200 uint8_t *pix = &pix1[8+FDEC_STRIDE - 1];
201 uint8_t *src = &fenc->lowres[0][i_pel_offset - 1];
202 int intra_penalty = 5 + 10 * b_bidir;
203 int satds[4], i_icost;
205 memcpy( pix-FDEC_STRIDE, src-i_stride, 9 );
207 pix[i*FDEC_STRIDE] = src[i*i_stride];
210 if( h->pixf.intra_satd_x3_8x8c && h->pixf.mbcmp[0] == h->pixf.satd[0] )
212 h->pixf.intra_satd_x3_8x8c( h->mb.pic.p_fenc[0], pix, satds );
213 h->predict_8x8c[I_PRED_CHROMA_P]( pix );
214 satds[I_PRED_CHROMA_P] =
215 h->pixf.satd[PIXEL_8x8]( pix, FDEC_STRIDE, h->mb.pic.p_fenc[0], FENC_STRIDE );
221 h->predict_8x8c[i]( pix );
222 satds[i] = h->pixf.mbcmp[PIXEL_8x8]( pix, FDEC_STRIDE, h->mb.pic.p_fenc[0], FENC_STRIDE );
225 i_icost = X264_MIN4( satds[0], satds[1], satds[2], satds[3] ) + intra_penalty;
226 if( i_icost < i_bcost )
230 && i_mb_x > 0 && i_mb_x < h->sps->i_mb_width - 1
231 && i_mb_y > 0 && i_mb_y < h->sps->i_mb_height - 1 )
233 fenc->i_intra_mbs[b-p0]++;
236 i_bcost = i_bcost * 9 / 8; // arbitrary penalty for I-blocks in and after B-frames
245 int x264_slicetype_frame_cost( x264_t *h, x264_mb_analysis_t *a,
246 x264_frame_t **frames, int p0, int p1, int b )
249 int dist_scale_factor = 128;
250 int *row_satd = frames[b]->i_row_satds[b-p0][p1-b];
252 /* Check whether we already evaluated this frame
253 * If we have tried this frame as P, then we have also tried
254 * the preceding frames as B. (is this still true?) */
255 if( frames[b]->i_cost_est[b-p0][p1-b] >= 0 )
256 return frames[b]->i_cost_est[b-p0][p1-b];
258 /* Init MVs so that we don't have to check edge conditions when loading predictors. */
259 /* FIXME: not needed every time */
260 memset( frames[b]->mv[0], 0, h->sps->i_mb_height * h->sps->i_mb_width * 2*sizeof(int16_t) );
262 memset( frames[b]->mv[1], 0, h->sps->i_mb_height * h->sps->i_mb_width * 2*sizeof(int16_t) );
265 frames[b]->i_intra_mbs[b-p0] = 0;
267 dist_scale_factor = ( ((b-p0) << 8) + ((p1-p0) >> 1) ) / (p1-p0);
269 /* the edge mbs seem to reduce the predictive quality of the
270 * whole frame's score, but are needed for a spatial distribution. */
271 if( h->param.rc.i_vbv_buffer_size )
273 for( h->mb.i_mb_y = 0; h->mb.i_mb_y < h->sps->i_mb_height; h->mb.i_mb_y++ )
275 row_satd[ h->mb.i_mb_y ] = 0;
276 for( h->mb.i_mb_x = 0; h->mb.i_mb_x < h->sps->i_mb_width; h->mb.i_mb_x++ )
278 int i_mb_cost = x264_slicetype_mb_cost( h, a, frames, p0, p1, b, dist_scale_factor );
279 row_satd[ h->mb.i_mb_y ] += i_mb_cost;
280 if( h->mb.i_mb_y > 0 && h->mb.i_mb_y < h->sps->i_mb_height - 1 &&
281 h->mb.i_mb_x > 0 && h->mb.i_mb_x < h->sps->i_mb_width - 1 )
283 i_score += i_mb_cost;
290 for( h->mb.i_mb_y = 1; h->mb.i_mb_y < h->sps->i_mb_height - 1; h->mb.i_mb_y++ )
291 for( h->mb.i_mb_x = 1; h->mb.i_mb_x < h->sps->i_mb_width - 1; h->mb.i_mb_x++ )
292 i_score += x264_slicetype_mb_cost( h, a, frames, p0, p1, b, dist_scale_factor );
296 i_score = i_score * 100 / (120 + h->param.i_bframe_bias);
298 frames[b]->i_cost_est[b-p0][p1-b] = i_score;
299 // fprintf( stderr, "frm %d %c(%d,%d): %6d I:%d \n", frames[b]->i_frame,
300 // (p1==0?'I':b<p1?'B':'P'), b-p0, p1-b, i_score, frames[b]->i_intra_mbs[b-p0] );
301 x264_cpu_restore( h->param.cpu );
305 void x264_slicetype_analyse( x264_t *h )
307 x264_mb_analysis_t a;
308 x264_frame_t *frames[X264_BFRAME_MAX+3] = { NULL, };
312 int i_mb_count = (h->sps->i_mb_width - 2) * (h->sps->i_mb_height - 2);
313 int cost1p0, cost2p0, cost1b1, cost2p1;
315 if( !h->frames.last_nonb )
317 frames[0] = h->frames.last_nonb;
318 for( j = 0; h->frames.next[j]; j++ )
319 frames[j+1] = h->frames.next[j];
320 keyint_limit = h->param.i_keyint_max - frames[0]->i_frame + h->frames.i_last_idr - 1;
321 num_frames = X264_MIN( j, keyint_limit );
322 if( num_frames == 0 )
324 if( num_frames == 1 )
327 frames[1]->i_type = X264_TYPE_P;
331 x264_lowres_context_init( h, &a );
333 cost2p1 = x264_slicetype_frame_cost( h, &a, frames, 0, 2, 2 );
334 if( frames[2]->i_intra_mbs[2] > i_mb_count / 2 )
337 cost1b1 = x264_slicetype_frame_cost( h, &a, frames, 0, 2, 1 );
338 cost1p0 = x264_slicetype_frame_cost( h, &a, frames, 0, 1, 1 );
339 cost2p0 = x264_slicetype_frame_cost( h, &a, frames, 1, 2, 2 );
340 // fprintf( stderr, "PP: %d + %d <=> BP: %d + %d \n",
341 // cost1p0, cost2p0, cost1b1, cost2p1 );
342 if( cost1p0 + cost2p0 < cost1b1 + cost2p1 )
345 // arbitrary and untuned
346 #define INTER_THRESH 300
347 #define P_SENS_BIAS (50 - h->param.i_bframe_bias)
348 frames[1]->i_type = X264_TYPE_B;
350 for( j = 2; j <= X264_MIN( h->param.i_bframe, num_frames-1 ); j++ )
352 int pthresh = X264_MAX(INTER_THRESH - P_SENS_BIAS * (j-1), INTER_THRESH/10);
353 int pcost = x264_slicetype_frame_cost( h, &a, frames, 0, j+1, j+1 );
354 // fprintf( stderr, "frm%d+%d: %d <=> %d, I:%d/%d \n",
355 // frames[0]->i_frame, j-1, pthresh, pcost/i_mb_count,
356 // frames[j+1]->i_intra_mbs[j+1], i_mb_count );
357 if( pcost > pthresh*i_mb_count || frames[j+1]->i_intra_mbs[j+1] > i_mb_count/3 )
359 frames[j]->i_type = X264_TYPE_P;
363 frames[j]->i_type = X264_TYPE_B;
367 void x264_slicetype_decide( x264_t *h )
373 if( h->frames.next[0] == NULL )
376 if( h->param.rc.b_stat_read )
378 /* Use the frame types from the first pass */
379 for( i = 0; h->frames.next[i] != NULL; i++ )
380 h->frames.next[i]->i_type =
381 x264_ratecontrol_slice_type( h, h->frames.next[i]->i_frame );
383 else if( h->param.i_bframe && h->param.b_bframe_adaptive )
384 x264_slicetype_analyse( h );
386 for( bframes = 0;; bframes++ )
388 frm = h->frames.next[bframes];
391 if( frm->i_frame - h->frames.i_last_idr >= h->param.i_keyint_max )
393 if( frm->i_type == X264_TYPE_AUTO )
394 frm->i_type = X264_TYPE_IDR;
395 if( frm->i_type != X264_TYPE_IDR )
396 x264_log( h, X264_LOG_WARNING, "specified frame type (%d) is not compatible with keyframe interval\n", frm->i_type );
398 if( frm->i_type == X264_TYPE_IDR )
404 h->frames.next[bframes]->i_type = X264_TYPE_P;
412 if( bframes == h->param.i_bframe
413 || h->frames.next[bframes+1] == NULL )
415 if( IS_X264_TYPE_B( frm->i_type ) )
416 x264_log( h, X264_LOG_WARNING, "specified frame type is not compatible with max B-frames\n" );
417 if( frm->i_type == X264_TYPE_AUTO
418 || IS_X264_TYPE_B( frm->i_type ) )
419 frm->i_type = X264_TYPE_P;
422 if( frm->i_type != X264_TYPE_AUTO && frm->i_type != X264_TYPE_B && frm->i_type != X264_TYPE_BREF )
425 frm->i_type = X264_TYPE_B;
429 int x264_rc_analyse_slice( x264_t *h )
431 x264_mb_analysis_t a;
432 x264_frame_t *frames[X264_BFRAME_MAX+2] = { NULL, };
436 x264_lowres_context_init( h, &a );
438 if( IS_X264_TYPE_I(h->fenc->i_type) )
442 else if( X264_TYPE_P == h->fenc->i_type )
445 while( h->frames.current[p1] && IS_X264_TYPE_B( h->frames.current[p1]->i_type ) )
452 p1 = (h->fref1[0]->i_poc - h->fref0[0]->i_poc)/2;
453 b = (h->fref1[0]->i_poc - h->fenc->i_poc)/2;
454 frames[p1] = h->fref1[0];
456 frames[p0] = h->fref0[0];
459 cost = x264_slicetype_frame_cost( h, &a, frames, p0, p1, b );
460 h->fenc->i_row_satd = h->fenc->i_row_satds[b-p0][p1-b];
461 h->fdec->i_row_satd = h->fdec->i_row_satds[b-p0][p1-b];
462 h->fdec->i_satd = cost;
463 memcpy( h->fdec->i_row_satd, h->fenc->i_row_satd, h->sps->i_mb_height * sizeof(int) );