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 *****************************************************************************/
28 #include "common/common.h"
29 #include "common/macroblock.h"
30 #include "common/cpu.h"
31 #include "macroblock.h"
35 static void x264_mb_analyse_load_costs_lowres( x264_t *h, x264_mb_analysis_t *a )
37 static int16_t *p_cost_mv;
42 x264_cpu_restore( h->param.cpu );
43 p_cost_mv = x264_malloc( (2*2*h->param.analyse.i_mv_range + 1) * sizeof(int16_t) );
44 p_cost_mv += 2*h->param.analyse.i_mv_range;
45 for( i = 0; i <= 2*h->param.analyse.i_mv_range; i++ )
47 p_cost_mv[i] = (int)( a->i_lambda * (1 + 2*log(2*i+1)/log(2)) );
50 a->p_cost_mv = p_cost_mv;
53 static void x264_lowres_context_init( x264_t *h, x264_mb_analysis_t *a )
55 a->i_qp = 12; // arbitrary, but low because SATD scores are 1/4 normal
56 a->i_lambda = i_qp0_cost_table[ a->i_qp ];
57 x264_mb_analyse_load_costs_lowres( h, a );
58 h->mb.i_me_method = X264_MIN( X264_ME_HEX, h->param.analyse.i_me_method ); // maybe dia?
59 h->mb.i_subpel_refine = 4; // 3 should be enough, but not tweaking for speed now
60 h->mb.b_chroma_me = 0;
62 h->mb.mv_min_fpel[0] =
63 h->mb.mv_min_fpel[1] = -16;
64 h->mb.mv_max_fpel[0] =
65 h->mb.mv_max_fpel[1] = 16;
67 h->mb.mv_min[1] = -4*32;
69 h->mb.mv_max[1] = 4*32;
72 int x264_slicetype_mb_cost( x264_t *h, x264_mb_analysis_t *a,
73 x264_frame_t **frames, int p0, int p1, int b,
74 int dist_scale_factor )
76 x264_frame_t *fref0 = frames[p0];
77 x264_frame_t *fref1 = frames[p1];
78 x264_frame_t *fenc = frames[b];
79 const int b_bidir = (b < p1);
80 const int i_mb_x = h->mb.i_mb_x;
81 const int i_mb_y = h->mb.i_mb_y;
82 const int i_mb_stride = h->sps->i_mb_width;
83 const int i_mb_xy = i_mb_x + i_mb_y * i_mb_stride;
84 const int i_stride = fenc->i_stride_lowres;
85 const int i_pel_offset = 8 * ( i_mb_x + i_mb_y * i_stride );
87 uint8_t pix1[9*9], pix2[8*8];
90 int i_bcost = COST_MAX;
94 if( !p0 && !p1 && !b )
97 #define LOAD_HPELS_LUMA(dst, src) \
99 (dst)[0] = &(src)[0][i_pel_offset]; \
100 (dst)[1] = &(src)[1][i_pel_offset]; \
101 (dst)[2] = &(src)[2][i_pel_offset]; \
102 (dst)[3] = &(src)[3][i_pel_offset]; \
104 #define SAVE_MVS( mv0, mv1 ) \
106 fenc->mv[0][i_mb_xy][0] = mv0[0]; \
107 fenc->mv[0][i_mb_xy][1] = mv0[1]; \
110 fenc->mv[1][i_mb_xy][0] = mv1[0]; \
111 fenc->mv[1][i_mb_xy][1] = mv1[1]; \
114 #define TRY_BIDIR( mv0, mv1, penalty ) \
119 h->mc.mc_luma( m[0].p_fref, m[0].i_stride[0], pix1, 8, \
120 (mv0)[0], (mv0)[1], 8, 8 ); \
121 src2 = h->mc.get_ref( m[1].p_fref, m[1].i_stride[0], pix2, &stride2, \
122 (mv1)[0], (mv1)[1], 8, 8 ); \
123 h->pixf.avg[PIXEL_8x8]( pix1, 8, src2, stride2 ); \
124 i_cost = penalty + h->pixf.mbcmp[PIXEL_8x8]( \
125 m[0].p_fenc[0], m[0].i_stride[0], pix1, 8 ); \
126 if( i_bcost > i_cost ) \
129 SAVE_MVS( mv0, mv1 ); \
133 m[0].i_pixel = PIXEL_8x8;
134 m[0].p_cost_mv = a->p_cost_mv;
135 m[0].i_stride[0] = i_stride;
136 m[0].p_fenc[0] = &fenc->lowres[0][ i_pel_offset ];
137 LOAD_HPELS_LUMA( m[0].p_fref, fref0->lowres );
141 int16_t *mvr = fref1->mv[0][i_mb_xy];
146 LOAD_HPELS_LUMA( m[1].p_fref, fref1->lowres );
148 dmv[0][0] = ( mvr[0] * dist_scale_factor + 128 ) >> 8;
149 dmv[0][1] = ( mvr[1] * dist_scale_factor + 128 ) >> 8;
150 dmv[1][0] = dmv[0][0] - mvr[0];
151 dmv[1][1] = dmv[0][1] - mvr[1];
153 TRY_BIDIR( dmv[0], dmv[1], 0 );
154 TRY_BIDIR( mv0, mv0, 0 );
155 // if( i_bcost < 60 ) // arbitrary threshold
159 i_cost_bak = i_bcost;
160 for( l = 0; l < 1 + b_bidir; l++ )
162 int16_t (*fenc_mv)[2] = &fenc->mv[0][i_mb_xy];
163 mvc[0][0] = fenc_mv[-1][0];
164 mvc[0][1] = fenc_mv[-1][1];
165 mvc[1][0] = fenc_mv[-i_mb_stride][0];
166 mvc[1][1] = fenc_mv[-i_mb_stride][1];
167 mvc[2][0] = fenc_mv[-i_mb_stride+1][0];
168 mvc[2][1] = fenc_mv[-i_mb_stride+1][1];
169 mvc[3][0] = fenc_mv[-i_mb_stride-1][0];
170 mvc[3][1] = fenc_mv[-i_mb_stride-1][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] );
175 x264_me_search( h, &m[l], mvc, i_mvc );
177 i_bcost = X264_MIN( i_bcost, m[l].cost + 3 );
181 TRY_BIDIR( m[0].mv, m[1].mv, 5 );
183 if( i_bcost < i_cost_bak )
184 SAVE_MVS( m[0].mv, m[1].mv );
188 uint8_t *src = &fenc->lowres[0][ i_pel_offset - i_stride - 1 ];
189 int intra_penalty = 5 + 10 * b_bidir;
190 i_cost_bak = i_bcost;
192 memcpy( pix1, src, 9 );
193 for( i=1; i<9; i++, src += i_stride )
195 src = &fenc->lowres[0][ i_pel_offset ];
197 for( i = I_PRED_CHROMA_DC; i <= I_PRED_CHROMA_P; i++ )
200 h->predict_8x8c[i]( &pix1[10], 9 );
201 i_cost = h->pixf.mbcmp[PIXEL_8x8]( &pix1[10], 9, src, i_stride ) + intra_penalty;
202 i_bcost = X264_MIN( i_bcost, i_cost );
204 if( i_bcost != i_cost_bak )
207 fenc->i_intra_mbs[b-p0]++;
209 i_bcost = i_bcost * 9 / 8; // arbitray penalty for I-blocks in and after B-frames
218 int x264_slicetype_frame_cost( x264_t *h, x264_mb_analysis_t *a,
219 x264_frame_t **frames, int p0, int p1, int b )
222 int dist_scale_factor = 128;
224 /* Check whether we already evaluated this frame
225 * If we have tried this frame as P, then we have also tried
226 * the preceding frames as B. (is this still true?) */
227 if( frames[b]->i_cost_est[b-p0][p1-b] >= 0 )
228 return frames[b]->i_cost_est[b-p0][p1-b];
230 /* Init MVs so that we don't have to check edge conditions when loading predictors. */
231 /* FIXME: not needed every time */
232 memset( frames[p1]->mv[0], 0, h->sps->i_mb_height * h->sps->i_mb_width * 2*sizeof(int) );
234 memset( frames[p1]->mv[1], 0, h->sps->i_mb_height * h->sps->i_mb_width * 2*sizeof(int) );
237 frames[b]->i_intra_mbs[b-p0] = 0;
239 dist_scale_factor = ( ((b-p0) << 8) + ((p1-p0) >> 1) ) / (p1-p0);
241 /* Skip the outermost ring of macroblocks, to simplify mv range and intra prediction. */
242 for( h->mb.i_mb_y = 1; h->mb.i_mb_y < h->sps->i_mb_height - 1; h->mb.i_mb_y++ )
243 for( h->mb.i_mb_x = 1; h->mb.i_mb_x < h->sps->i_mb_width - 1; h->mb.i_mb_x++ )
244 i_score += x264_slicetype_mb_cost( h, a, frames, p0, p1, b, dist_scale_factor );
247 i_score = i_score * 100 / (120 + h->param.i_bframe_bias);
249 frames[b]->i_cost_est[b-p0][p1-b] = i_score;
250 // fprintf( stderr, "frm %d %c(%d,%d): %6d I:%d \n", frames[b]->i_frame,
251 // (p1==0?'I':b<p1?'B':'P'), b-p0, p1-b, i_score, frames[b]->i_intra_mbs[b-p0] );
252 x264_cpu_restore( h->param.cpu );
256 void x264_slicetype_analyse( x264_t *h )
258 x264_mb_analysis_t a;
259 x264_frame_t *frames[X264_BFRAME_MAX+3] = { NULL, };
264 if( !h->frames.last_nonb )
266 frames[0] = h->frames.last_nonb;
267 for( j = 0; h->frames.next[j]; j++ )
268 frames[j+1] = h->frames.next[j];
269 keyint_limit = h->param.i_keyint_max - frames[0]->i_frame + h->frames.i_last_idr - 1;
270 num_frames = X264_MIN( j, keyint_limit );
271 if( num_frames == 0 )
273 if( num_frames == 1 )
276 frames[1]->i_type = X264_TYPE_P;
280 x264_lowres_context_init( h, &a );
283 /* BFS over possible frame types for minimum total SATD cost.
284 * requires higher encoding delay to be effective. */
289 char path[X264_BFRAME_MAX+1];
290 } paths[X264_BFRAME_MAX+1];
292 for( p1 = 1; frames[p1]; p1++ )
293 for( p0 = X264_MAX(0, p1 - h->param.i_bframe - 1); p0 < p1; p0++ )
294 for( b = p0+1; b <= p1; b++ )
295 x264_slicetype_frame_cost( h, &a, frames, p0, p1, b );
300 for( j = 1; j <= p1; j++ )
303 paths[j].score = INT_MAX;
305 for( k = 1; k <= X264_MIN( j, h->param.i_bframe + 1 ); k++ )
307 int64_t sum = paths[j-k].score;
308 for( i = 0; i < k; i++ )
309 sum += frames[j-k+i+1]->i_cost_est[(j-k+i+1) - (j-k)][j - (j-k+i+1)];
310 if( sum < paths[j].score )
312 paths[j].score = sum;
313 for( i = 0; i < j-k; i++ )
314 paths[j].path[i] = paths[j-k].path[i];
315 for( i = j-k; i < j-1; i++ )
316 paths[j].path[i] = 'B';
317 paths[j].path[j-1] = 'P';
318 paths[j].path[j] = 0;
319 fprintf( stderr, "path: %-8s %7lld \n", paths[j].path, sum );
324 for( j = 0; paths[p1].path[j] == 'B'; j++ )
325 frames[j+1]->i_type = X264_TYPE_B;
326 frames[j+1]->i_type = X264_TYPE_P;
330 int i_mb_count = (h->sps->i_mb_width - 2) * (h->sps->i_mb_height - 2);
331 int cost1p0, cost2p0, cost1b1, cost2p1;
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 cost2p0 = x264_slicetype_frame_cost( h, &a, frames, 1, 2, 2 );
338 cost1p0 = x264_slicetype_frame_cost( h, &a, frames, 0, 1, 1 );
339 cost1b1 = x264_slicetype_frame_cost( h, &a, frames, 0, 2, 1 );
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;
369 void x264_slicetype_decide( x264_t *h )
375 if( h->frames.next[0] == NULL )
378 if( h->param.rc.b_stat_read )
380 /* Use the frame types from the first pass */
381 for( i = 0; h->frames.next[i] != NULL; i++ )
382 h->frames.next[i]->i_type =
383 x264_ratecontrol_slice_type( h, h->frames.next[i]->i_frame );
385 else if( h->param.i_bframe && h->param.b_bframe_adaptive )
386 x264_slicetype_analyse( h );
388 for( bframes = 0;; bframes++ )
390 frm = h->frames.next[bframes];
393 if( frm->i_frame - h->frames.i_last_idr >= h->param.i_keyint_max )
395 if( frm->i_type == X264_TYPE_AUTO )
396 frm->i_type = X264_TYPE_IDR;
397 if( frm->i_type != X264_TYPE_IDR )
398 x264_log( h, X264_LOG_ERROR, "specified frame type (%d) is not compatible with keyframe interval\n", frm->i_type );
400 if( frm->i_type == X264_TYPE_IDR )
406 h->frames.next[bframes]->i_type = X264_TYPE_P;
414 if( bframes == h->param.i_bframe
415 || h->frames.next[bframes+1] == NULL )
417 if( IS_X264_TYPE_B( frm->i_type ) )
418 x264_log( h, X264_LOG_ERROR, "specified frame type is not compatible with max B-frames\n" );
419 if( frm->i_type == X264_TYPE_AUTO
420 || IS_X264_TYPE_B( frm->i_type ) )
421 frm->i_type = X264_TYPE_P;
424 if( frm->i_type != X264_TYPE_AUTO && frm->i_type != X264_TYPE_B && frm->i_type != X264_TYPE_BREF )
427 frm->i_type = X264_TYPE_B;
431 int x264_rc_analyse_slice( x264_t *h )
434 x264_mb_analysis_t a;
435 x264_frame_t *frames[X264_BFRAME_MAX+2] = { NULL, };
437 if( IS_X264_TYPE_I(h->fenc->i_type) )
438 return x264_slicetype_frame_cost( h, &a, &h->fenc, 0, 0, 0 );
440 while( h->frames.current[p1] && IS_X264_TYPE_B( h->frames.current[p1]->i_type ) )
443 if( h->fenc->i_cost_est[p1][0] >= 0 )
444 return h->fenc->i_cost_est[p1][0];
446 frames[0] = h->fref0[0];
447 frames[p1] = h->fenc;
448 x264_lowres_context_init( h, &a );
450 return x264_slicetype_frame_cost( h, &a, frames, 0, p1, p1 );