← Back to branch summary

~diresu/blender/blender-command-port

« back to all changes in this revision

Viewing changes to extern/x264/encoder/encoder.c

  • Committer: theeth
  • Date: 2008-10-14 16:52:04 UTC
  • Revision ID: vcs-imports@canonical.com-20081014165204-r32w2gm6s0osvdhn
copy back trunk

Show diffs side-by-side

added added

removed removed

Lines of Context:
extern/x264/encoder/encoder.c
 
1
/*****************************************************************************
 
2
 * x264: h264 encoder
 
3
 *****************************************************************************
 
4
 * Copyright (C) 2003 Laurent Aimar
 
5
 * $Id: encoder.c,v 1.1 2004/06/03 19:27:08 fenrir Exp $
 
6
 *
 
7
 * Authors: Laurent Aimar <fenrir@via.ecp.fr>
 
8
 *
 
9
 * This program is free software; you can redistribute it and/or modify
 
10
 * it under the terms of the GNU General Public License as published by
 
11
 * the Free Software Foundation; either version 2 of the License, or
 
12
 * (at your option) any later version.
 
13
 *
 
14
 * This program is distributed in the hope that it will be useful,
 
15
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 
16
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 
17
 * GNU General Public License for more details.
 
18
 *
 
19
 * You should have received a copy of the GNU General Public License
 
20
 * along with this program; if not, write to the Free Software
 
21
 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA  02111, USA.
 
22
 *****************************************************************************/
 
23
 
 
24
#include <math.h>
 
25
 
 
26
#include "common/common.h"
 
27
#include "common/cpu.h"
 
28
 
 
29
#include "set.h"
 
30
#include "analyse.h"
 
31
#include "ratecontrol.h"
 
32
#include "macroblock.h"
 
33
 
 
34
#if VISUALIZE
 
35
#include "common/visualize.h"
 
36
#endif
 
37
 
 
38
//#define DEBUG_MB_TYPE
 
39
//#define DEBUG_DUMP_FRAME
 
40
//#define DEBUG_BENCHMARK
 
41
 
 
42
#ifdef DEBUG_BENCHMARK
 
43
static int64_t i_mtime_encode_frame = 0;
 
44
static int64_t i_mtime_analyse = 0;
 
45
static int64_t i_mtime_encode = 0;
 
46
static int64_t i_mtime_write = 0;
 
47
static int64_t i_mtime_filter = 0;
 
48
#define TIMER_START( d ) \
 
49
    { \
 
50
        int64_t d##start = x264_mdate();
 
51
 
 
52
#define TIMER_STOP( d ) \
 
53
        d += x264_mdate() - d##start;\
 
54
    }
 
55
#else
 
56
#define TIMER_START( d )
 
57
#define TIMER_STOP( d )
 
58
#endif
 
59
 
 
60
#define NALU_OVERHEAD 5 // startcode + NAL type costs 5 bytes per frame
 
61
 
 
62
static void x264_encoder_frame_end( x264_t *h, x264_t *thread_current,
 
63
                                    x264_nal_t **pp_nal, int *pi_nal,
 
64
                                    x264_picture_t *pic_out );
 
65
 
 
66
/****************************************************************************
 
67
 *
 
68
 ******************************* x264 libs **********************************
 
69
 *
 
70
 ****************************************************************************/
 
71
static float x264_psnr( int64_t i_sqe, int64_t i_size )
 
72
{
 
73
    double f_mse = (double)i_sqe / ((double)65025.0 * (double)i_size);
 
74
    if( f_mse <= 0.0000000001 ) /* Max 100dB */
 
75
        return 100;
 
76
 
 
77
    return (float)(-10.0 * log( f_mse ) / log( 10.0 ));
 
78
}
 
79
 
 
80
#ifdef DEBUG_DUMP_FRAME
 
81
static void x264_frame_dump( x264_t *h, x264_frame_t *fr, char *name )
 
82
{
 
83
    FILE *f = fopen( name, "r+b" );
 
84
    int i, y;
 
85
    if( !f )
 
86
        return;
 
87
 
 
88
    /* Write the frame in display order */
 
89
    fseek( f, fr->i_frame * h->param.i_height * h->param.i_width * 3 / 2, SEEK_SET );
 
90
 
 
91
    for( i = 0; i < fr->i_plane; i++ )
 
92
    {
 
93
        for( y = 0; y < h->param.i_height / ( i == 0 ? 1 : 2 ); y++ )
 
94
        {
 
95
            fwrite( &fr->plane[i][y*fr->i_stride[i]], 1, h->param.i_width / ( i == 0 ? 1 : 2 ), f );
 
96
        }
 
97
    }
 
98
    fclose( f );
 
99
}
 
100
#endif
 
101
 
 
102
 
 
103
/* Fill "default" values */
 
104
static void x264_slice_header_init( x264_t *h, x264_slice_header_t *sh,
 
105
                                    x264_sps_t *sps, x264_pps_t *pps,
 
106
                                    int i_idr_pic_id, int i_frame, int i_qp )
 
107
{
 
108
    x264_param_t *param = &h->param;
 
109
    int i;
 
110
 
 
111
    /* First we fill all field */
 
112
    sh->sps = sps;
 
113
    sh->pps = pps;
 
114
 
 
115
    sh->i_first_mb  = 0;
 
116
    sh->i_last_mb   = h->sps->i_mb_width * h->sps->i_mb_height;
 
117
    sh->i_pps_id    = pps->i_id;
 
118
 
 
119
    sh->i_frame_num = i_frame;
 
120
 
 
121
    sh->b_mbaff = h->param.b_interlaced;
 
122
    sh->b_field_pic = 0;    /* no field support for now */
 
123
    sh->b_bottom_field = 0; /* not yet used */
 
124
 
 
125
    sh->i_idr_pic_id = i_idr_pic_id;
 
126
 
 
127
    /* poc stuff, fixed later */
 
128
    sh->i_poc_lsb = 0;
 
129
    sh->i_delta_poc_bottom = 0;
 
130
    sh->i_delta_poc[0] = 0;
 
131
    sh->i_delta_poc[1] = 0;
 
132
 
 
133
    sh->i_redundant_pic_cnt = 0;
 
134
 
 
135
    if( !h->mb.b_direct_auto_read )
 
136
    {
 
137
        if( h->mb.b_direct_auto_write )
 
138
            sh->b_direct_spatial_mv_pred = ( h->stat.i_direct_score[1] > h->stat.i_direct_score[0] );
 
139
        else
 
140
            sh->b_direct_spatial_mv_pred = ( param->analyse.i_direct_mv_pred == X264_DIRECT_PRED_SPATIAL );
 
141
    }
 
142
    /* else b_direct_spatial_mv_pred was read from the 2pass statsfile */
 
143
 
 
144
    sh->b_num_ref_idx_override = 0;
 
145
    sh->i_num_ref_idx_l0_active = 1;
 
146
    sh->i_num_ref_idx_l1_active = 1;
 
147
 
 
148
    sh->b_ref_pic_list_reordering_l0 = h->b_ref_reorder[0];
 
149
    sh->b_ref_pic_list_reordering_l1 = h->b_ref_reorder[1];
 
150
 
 
151
    /* If the ref list isn't in the default order, construct reordering header */
 
152
    /* List1 reordering isn't needed yet */
 
153
    if( sh->b_ref_pic_list_reordering_l0 )
 
154
    {
 
155
        int pred_frame_num = i_frame;
 
156
        for( i = 0; i < h->i_ref0; i++ )
 
157
        {
 
158
            int diff = h->fref0[i]->i_frame_num - pred_frame_num;
 
159
            if( diff == 0 )
 
160
                x264_log( h, X264_LOG_ERROR, "diff frame num == 0\n" );
 
161
            sh->ref_pic_list_order[0][i].idc = ( diff > 0 );
 
162
            sh->ref_pic_list_order[0][i].arg = abs( diff ) - 1;
 
163
            pred_frame_num = h->fref0[i]->i_frame_num;
 
164
        }
 
165
    }
 
166
 
 
167
    sh->i_cabac_init_idc = param->i_cabac_init_idc;
 
168
 
 
169
    sh->i_qp = i_qp;
 
170
    sh->i_qp_delta = i_qp - pps->i_pic_init_qp;
 
171
    sh->b_sp_for_swidth = 0;
 
172
    sh->i_qs_delta = 0;
 
173
 
 
174
    /* If effective qp <= 15, deblocking would have no effect anyway */
 
175
    if( param->b_deblocking_filter
 
176
        && ( h->mb.b_variable_qp
 
177
        || 15 < i_qp + 2 * X264_MAX(param->i_deblocking_filter_alphac0, param->i_deblocking_filter_beta) ) )
 
178
    {
 
179
        sh->i_disable_deblocking_filter_idc = 0;
 
180
    }
 
181
    else
 
182
    {
 
183
        sh->i_disable_deblocking_filter_idc = 1;
 
184
    }
 
185
    sh->i_alpha_c0_offset = param->i_deblocking_filter_alphac0 << 1;
 
186
    sh->i_beta_offset = param->i_deblocking_filter_beta << 1;
 
187
}
 
188
 
 
189
static void x264_slice_header_write( bs_t *s, x264_slice_header_t *sh, int i_nal_ref_idc )
 
190
{
 
191
    int i;
 
192
 
 
193
    if( sh->b_mbaff )
 
194
    {
 
195
        assert( sh->i_first_mb % (2*sh->sps->i_mb_width) == 0 );
 
196
        bs_write_ue( s, sh->i_first_mb >> 1 );
 
197
    }
 
198
    else
 
199
        bs_write_ue( s, sh->i_first_mb );
 
200
 
 
201
    bs_write_ue( s, sh->i_type + 5 );   /* same type things */
 
202
    bs_write_ue( s, sh->i_pps_id );
 
203
    bs_write( s, sh->sps->i_log2_max_frame_num, sh->i_frame_num );
 
204
 
 
205
    if( !sh->sps->b_frame_mbs_only )
 
206
    {
 
207
        bs_write1( s, sh->b_field_pic );
 
208
        if ( sh->b_field_pic )
 
209
            bs_write1( s, sh->b_bottom_field );
 
210
    }
 
211
 
 
212
    if( sh->i_idr_pic_id >= 0 ) /* NAL IDR */
 
213
    {
 
214
        bs_write_ue( s, sh->i_idr_pic_id );
 
215
    }
 
216
 
 
217
    if( sh->sps->i_poc_type == 0 )
 
218
    {
 
219
        bs_write( s, sh->sps->i_log2_max_poc_lsb, sh->i_poc_lsb );
 
220
        if( sh->pps->b_pic_order && !sh->b_field_pic )
 
221
        {
 
222
            bs_write_se( s, sh->i_delta_poc_bottom );
 
223
        }
 
224
    }
 
225
    else if( sh->sps->i_poc_type == 1 && !sh->sps->b_delta_pic_order_always_zero )
 
226
    {
 
227
        bs_write_se( s, sh->i_delta_poc[0] );
 
228
        if( sh->pps->b_pic_order && !sh->b_field_pic )
 
229
        {
 
230
            bs_write_se( s, sh->i_delta_poc[1] );
 
231
        }
 
232
    }
 
233
 
 
234
    if( sh->pps->b_redundant_pic_cnt )
 
235
    {
 
236
        bs_write_ue( s, sh->i_redundant_pic_cnt );
 
237
    }
 
238
 
 
239
    if( sh->i_type == SLICE_TYPE_B )
 
240
    {
 
241
        bs_write1( s, sh->b_direct_spatial_mv_pred );
 
242
    }
 
243
    if( sh->i_type == SLICE_TYPE_P || sh->i_type == SLICE_TYPE_SP || sh->i_type == SLICE_TYPE_B )
 
244
    {
 
245
        bs_write1( s, sh->b_num_ref_idx_override );
 
246
        if( sh->b_num_ref_idx_override )
 
247
        {
 
248
            bs_write_ue( s, sh->i_num_ref_idx_l0_active - 1 );
 
249
            if( sh->i_type == SLICE_TYPE_B )
 
250
            {
 
251
                bs_write_ue( s, sh->i_num_ref_idx_l1_active - 1 );
 
252
            }
 
253
        }
 
254
    }
 
255
 
 
256
    /* ref pic list reordering */
 
257
    if( sh->i_type != SLICE_TYPE_I )
 
258
    {
 
259
        bs_write1( s, sh->b_ref_pic_list_reordering_l0 );
 
260
        if( sh->b_ref_pic_list_reordering_l0 )
 
261
        {
 
262
            for( i = 0; i < sh->i_num_ref_idx_l0_active; i++ )
 
263
            {
 
264
                bs_write_ue( s, sh->ref_pic_list_order[0][i].idc );
 
265
                bs_write_ue( s, sh->ref_pic_list_order[0][i].arg );
 
266
                        
 
267
            }
 
268
            bs_write_ue( s, 3 );
 
269
        }
 
270
    }
 
271
    if( sh->i_type == SLICE_TYPE_B )
 
272
    {
 
273
        bs_write1( s, sh->b_ref_pic_list_reordering_l1 );
 
274
        if( sh->b_ref_pic_list_reordering_l1 )
 
275
        {
 
276
            for( i = 0; i < sh->i_num_ref_idx_l1_active; i++ )
 
277
            {
 
278
                bs_write_ue( s, sh->ref_pic_list_order[1][i].idc );
 
279
                bs_write_ue( s, sh->ref_pic_list_order[1][i].arg );
 
280
            }
 
281
            bs_write_ue( s, 3 );
 
282
        }
 
283
    }
 
284
 
 
285
    if( ( sh->pps->b_weighted_pred && ( sh->i_type == SLICE_TYPE_P || sh->i_type == SLICE_TYPE_SP ) ) ||
 
286
        ( sh->pps->b_weighted_bipred == 1 && sh->i_type == SLICE_TYPE_B ) )
 
287
    {
 
288
        /* FIXME */
 
289
    }
 
290
 
 
291
    if( i_nal_ref_idc != 0 )
 
292
    {
 
293
        if( sh->i_idr_pic_id >= 0 )
 
294
        {
 
295
            bs_write1( s, 0 );  /* no output of prior pics flag */
 
296
            bs_write1( s, 0 );  /* long term reference flag */
 
297
        }
 
298
        else
 
299
        {
 
300
            bs_write1( s, 0 );  /* adaptive_ref_pic_marking_mode_flag */
 
301
        }
 
302
    }
 
303
 
 
304
    if( sh->pps->b_cabac && sh->i_type != SLICE_TYPE_I )
 
305
    {
 
306
        bs_write_ue( s, sh->i_cabac_init_idc );
 
307
    }
 
308
    bs_write_se( s, sh->i_qp_delta );      /* slice qp delta */
 
309
 
 
310
    if( sh->pps->b_deblocking_filter_control )
 
311
    {
 
312
        bs_write_ue( s, sh->i_disable_deblocking_filter_idc );
 
313
        if( sh->i_disable_deblocking_filter_idc != 1 )
 
314
        {
 
315
            bs_write_se( s, sh->i_alpha_c0_offset >> 1 );
 
316
            bs_write_se( s, sh->i_beta_offset >> 1 );
 
317
        }
 
318
    }
 
319
}
 
320
 
 
321
/****************************************************************************
 
322
 *
 
323
 ****************************************************************************
 
324
 ****************************** External API*********************************
 
325
 ****************************************************************************
 
326
 *
 
327
 ****************************************************************************/
 
328
 
 
329
static int x264_validate_parameters( x264_t *h )
 
330
{
 
331
    if( h->param.i_width <= 0 || h->param.i_height <= 0 )
 
332
    {
 
333
        x264_log( h, X264_LOG_ERROR, "invalid width x height (%dx%d)\n",
 
334
                  h->param.i_width, h->param.i_height );
 
335
        return -1;
 
336
    }
 
337
 
 
338
    if( h->param.i_width % 2 || h->param.i_height % 2 )
 
339
    {
 
340
        x264_log( h, X264_LOG_ERROR, "width or height not divisible by 2 (%dx%d)\n",
 
341
                  h->param.i_width, h->param.i_height );
 
342
        return -1;
 
343
    }
 
344
    if( h->param.i_csp != X264_CSP_I420 )
 
345
    {
 
346
        x264_log( h, X264_LOG_ERROR, "invalid CSP (only I420 supported)\n" );
 
347
        return -1;
 
348
    }
 
349
 
 
350
    if( h->param.i_threads == 0 )
 
351
        h->param.i_threads = x264_cpu_num_processors() * 3/2;
 
352
    h->param.i_threads = x264_clip3( h->param.i_threads, 1, X264_THREAD_MAX );
 
353
    if( h->param.i_threads > 1 )
 
354
    {
 
355
#ifndef HAVE_PTHREAD
 
356
        x264_log( h, X264_LOG_WARNING, "not compiled with pthread support!\n");
 
357
        h->param.i_threads = 1;
 
358
#else
 
359
        if( h->param.i_scenecut_threshold >= 0 )
 
360
            h->param.b_pre_scenecut = 1;
 
361
#endif
 
362
    }
 
363
 
 
364
    if( h->param.b_interlaced )
 
365
    {
 
366
        if( h->param.analyse.i_me_method >= X264_ME_ESA )
 
367
        {
 
368
            x264_log( h, X264_LOG_WARNING, "interlace + me=esa is not implemented\n" );
 
369
            h->param.analyse.i_me_method = X264_ME_UMH;
 
370
        }
 
371
        if( h->param.analyse.i_direct_mv_pred > X264_DIRECT_PRED_SPATIAL )
 
372
        {
 
373
            x264_log( h, X264_LOG_WARNING, "interlace + direct=temporal is not implemented\n" );
 
374
            h->param.analyse.i_direct_mv_pred = X264_DIRECT_PRED_SPATIAL;
 
375
        }
 
376
    }
 
377
 
 
378
    if( h->param.rc.i_rc_method < 0 || h->param.rc.i_rc_method > 2 )
 
379
    {
 
380
        x264_log( h, X264_LOG_ERROR, "no ratecontrol method specified\n" );
 
381
        return -1;
 
382
    }
 
383
    h->param.rc.f_rf_constant = x264_clip3f( h->param.rc.f_rf_constant, 0, 51 );
 
384
    h->param.rc.i_qp_constant = x264_clip3( h->param.rc.i_qp_constant, 0, 51 );
 
385
    if( h->param.rc.i_rc_method == X264_RC_CRF )
 
386
        h->param.rc.i_qp_constant = h->param.rc.f_rf_constant;
 
387
    if( (h->param.rc.i_rc_method == X264_RC_CQP || h->param.rc.i_rc_method == X264_RC_CRF)
 
388
        && h->param.rc.i_qp_constant == 0 )
 
389
    {
 
390
        h->mb.b_lossless = 1;
 
391
        h->param.i_cqm_preset = X264_CQM_FLAT;
 
392
        h->param.psz_cqm_file = NULL;
 
393
        h->param.rc.i_rc_method = X264_RC_CQP;
 
394
        h->param.rc.f_ip_factor = 1;
 
395
        h->param.rc.f_pb_factor = 1;
 
396
        h->param.analyse.b_transform_8x8 = 0;
 
397
        h->param.analyse.b_psnr = 0;
 
398
        h->param.analyse.b_ssim = 0;
 
399
        h->param.analyse.i_chroma_qp_offset = 0;
 
400
        h->param.analyse.i_trellis = 0;
 
401
        h->param.analyse.b_fast_pskip = 0;
 
402
        h->param.analyse.i_noise_reduction = 0;
 
403
        h->param.analyse.i_subpel_refine = x264_clip3( h->param.analyse.i_subpel_refine, 1, 6 );
 
404
    }
 
405
    if( h->param.rc.i_rc_method == X264_RC_CQP )
 
406
    {
 
407
        float qp_p = h->param.rc.i_qp_constant;
 
408
        float qp_i = qp_p - 6*log(h->param.rc.f_ip_factor)/log(2);
 
409
        float qp_b = qp_p + 6*log(h->param.rc.f_pb_factor)/log(2);
 
410
        h->param.rc.i_qp_min = x264_clip3( (int)(X264_MIN3( qp_p, qp_i, qp_b )), 0, 51 );
 
411
        h->param.rc.i_qp_max = x264_clip3( (int)(X264_MAX3( qp_p, qp_i, qp_b ) + .999), 0, 51 );
 
412
    }
 
413
 
 
414
    if( ( h->param.i_width % 16 || h->param.i_height % 16 )
 
415
        && h->param.i_height != 1080 && !h->mb.b_lossless )
 
416
    {
 
417
        // There's nothing special about 1080 in that the warning still applies to it,
 
418
        // but chances are the user can't help it if his content is already 1080p,
 
419
        // so there's no point in warning in that case.
 
420
        x264_log( h, X264_LOG_WARNING, 
 
421
                  "width or height not divisible by 16 (%dx%d), compression will suffer.\n",
 
422
                  h->param.i_width, h->param.i_height );
 
423
    }
 
424
 
 
425
    h->param.i_frame_reference = x264_clip3( h->param.i_frame_reference, 1, 16 );
 
426
    if( h->param.i_keyint_max <= 0 )
 
427
        h->param.i_keyint_max = 1;
 
428
    h->param.i_keyint_min = x264_clip3( h->param.i_keyint_min, 1, h->param.i_keyint_max/2+1 );
 
429
 
 
430
    h->param.i_bframe = x264_clip3( h->param.i_bframe, 0, X264_BFRAME_MAX );
 
431
    h->param.i_bframe_bias = x264_clip3( h->param.i_bframe_bias, -90, 100 );
 
432
    h->param.b_bframe_pyramid = h->param.b_bframe_pyramid && h->param.i_bframe > 1;
 
433
    h->param.b_bframe_adaptive = h->param.b_bframe_adaptive && h->param.i_bframe > 0;
 
434
    h->param.analyse.b_weighted_bipred = h->param.analyse.b_weighted_bipred && h->param.i_bframe > 0;
 
435
    h->mb.b_direct_auto_write = h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO
 
436
                                && h->param.i_bframe
 
437
                                && ( h->param.rc.b_stat_write || !h->param.rc.b_stat_read );
 
438
    if( h->param.i_scenecut_threshold < 0 )
 
439
        h->param.b_pre_scenecut = 0;
 
440
 
 
441
    h->param.i_deblocking_filter_alphac0 = x264_clip3( h->param.i_deblocking_filter_alphac0, -6, 6 );
 
442
    h->param.i_deblocking_filter_beta    = x264_clip3( h->param.i_deblocking_filter_beta, -6, 6 );
 
443
    h->param.analyse.i_luma_deadzone[0] = x264_clip3( h->param.analyse.i_luma_deadzone[0], 0, 32 );
 
444
    h->param.analyse.i_luma_deadzone[1] = x264_clip3( h->param.analyse.i_luma_deadzone[1], 0, 32 );
 
445
 
 
446
    h->param.i_cabac_init_idc = x264_clip3( h->param.i_cabac_init_idc, 0, 2 );
 
447
 
 
448
    if( h->param.i_cqm_preset < X264_CQM_FLAT || h->param.i_cqm_preset > X264_CQM_CUSTOM )
 
449
        h->param.i_cqm_preset = X264_CQM_FLAT;
 
450
 
 
451
    if( h->param.analyse.i_me_method < X264_ME_DIA ||
 
452
        h->param.analyse.i_me_method > X264_ME_TESA )
 
453
        h->param.analyse.i_me_method = X264_ME_HEX;
 
454
    if( h->param.analyse.i_me_range < 4 )
 
455
        h->param.analyse.i_me_range = 4;
 
456
    if( h->param.analyse.i_me_range > 16 && h->param.analyse.i_me_method <= X264_ME_HEX )
 
457
        h->param.analyse.i_me_range = 16;
 
458
    if( h->param.analyse.i_me_method == X264_ME_TESA &&
 
459
        (h->mb.b_lossless || h->param.analyse.i_subpel_refine <= 1) )
 
460
        h->param.analyse.i_me_method = X264_ME_ESA;
 
461
    h->param.analyse.i_subpel_refine = x264_clip3( h->param.analyse.i_subpel_refine, 1, 7 );
 
462
    h->param.analyse.b_bframe_rdo = h->param.analyse.b_bframe_rdo && h->param.analyse.i_subpel_refine >= 6;
 
463
    h->param.analyse.b_mixed_references = h->param.analyse.b_mixed_references && h->param.i_frame_reference > 1;
 
464
    h->param.analyse.inter &= X264_ANALYSE_PSUB16x16|X264_ANALYSE_PSUB8x8|X264_ANALYSE_BSUB16x16|
 
465
                              X264_ANALYSE_I4x4|X264_ANALYSE_I8x8;
 
466
    h->param.analyse.intra &= X264_ANALYSE_I4x4|X264_ANALYSE_I8x8;
 
467
    if( !(h->param.analyse.inter & X264_ANALYSE_PSUB16x16) )
 
468
        h->param.analyse.inter &= ~X264_ANALYSE_PSUB8x8;
 
469
    if( !h->param.analyse.b_transform_8x8 )
 
470
    {
 
471
        h->param.analyse.inter &= ~X264_ANALYSE_I8x8;
 
472
        h->param.analyse.intra &= ~X264_ANALYSE_I8x8;
 
473
    }
 
474
    h->param.analyse.i_chroma_qp_offset = x264_clip3(h->param.analyse.i_chroma_qp_offset, -12, 12);
 
475
    if( !h->param.b_cabac )
 
476
        h->param.analyse.i_trellis = 0;
 
477
    h->param.analyse.i_trellis = x264_clip3( h->param.analyse.i_trellis, 0, 2 );
 
478
    h->param.analyse.i_noise_reduction = x264_clip3( h->param.analyse.i_noise_reduction, 0, 1<<16 );
 
479
 
 
480
    {
 
481
        const x264_level_t *l = x264_levels;
 
482
        while( l->level_idc != 0 && l->level_idc != h->param.i_level_idc )
 
483
            l++;
 
484
        if( l->level_idc == 0 )
 
485
        {
 
486
            x264_log( h, X264_LOG_ERROR, "invalid level_idc: %d\n", h->param.i_level_idc );
 
487
            return -1;
 
488
        }
 
489
        if( h->param.analyse.i_mv_range <= 0 )
 
490
            h->param.analyse.i_mv_range = l->mv_range;
 
491
        else
 
492
            h->param.analyse.i_mv_range = x264_clip3(h->param.analyse.i_mv_range, 32, 512);
 
493
        if( h->param.analyse.i_direct_8x8_inference < 0 )
 
494
            h->param.analyse.i_direct_8x8_inference = l->direct8x8;
 
495
    }
 
496
 
 
497
    if( h->param.i_threads > 1 )
 
498
    {
 
499
        int r = h->param.analyse.i_mv_range_thread;
 
500
        int r2;
 
501
        if( r <= 0 )
 
502
        {
 
503
            // half of the available space is reserved and divided evenly among the threads,
 
504
            // the rest is allocated to whichever thread is far enough ahead to use it.
 
505
            // reserving more space increases quality for some videos, but costs more time
 
506
            // in thread synchronization.
 
507
            int max_range = (h->param.i_height + X264_THREAD_HEIGHT) / h->param.i_threads - X264_THREAD_HEIGHT;
 
508
            r = max_range / 2;
 
509
        }
 
510
        r = X264_MAX( r, h->param.analyse.i_me_range );
 
511
        r = X264_MIN( r, h->param.analyse.i_mv_range );
 
512
        // round up to use the whole mb row
 
513
        r2 = (r & ~15) + ((-X264_THREAD_HEIGHT) & 15);
 
514
        if( r2 < r )
 
515
            r2 += 16;
 
516
        x264_log( h, X264_LOG_DEBUG, "using mv_range_thread = %d\n", r2 );
 
517
        h->param.analyse.i_mv_range_thread = r2;
 
518
    }
 
519
 
 
520
    if( h->param.rc.f_qblur < 0 )
 
521
        h->param.rc.f_qblur = 0;
 
522
    if( h->param.rc.f_complexity_blur < 0 )
 
523
        h->param.rc.f_complexity_blur = 0;
 
524
 
 
525
    h->param.i_sps_id &= 31;
 
526
 
 
527
    if( h->param.i_log_level < X264_LOG_INFO )
 
528
    {
 
529
        h->param.analyse.b_psnr = 0;
 
530
        h->param.analyse.b_ssim = 0;
 
531
    }
 
532
 
 
533
    /* ensure the booleans are 0 or 1 so they can be used in math */
 
534
#define BOOLIFY(x) h->param.x = !!h->param.x
 
535
    BOOLIFY( b_cabac );
 
536
    BOOLIFY( b_deblocking_filter );
 
537
    BOOLIFY( b_interlaced );
 
538
    BOOLIFY( analyse.b_transform_8x8 );
 
539
    BOOLIFY( analyse.i_direct_8x8_inference );
 
540
    BOOLIFY( analyse.b_bidir_me );
 
541
    BOOLIFY( analyse.b_chroma_me );
 
542
    BOOLIFY( analyse.b_fast_pskip );
 
543
    BOOLIFY( rc.b_stat_write );
 
544
    BOOLIFY( rc.b_stat_read );
 
545
#undef BOOLIFY
 
546
 
 
547
    return 0;
 
548
}
 
549
 
 
550
static void mbcmp_init( x264_t *h )
 
551
{
 
552
    int satd = !h->mb.b_lossless && h->param.analyse.i_subpel_refine > 1;
 
553
    memcpy( h->pixf.mbcmp, satd ? h->pixf.satd : h->pixf.sad, sizeof(h->pixf.mbcmp) );
 
554
    satd &= h->param.analyse.i_me_method == X264_ME_TESA;
 
555
    memcpy( h->pixf.fpelcmp, satd ? h->pixf.satd : h->pixf.sad, sizeof(h->pixf.fpelcmp) );
 
556
    memcpy( h->pixf.fpelcmp_x3, satd ? h->pixf.satd_x3 : h->pixf.sad_x3, sizeof(h->pixf.fpelcmp_x3) );
 
557
    memcpy( h->pixf.fpelcmp_x4, satd ? h->pixf.satd_x4 : h->pixf.sad_x4, sizeof(h->pixf.fpelcmp_x4) );
 
558
}
 
559
 
 
560
/****************************************************************************
 
561
 * x264_encoder_open:
 
562
 ****************************************************************************/
 
563
x264_t *x264_encoder_open   ( x264_param_t *param )
 
564
{
 
565
    x264_t *h = x264_malloc( sizeof( x264_t ) );
 
566
    int i;
 
567
 
 
568
    memset( h, 0, sizeof( x264_t ) );
 
569
 
 
570
    /* Create a copy of param */
 
571
    memcpy( &h->param, param, sizeof( x264_param_t ) );
 
572
 
 
573
    if( x264_validate_parameters( h ) < 0 )
 
574
    {
 
575
        x264_free( h );
 
576
        return NULL;
 
577
    }
 
578
 
 
579
    if( h->param.psz_cqm_file )
 
580
        if( x264_cqm_parse_file( h, h->param.psz_cqm_file ) < 0 )
 
581
        {
 
582
            x264_free( h );
 
583
            return NULL;
 
584
        }
 
585
 
 
586
    if( h->param.rc.psz_stat_out )
 
587
        h->param.rc.psz_stat_out = strdup( h->param.rc.psz_stat_out );
 
588
    if( h->param.rc.psz_stat_in )
 
589
        h->param.rc.psz_stat_in = strdup( h->param.rc.psz_stat_in );
 
590
    if( h->param.rc.psz_rc_eq )
 
591
        h->param.rc.psz_rc_eq = strdup( h->param.rc.psz_rc_eq );
 
592
 
 
593
    /* VUI */
 
594
    if( h->param.vui.i_sar_width > 0 && h->param.vui.i_sar_height > 0 )
 
595
    {
 
596
        int i_w = param->vui.i_sar_width;
 
597
        int i_h = param->vui.i_sar_height;
 
598
 
 
599
        x264_reduce_fraction( &i_w, &i_h );
 
600
 
 
601
        while( i_w > 65535 || i_h > 65535 )
 
602
        {
 
603
            i_w /= 2;
 
604
            i_h /= 2;
 
605
        }
 
606
 
 
607
        h->param.vui.i_sar_width = 0;
 
608
        h->param.vui.i_sar_height = 0;
 
609
        if( i_w == 0 || i_h == 0 )
 
610
        {
 
611
            x264_log( h, X264_LOG_WARNING, "cannot create valid sample aspect ratio\n" );
 
612
        }
 
613
        else
 
614
        {
 
615
            x264_log( h, X264_LOG_INFO, "using SAR=%d/%d\n", i_w, i_h );
 
616
            h->param.vui.i_sar_width = i_w;
 
617
            h->param.vui.i_sar_height = i_h;
 
618
        }
 
619
    }
 
620
 
 
621
    x264_reduce_fraction( &h->param.i_fps_num, &h->param.i_fps_den );
 
622
 
 
623
    /* Init x264_t */
 
624
    h->i_frame = 0;
 
625
    h->i_frame_num = 0;
 
626
    h->i_idr_pic_id = 0;
 
627
 
 
628
    h->sps = &h->sps_array[0];
 
629
    x264_sps_init( h->sps, h->param.i_sps_id, &h->param );
 
630
 
 
631
    h->pps = &h->pps_array[0];
 
632
    x264_pps_init( h->pps, h->param.i_sps_id, &h->param, h->sps);
 
633
 
 
634
    x264_validate_levels( h );
 
635
 
 
636
    if( x264_cqm_init( h ) < 0 )
 
637
    {
 
638
        x264_free( h );
 
639
        return NULL;
 
640
    }
 
641
    
 
642
    h->mb.i_mb_count = h->sps->i_mb_width * h->sps->i_mb_height;
 
643
 
 
644
    /* Init frames. */
 
645
    h->frames.i_delay = h->param.i_bframe + h->param.i_threads - 1;
 
646
    h->frames.i_max_ref0 = h->param.i_frame_reference;
 
647
    h->frames.i_max_ref1 = h->sps->vui.i_num_reorder_frames;
 
648
    h->frames.i_max_dpb  = h->sps->vui.i_max_dec_frame_buffering;
 
649
    h->frames.b_have_lowres = !h->param.rc.b_stat_read
 
650
        && ( h->param.rc.i_rc_method == X264_RC_ABR
 
651
          || h->param.rc.i_rc_method == X264_RC_CRF
 
652
          || h->param.b_bframe_adaptive
 
653
          || h->param.b_pre_scenecut );
 
654
 
 
655
    h->frames.i_last_idr = - h->param.i_keyint_max;
 
656
    h->frames.i_input    = 0;
 
657
    h->frames.last_nonb  = NULL;
 
658
 
 
659
    h->i_ref0 = 0;
 
660
    h->i_ref1 = 0;
 
661
 
 
662
    x264_rdo_init( );
 
663
 
 
664
    /* init CPU functions */
 
665
    x264_predict_16x16_init( h->param.cpu, h->predict_16x16 );
 
666
    x264_predict_8x8c_init( h->param.cpu, h->predict_8x8c );
 
667
    x264_predict_8x8_init( h->param.cpu, h->predict_8x8 );
 
668
    x264_predict_4x4_init( h->param.cpu, h->predict_4x4 );
 
669
 
 
670
    x264_pixel_init( h->param.cpu, &h->pixf );
 
671
    x264_dct_init( h->param.cpu, &h->dctf );
 
672
    x264_zigzag_init( h->param.cpu, &h->zigzagf, h->param.b_interlaced );
 
673
    x264_mc_init( h->param.cpu, &h->mc );
 
674
    x264_quant_init( h, h->param.cpu, &h->quantf );
 
675
    x264_deblock_init( h->param.cpu, &h->loopf );
 
676
    x264_dct_init_weights();
 
677
 
 
678
    mbcmp_init( h );
 
679
 
 
680
    x264_log( h, X264_LOG_INFO, "using cpu capabilities: %s%s%s%s%s%s%s%s%s%s\n",
 
681
             param->cpu&X264_CPU_MMX ? "MMX " : "",
 
682
             param->cpu&X264_CPU_MMXEXT ? "MMXEXT " : "",
 
683
             param->cpu&X264_CPU_SSE ? "SSE " : "",
 
684
             param->cpu&X264_CPU_SSE2 ? "SSE2 " : "",
 
685
             param->cpu&X264_CPU_SSE3 ? "SSE3 " : "",
 
686
             param->cpu&X264_CPU_SSSE3 ? "SSSE3 " : "",
 
687
             param->cpu&X264_CPU_3DNOW ? "3DNow! " : "",
 
688
             param->cpu&X264_CPU_ALTIVEC ? "Altivec " : "",
 
689
             param->cpu&X264_CPU_CACHELINE_SPLIT ?
 
690
                 param->cpu&X264_CPU_CACHELINE_32 ? "Cache32 " :
 
691
                 param->cpu&X264_CPU_CACHELINE_64 ? "Cache64 " : "Cache? " : "",
 
692
             param->cpu ? "" : "none!" );
 
693
 
 
694
    h->out.i_nal = 0;
 
695
    h->out.i_bitstream = X264_MAX( 1000000, h->param.i_width * h->param.i_height * 4
 
696
        * ( h->param.rc.i_rc_method == X264_RC_ABR ? pow( 0.95, h->param.rc.i_qp_min )
 
697
          : pow( 0.95, h->param.rc.i_qp_constant ) * X264_MAX( 1, h->param.rc.f_ip_factor )));
 
698
 
 
699
    h->thread[0] = h;
 
700
    h->i_thread_num = 0;
 
701
    for( i = 1; i < h->param.i_threads; i++ )
 
702
        h->thread[i] = x264_malloc( sizeof(x264_t) );
 
703
 
 
704
    for( i = 0; i < h->param.i_threads; i++ )
 
705
    {
 
706
        if( i > 0 )
 
707
            *h->thread[i] = *h;
 
708
        h->thread[i]->fdec = x264_frame_pop_unused( h );
 
709
        h->thread[i]->out.p_bitstream = x264_malloc( h->out.i_bitstream );
 
710
        if( x264_macroblock_cache_init( h->thread[i] ) < 0 )
 
711
            return NULL;
 
712
    }
 
713
 
 
714
    if( x264_ratecontrol_new( h ) < 0 )
 
715
        return NULL;
 
716
 
 
717
#ifdef DEBUG_DUMP_FRAME
 
718
    {
 
719
        /* create or truncate the reconstructed video file */
 
720
        FILE *f = fopen( "fdec.yuv", "w" );
 
721
        if( f )
 
722
            fclose( f );
 
723
        else
 
724
        {
 
725
            x264_log( h, X264_LOG_ERROR, "can't write to fdec.yuv\n" );
 
726
            x264_free( h );
 
727
            return NULL;
 
728
        }
 
729
    }
 
730
#endif
 
731
 
 
732
    return h;
 
733
}
 
734
 
 
735
/****************************************************************************
 
736
 * x264_encoder_reconfig:
 
737
 ****************************************************************************/
 
738
int x264_encoder_reconfig( x264_t *h, x264_param_t *param )
 
739
{
 
740
#define COPY(var) h->param.var = param->var
 
741
    COPY( i_frame_reference ); // but never uses more refs than initially specified
 
742
    COPY( i_bframe_bias );
 
743
    if( h->param.i_scenecut_threshold >= 0 && param->i_scenecut_threshold >= 0 )
 
744
        COPY( i_scenecut_threshold ); // can't turn it on or off, only vary the threshold
 
745
    COPY( b_deblocking_filter );
 
746
    COPY( i_deblocking_filter_alphac0 );
 
747
    COPY( i_deblocking_filter_beta );
 
748
    COPY( analyse.intra );
 
749
    COPY( analyse.inter );
 
750
    COPY( analyse.i_direct_mv_pred );
 
751
    COPY( analyse.i_me_method );
 
752
    COPY( analyse.i_me_range );
 
753
    COPY( analyse.i_noise_reduction );
 
754
    COPY( analyse.i_subpel_refine );
 
755
    COPY( analyse.i_trellis );
 
756
    COPY( analyse.b_bidir_me );
 
757
    COPY( analyse.b_bframe_rdo );
 
758
    COPY( analyse.b_chroma_me );
 
759
    COPY( analyse.b_dct_decimate );
 
760
    COPY( analyse.b_fast_pskip );
 
761
    COPY( analyse.b_mixed_references );
 
762
    // can only twiddle these if they were enabled to begin with:
 
763
    if( h->pps->b_transform_8x8_mode )
 
764
        COPY( analyse.b_transform_8x8 );
 
765
    if( h->frames.i_max_ref1 > 1 )
 
766
        COPY( b_bframe_pyramid );
 
767
#undef COPY
 
768
 
 
769
    mbcmp_init( h );
 
770
 
 
771
    return x264_validate_parameters( h );
 
772
}
 
773
 
 
774
/* internal usage */
 
775
static void x264_nal_start( x264_t *h, int i_type, int i_ref_idc )
 
776
{
 
777
    x264_nal_t *nal = &h->out.nal[h->out.i_nal];
 
778
 
 
779
    nal->i_ref_idc = i_ref_idc;
 
780
    nal->i_type    = i_type;
 
781
 
 
782
    nal->i_payload= 0;
 
783
    nal->p_payload= &h->out.p_bitstream[bs_pos( &h->out.bs ) / 8];
 
784
}
 
785
static void x264_nal_end( x264_t *h )
 
786
{
 
787
    x264_nal_t *nal = &h->out.nal[h->out.i_nal];
 
788
    nal->i_payload = &h->out.p_bitstream[bs_pos( &h->out.bs ) / 8] - nal->p_payload;
 
789
    h->out.i_nal++;
 
790
}
 
791
 
 
792
/****************************************************************************
 
793
 * x264_encoder_headers:
 
794
 ****************************************************************************/
 
795
int x264_encoder_headers( x264_t *h, x264_nal_t **pp_nal, int *pi_nal )
 
796
{
 
797
    /* init bitstream context */
 
798
    h->out.i_nal = 0;
 
799
    bs_init( &h->out.bs, h->out.p_bitstream, h->out.i_bitstream );
 
800
 
 
801
    /* Put SPS and PPS */
 
802
    if( h->i_frame == 0 )
 
803
    {
 
804
        /* identify ourself */
 
805
        x264_nal_start( h, NAL_SEI, NAL_PRIORITY_DISPOSABLE );
 
806
        x264_sei_version_write( h, &h->out.bs );
 
807
        x264_nal_end( h );
 
808
 
 
809
        /* generate sequence parameters */
 
810
        x264_nal_start( h, NAL_SPS, NAL_PRIORITY_HIGHEST );
 
811
        x264_sps_write( &h->out.bs, h->sps );
 
812
        x264_nal_end( h );
 
813
 
 
814
        /* generate picture parameters */
 
815
        x264_nal_start( h, NAL_PPS, NAL_PRIORITY_HIGHEST );
 
816
        x264_pps_write( &h->out.bs, h->pps );
 
817
        x264_nal_end( h );
 
818
    }
 
819
    /* now set output*/
 
820
    *pi_nal = h->out.i_nal;
 
821
    *pp_nal = &h->out.nal[0];
 
822
    h->out.i_nal = 0;
 
823
 
 
824
    return 0;
 
825
}
 
826
 
 
827
static inline void x264_reference_build_list( x264_t *h, int i_poc )
 
828
{
 
829
    int i;
 
830
    int b_ok;
 
831
 
 
832
    /* build ref list 0/1 */
 
833
    h->i_ref0 = 0;
 
834
    h->i_ref1 = 0;
 
835
    for( i = 0; h->frames.reference[i]; i++ )
 
836
    {
 
837
        if( h->frames.reference[i]->i_poc < i_poc )
 
838
        {
 
839
            h->fref0[h->i_ref0++] = h->frames.reference[i];
 
840
        }
 
841
        else if( h->frames.reference[i]->i_poc > i_poc )
 
842
        {
 
843
            h->fref1[h->i_ref1++] = h->frames.reference[i];
 
844
        }
 
845
    }
 
846
 
 
847
    /* Order ref0 from higher to lower poc */
 
848
    do
 
849
    {
 
850
        b_ok = 1;
 
851
        for( i = 0; i < h->i_ref0 - 1; i++ )
 
852
        {
 
853
            if( h->fref0[i]->i_poc < h->fref0[i+1]->i_poc )
 
854
            {
 
855
                XCHG( x264_frame_t*, h->fref0[i], h->fref0[i+1] );
 
856
                b_ok = 0;
 
857
                break;
 
858
            }
 
859
        }
 
860
    } while( !b_ok );
 
861
    /* Order ref1 from lower to higher poc (bubble sort) for B-frame */
 
862
    do
 
863
    {
 
864
        b_ok = 1;
 
865
        for( i = 0; i < h->i_ref1 - 1; i++ )
 
866
        {
 
867
            if( h->fref1[i]->i_poc > h->fref1[i+1]->i_poc )
 
868
            {
 
869
                XCHG( x264_frame_t*, h->fref1[i], h->fref1[i+1] );
 
870
                b_ok = 0;
 
871
                break;
 
872
            }
 
873
        }
 
874
    } while( !b_ok );
 
875
 
 
876
    /* In the standard, a P-frame's ref list is sorted by frame_num.
 
877
     * We use POC, but check whether explicit reordering is needed */
 
878
    h->b_ref_reorder[0] =
 
879
    h->b_ref_reorder[1] = 0;
 
880
    if( h->sh.i_type == SLICE_TYPE_P )
 
881
    {
 
882
        for( i = 0; i < h->i_ref0 - 1; i++ )
 
883
            if( h->fref0[i]->i_frame_num < h->fref0[i+1]->i_frame_num )
 
884
            {
 
885
                h->b_ref_reorder[0] = 1;
 
886
                break;
 
887
            }
 
888
    }
 
889
 
 
890
    h->i_ref1 = X264_MIN( h->i_ref1, h->frames.i_max_ref1 );
 
891
    h->i_ref0 = X264_MIN( h->i_ref0, h->frames.i_max_ref0 );
 
892
    h->i_ref0 = X264_MIN( h->i_ref0, h->param.i_frame_reference ); // if reconfig() has lowered the limit
 
893
    assert( h->i_ref0 + h->i_ref1 <= 16 );
 
894
    h->mb.pic.i_fref[0] = h->i_ref0;
 
895
    h->mb.pic.i_fref[1] = h->i_ref1;
 
896
}
 
897
 
 
898
static void x264_fdec_filter_row( x264_t *h, int mb_y )
 
899
{
 
900
    /* mb_y is the mb to be encoded next, not the mb to be filtered here */
 
901
    int b_hpel = h->fdec->b_kept_as_ref;
 
902
    int b_deblock = !h->sh.i_disable_deblocking_filter_idc;
 
903
    int b_end = mb_y == h->sps->i_mb_height;
 
904
    int min_y = mb_y - (1 << h->sh.b_mbaff);
 
905
#ifndef DEBUG_DUMP_FRAME
 
906
    b_deblock &= b_hpel;
 
907
#endif
 
908
    if( mb_y & h->sh.b_mbaff )
 
909
        return;
 
910
    if( min_y < 0 )
 
911
        return;
 
912
 
 
913
    if( !b_end )
 
914
    {
 
915
        int i, j;
 
916
        for( j=0; j<=h->sh.b_mbaff; j++ )
 
917
            for( i=0; i<3; i++ )
 
918
            {
 
919
                memcpy( h->mb.intra_border_backup[j][i],
 
920
                        h->fdec->plane[i] + ((mb_y*16 >> !!i) + j - 1 - h->sh.b_mbaff) * h->fdec->i_stride[i],
 
921
                        h->sps->i_mb_width*16 >> !!i );
 
922
            }
 
923
    }
 
924
 
 
925
    if( b_deblock )
 
926
    {
 
927
        int max_y = b_end ? h->sps->i_mb_height : mb_y;
 
928
        int y;
 
929
        for( y = min_y; y < max_y; y += (1 << h->sh.b_mbaff) )
 
930
            x264_frame_deblock_row( h, y );
 
931
    }
 
932
 
 
933
    if( b_hpel )
 
934
    {
 
935
        x264_frame_expand_border( h, h->fdec, min_y, b_end );
 
936
        x264_frame_filter( h, h->fdec, min_y, b_end );
 
937
        x264_frame_expand_border_filtered( h, h->fdec, min_y, b_end );
 
938
    }
 
939
 
 
940
    if( h->param.i_threads > 1 && h->fdec->b_kept_as_ref )
 
941
    {
 
942
        x264_frame_cond_broadcast( h->fdec, mb_y*16 + (b_end ? 10000 : -(X264_THREAD_HEIGHT << h->sh.b_mbaff)) );
 
943
    }
 
944
}
 
945
 
 
946
static inline void x264_reference_update( x264_t *h )
 
947
{
 
948
    int i;
 
949
 
 
950
    if( h->fdec->i_frame >= 0 )
 
951
        h->i_frame++;
 
952
 
 
953
    if( !h->fdec->b_kept_as_ref )
 
954
    {
 
955
        if( h->param.i_threads > 1 )
 
956
        {
 
957
            x264_frame_push_unused( h, h->fdec );
 
958
            h->fdec = x264_frame_pop_unused( h );
 
959
        }
 
960
        return;
 
961
    }
 
962
 
 
963
    /* move lowres copy of the image to the ref frame */
 
964
    for( i = 0; i < 4; i++)
 
965
    {
 
966
        XCHG( uint8_t*, h->fdec->lowres[i], h->fenc->lowres[i] );
 
967
        XCHG( uint8_t*, h->fdec->buffer_lowres[i], h->fenc->buffer_lowres[i] );
 
968
    }
 
969
 
 
970
    /* adaptive B decision needs a pointer, since it can't use the ref lists */
 
971
    if( h->sh.i_type != SLICE_TYPE_B )
 
972
        h->frames.last_nonb = h->fdec;
 
973
 
 
974
    /* move frame in the buffer */
 
975
    x264_frame_push( h->frames.reference, h->fdec );
 
976
    if( h->frames.reference[h->frames.i_max_dpb] )
 
977
        x264_frame_push_unused( h, x264_frame_shift( h->frames.reference ) );
 
978
    h->fdec = x264_frame_pop_unused( h );
 
979
}
 
980
 
 
981
static inline void x264_reference_reset( x264_t *h )
 
982
{
 
983
    while( h->frames.reference[0] )
 
984
        x264_frame_push_unused( h, x264_frame_pop( h->frames.reference ) );
 
985
    h->fdec->i_poc =
 
986
    h->fenc->i_poc = 0;
 
987
}
 
988
 
 
989
static inline void x264_slice_init( x264_t *h, int i_nal_type, int i_global_qp )
 
990
{
 
991
    /* ------------------------ Create slice header  ----------------------- */
 
992
    if( i_nal_type == NAL_SLICE_IDR )
 
993
    {
 
994
        x264_slice_header_init( h, &h->sh, h->sps, h->pps, h->i_idr_pic_id, h->i_frame_num, i_global_qp );
 
995
 
 
996
        /* increment id */
 
997
        h->i_idr_pic_id = ( h->i_idr_pic_id + 1 ) % 65536;
 
998
    }
 
999
    else
 
1000
    {
 
1001
        x264_slice_header_init( h, &h->sh, h->sps, h->pps, -1, h->i_frame_num, i_global_qp );
 
1002
 
 
1003
        /* always set the real higher num of ref frame used */
 
1004
        h->sh.b_num_ref_idx_override = 1;
 
1005
        h->sh.i_num_ref_idx_l0_active = h->i_ref0 <= 0 ? 1 : h->i_ref0;
 
1006
        h->sh.i_num_ref_idx_l1_active = h->i_ref1 <= 0 ? 1 : h->i_ref1;
 
1007
    }
 
1008
 
 
1009
    h->fdec->i_frame_num = h->sh.i_frame_num;
 
1010
 
 
1011
    if( h->sps->i_poc_type == 0 )
 
1012
    {
 
1013
        h->sh.i_poc_lsb = h->fdec->i_poc & ( (1 << h->sps->i_log2_max_poc_lsb) - 1 );
 
1014
        h->sh.i_delta_poc_bottom = 0;   /* XXX won't work for field */
 
1015
    }
 
1016
    else if( h->sps->i_poc_type == 1 )
 
1017
    {
 
1018
        /* FIXME TODO FIXME */
 
1019
    }
 
1020
    else
 
1021
    {
 
1022
        /* Nothing to do ? */
 
1023
    }
 
1024
 
 
1025
    x264_macroblock_slice_init( h );
 
1026
}
 
1027
 
 
1028
static void x264_slice_write( x264_t *h )
 
1029
{
 
1030
    int i_skip;
 
1031
    int mb_xy;
 
1032
    int i;
 
1033
 
 
1034
    /* init stats */
 
1035
    memset( &h->stat.frame, 0, sizeof(h->stat.frame) );
 
1036
 
 
1037
    /* Slice */
 
1038
    x264_nal_start( h, h->i_nal_type, h->i_nal_ref_idc );
 
1039
 
 
1040
    /* Slice header */
 
1041
    x264_slice_header_write( &h->out.bs, &h->sh, h->i_nal_ref_idc );
 
1042
    if( h->param.b_cabac )
 
1043
    {
 
1044
        /* alignment needed */
 
1045
        bs_align_1( &h->out.bs );
 
1046
 
 
1047
        /* init cabac */
 
1048
        x264_cabac_context_init( &h->cabac, h->sh.i_type, h->sh.i_qp, h->sh.i_cabac_init_idc );
 
1049
        x264_cabac_encode_init ( &h->cabac, h->out.bs.p, h->out.bs.p_end );
 
1050
    }
 
1051
    h->mb.i_last_qp = h->sh.i_qp;
 
1052
    h->mb.i_last_dqp = 0;
 
1053
 
 
1054
    for( mb_xy = h->sh.i_first_mb, i_skip = 0; mb_xy < h->sh.i_last_mb; )
 
1055
    {
 
1056
        const int i_mb_y = mb_xy / h->sps->i_mb_width;
 
1057
        const int i_mb_x = mb_xy % h->sps->i_mb_width;
 
1058
        int mb_spos = bs_pos(&h->out.bs) + x264_cabac_pos(&h->cabac);
 
1059
 
 
1060
        if( i_mb_x == 0 )
 
1061
            x264_fdec_filter_row( h, i_mb_y );
 
1062
 
 
1063
        /* load cache */
 
1064
        x264_macroblock_cache_load( h, i_mb_x, i_mb_y );
 
1065
 
 
1066
        /* analyse parameters
 
1067
         * Slice I: choose I_4x4 or I_16x16 mode
 
1068
         * Slice P: choose between using P mode or intra (4x4 or 16x16)
 
1069
         * */
 
1070
        TIMER_START( i_mtime_analyse );
 
1071
        x264_macroblock_analyse( h );
 
1072
        TIMER_STOP( i_mtime_analyse );
 
1073
 
 
1074
        /* encode this macroblock -> be careful it can change the mb type to P_SKIP if needed */
 
1075
        TIMER_START( i_mtime_encode );
 
1076
        x264_macroblock_encode( h );
 
1077
        TIMER_STOP( i_mtime_encode );
 
1078
 
 
1079
        TIMER_START( i_mtime_write );
 
1080
        if( h->param.b_cabac )
 
1081
        {
 
1082
            if( mb_xy > h->sh.i_first_mb && !(h->sh.b_mbaff && (i_mb_y&1)) )
 
1083
                x264_cabac_encode_terminal( &h->cabac );
 
1084
 
 
1085
            if( IS_SKIP( h->mb.i_type ) )
 
1086
                x264_cabac_mb_skip( h, 1 );
 
1087
            else
 
1088
            {
 
1089
                if( h->sh.i_type != SLICE_TYPE_I )
 
1090
                    x264_cabac_mb_skip( h, 0 );
 
1091
                x264_macroblock_write_cabac( h, &h->cabac );
 
1092
            }
 
1093
        }
 
1094
        else
 
1095
        {
 
1096
            if( IS_SKIP( h->mb.i_type ) )
 
1097
                i_skip++;
 
1098
            else
 
1099
            {
 
1100
                if( h->sh.i_type != SLICE_TYPE_I )
 
1101
                {
 
1102
                    bs_write_ue( &h->out.bs, i_skip );  /* skip run */
 
1103
                    i_skip = 0;
 
1104
                }
 
1105
                x264_macroblock_write_cavlc( h, &h->out.bs );
 
1106
            }
 
1107
        }
 
1108
        TIMER_STOP( i_mtime_write );
 
1109
 
 
1110
#if VISUALIZE
 
1111
        if( h->param.b_visualize )
 
1112
            x264_visualize_mb( h );
 
1113
#endif
 
1114
 
 
1115
        /* save cache */
 
1116
        x264_macroblock_cache_save( h );
 
1117
 
 
1118
        /* accumulate mb stats */
 
1119
        h->stat.frame.i_mb_count[h->mb.i_type]++;
 
1120
        if( !IS_SKIP(h->mb.i_type) && !IS_INTRA(h->mb.i_type) && !IS_DIRECT(h->mb.i_type) )
 
1121
        {
 
1122
            if( h->mb.i_partition != D_8x8 )
 
1123
                h->stat.frame.i_mb_count_size[ x264_mb_partition_pixel_table[ h->mb.i_partition ] ] += 4;
 
1124
            else
 
1125
                for( i = 0; i < 4; i++ )
 
1126
                    h->stat.frame.i_mb_count_size[ x264_mb_partition_pixel_table[ h->mb.i_sub_partition[i] ] ] ++;
 
1127
            if( h->param.i_frame_reference > 1 )
 
1128
            {
 
1129
                for( i = 0; i < 4; i++ )
 
1130
                {
 
1131
                    int i_ref = h->mb.cache.ref[0][ x264_scan8[4*i] ];
 
1132
                    if( i_ref >= 0 )
 
1133
                        h->stat.frame.i_mb_count_ref[i_ref] ++;
 
1134
                }
 
1135
            }
 
1136
        }
 
1137
        if( h->mb.i_cbp_luma && !IS_INTRA(h->mb.i_type) )
 
1138
        {
 
1139
            h->stat.frame.i_mb_count_8x8dct[0] ++;
 
1140
            h->stat.frame.i_mb_count_8x8dct[1] += h->mb.b_transform_8x8;
 
1141
        }
 
1142
 
 
1143
        x264_ratecontrol_mb( h, bs_pos(&h->out.bs) + x264_cabac_pos(&h->cabac) - mb_spos );
 
1144
 
 
1145
        if( h->sh.b_mbaff )
 
1146
        {
 
1147
            if( (i_mb_y&1) && i_mb_x == h->sps->i_mb_width - 1 )
 
1148
                mb_xy++;
 
1149
            else if( i_mb_y&1 )
 
1150
                mb_xy += 1 - h->sps->i_mb_width;
 
1151
            else
 
1152
                mb_xy += h->sps->i_mb_width;
 
1153
        }
 
1154
        else
 
1155
            mb_xy++;
 
1156
    }
 
1157
 
 
1158
    if( h->param.b_cabac )
 
1159
    {
 
1160
        x264_cabac_encode_flush( h, &h->cabac );
 
1161
        h->out.bs.p = h->cabac.p;
 
1162
    }
 
1163
    else
 
1164
    {
 
1165
        if( i_skip > 0 )
 
1166
            bs_write_ue( &h->out.bs, i_skip );  /* last skip run */
 
1167
        /* rbsp_slice_trailing_bits */
 
1168
        bs_rbsp_trailing( &h->out.bs );
 
1169
    }
 
1170
 
 
1171
    x264_nal_end( h );
 
1172
 
 
1173
    /* Compute misc bits */
 
1174
    h->stat.frame.i_misc_bits = bs_pos( &h->out.bs )
 
1175
                              + NALU_OVERHEAD * 8
 
1176
                              - h->stat.frame.i_itex_bits
 
1177
                              - h->stat.frame.i_ptex_bits
 
1178
                              - h->stat.frame.i_hdr_bits;
 
1179
}
 
1180
 
 
1181
static void x264_thread_sync_context( x264_t *dst, x264_t *src )
 
1182
{
 
1183
    x264_frame_t **f;
 
1184
    if( dst == src )
 
1185
        return;
 
1186
 
 
1187
    // reference counting
 
1188
    for( f = src->frames.reference; *f; f++ )
 
1189
        (*f)->i_reference_count++;
 
1190
    for( f = dst->frames.reference; *f; f++ )
 
1191
        x264_frame_push_unused( src, *f );
 
1192
    src->fdec->i_reference_count++;
 
1193
    x264_frame_push_unused( src, dst->fdec );
 
1194
 
 
1195
    // copy everything except the per-thread pointers and the constants.
 
1196
    memcpy( &dst->i_frame, &src->i_frame, offsetof(x264_t, mb.type) - offsetof(x264_t, i_frame) );
 
1197
    memcpy( &dst->mb.i_type, &src->mb.i_type, offsetof(x264_t, rc) - offsetof(x264_t, mb.i_type) );
 
1198
    dst->stat = src->stat;
 
1199
}
 
1200
 
 
1201
static void x264_thread_sync_stat( x264_t *dst, x264_t *src )
 
1202
{
 
1203
    if( dst == src )
 
1204
        return;
 
1205
    memcpy( &dst->stat.i_slice_count, &src->stat.i_slice_count, sizeof(dst->stat) - sizeof(dst->stat.frame) );
 
1206
}
 
1207
 
 
1208
static int x264_slices_write( x264_t *h )
 
1209
{
 
1210
    int i_frame_size;
 
1211
 
 
1212
#if VISUALIZE
 
1213
    if( h->param.b_visualize )
 
1214
        x264_visualize_init( h );
 
1215
#endif
 
1216
 
 
1217
    x264_stack_align( x264_slice_write, h );
 
1218
    i_frame_size = h->out.nal[h->out.i_nal-1].i_payload;
 
1219
    x264_fdec_filter_row( h, h->sps->i_mb_height );
 
1220
 
 
1221
#if VISUALIZE
 
1222
    if( h->param.b_visualize )
 
1223
    {
 
1224
        x264_visualize_show( h );
 
1225
        x264_visualize_close( h );
 
1226
    }
 
1227
#endif
 
1228
 
 
1229
    h->out.i_frame_size = i_frame_size;
 
1230
    return 0;
 
1231
}
 
1232
 
 
1233
/****************************************************************************
 
1234
 * x264_encoder_encode:
 
1235
 *  XXX: i_poc   : is the poc of the current given picture
 
1236
 *       i_frame : is the number of the frame being coded
 
1237
 *  ex:  type frame poc
 
1238
 *       I      0   2*0
 
1239
 *       P      1   2*3
 
1240
 *       B      2   2*1
 
1241
 *       B      3   2*2
 
1242
 *       P      4   2*6
 
1243
 *       B      5   2*4
 
1244
 *       B      6   2*5
 
1245
 ****************************************************************************/
 
1246
int     x264_encoder_encode( x264_t *h,
 
1247
                             x264_nal_t **pp_nal, int *pi_nal,
 
1248
                             x264_picture_t *pic_in,
 
1249
                             x264_picture_t *pic_out )
 
1250
{
 
1251
    x264_t *thread_current, *thread_prev, *thread_oldest;
 
1252
    int     i_nal_type;
 
1253
    int     i_nal_ref_idc;
 
1254
 
 
1255
    int   i_global_qp;
 
1256
 
 
1257
    if( h->param.i_threads > 1)
 
1258
    {
 
1259
        int i = ++h->i_thread_phase;
 
1260
        int t = h->param.i_threads;
 
1261
        thread_current = h->thread[ i%t ];
 
1262
        thread_prev    = h->thread[ (i-1)%t ];
 
1263
        thread_oldest  = h->thread[ (i+1)%t ];
 
1264
        x264_thread_sync_context( thread_current, thread_prev );
 
1265
        x264_thread_sync_ratecontrol( thread_current, thread_prev, thread_oldest );
 
1266
        h = thread_current;
 
1267
//      fprintf(stderr, "current: %p  prev: %p  oldest: %p \n", thread_current, thread_prev, thread_oldest);
 
1268
    }
 
1269
    else
 
1270
    {
 
1271
        thread_current =
 
1272
        thread_prev    =
 
1273
        thread_oldest  = h;
 
1274
    }
 
1275
 
 
1276
    // ok to call this before encoding any frames, since the initial values of fdec have b_kept_as_ref=0
 
1277
    x264_reference_update( h );
 
1278
    h->fdec->i_lines_completed = -1;
 
1279
 
 
1280
    /* no data out */
 
1281
    *pi_nal = 0;
 
1282
    *pp_nal = NULL;
 
1283
 
 
1284
    /* ------------------- Setup new frame from picture -------------------- */
 
1285
    TIMER_START( i_mtime_encode_frame );
 
1286
    if( pic_in != NULL )
 
1287
    {
 
1288
        /* 1: Copy the picture to a frame and move it to a buffer */
 
1289
        x264_frame_t *fenc = x264_frame_pop_unused( h );
 
1290
 
 
1291
        if( x264_frame_copy_picture( h, fenc, pic_in ) < 0 )
 
1292
            return -1;
 
1293
 
 
1294
        if( h->param.i_width != 16 * h->sps->i_mb_width ||
 
1295
            h->param.i_height != 16 * h->sps->i_mb_height )
 
1296
            x264_frame_expand_border_mod16( h, fenc );
 
1297
 
 
1298
        fenc->i_frame = h->frames.i_input++;
 
1299
 
 
1300
        x264_frame_push( h->frames.next, fenc );
 
1301
 
 
1302
        if( h->frames.b_have_lowres )
 
1303
            x264_frame_init_lowres( h, fenc );
 
1304
 
 
1305
        if( h->frames.i_input <= h->frames.i_delay + 1 - h->param.i_threads )
 
1306
        {
 
1307
            /* Nothing yet to encode */
 
1308
            /* waiting for filling bframe buffer */
 
1309
            pic_out->i_type = X264_TYPE_AUTO;
 
1310
            return 0;
 
1311
        }
 
1312
    }
 
1313
 
 
1314
    if( h->frames.current[0] == NULL )
 
1315
    {
 
1316
        int bframes = 0;
 
1317
        /* 2: Select frame types */
 
1318
        if( h->frames.next[0] == NULL )
 
1319
        {
 
1320
            x264_encoder_frame_end( thread_oldest, thread_current, pp_nal, pi_nal, pic_out );
 
1321
            return 0;
 
1322
        }
 
1323
 
 
1324
        x264_slicetype_decide( h );
 
1325
 
 
1326
        /* 3: move some B-frames and 1 non-B to encode queue */
 
1327
        while( IS_X264_TYPE_B( h->frames.next[bframes]->i_type ) )
 
1328
            bframes++;
 
1329
        x264_frame_push( h->frames.current, x264_frame_shift( &h->frames.next[bframes] ) );
 
1330
        /* FIXME: when max B-frames > 3, BREF may no longer be centered after GOP closing */
 
1331
        if( h->param.b_bframe_pyramid && bframes > 1 )
 
1332
        {
 
1333
            x264_frame_t *mid = x264_frame_shift( &h->frames.next[bframes/2] );
 
1334
            mid->i_type = X264_TYPE_BREF;
 
1335
            x264_frame_push( h->frames.current, mid );
 
1336
            bframes--;
 
1337
        }
 
1338
        while( bframes-- )
 
1339
            x264_frame_push( h->frames.current, x264_frame_shift( h->frames.next ) );
 
1340
    }
 
1341
    TIMER_STOP( i_mtime_encode_frame );
 
1342
 
 
1343
    /* ------------------- Get frame to be encoded ------------------------- */
 
1344
    /* 4: get picture to encode */
 
1345
    h->fenc = x264_frame_shift( h->frames.current );
 
1346
    if( h->fenc == NULL )
 
1347
    {
 
1348
        /* Nothing yet to encode (ex: waiting for I/P with B frames) */
 
1349
        /* waiting for filling bframe buffer */
 
1350
        pic_out->i_type = X264_TYPE_AUTO;
 
1351
        return 0;
 
1352
    }
 
1353
 
 
1354
do_encode:
 
1355
 
 
1356
    if( h->fenc->i_type == X264_TYPE_IDR )
 
1357
    {
 
1358
        h->frames.i_last_idr = h->fenc->i_frame;
 
1359
    }
 
1360
 
 
1361
    /* ------------------- Setup frame context ----------------------------- */
 
1362
    /* 5: Init data dependent of frame type */
 
1363
    TIMER_START( i_mtime_encode_frame );
 
1364
    if( h->fenc->i_type == X264_TYPE_IDR )
 
1365
    {
 
1366
        /* reset ref pictures */
 
1367
        x264_reference_reset( h );
 
1368
 
 
1369
        i_nal_type    = NAL_SLICE_IDR;
 
1370
        i_nal_ref_idc = NAL_PRIORITY_HIGHEST;
 
1371
        h->sh.i_type = SLICE_TYPE_I;
 
1372
    }
 
1373
    else if( h->fenc->i_type == X264_TYPE_I )
 
1374
    {
 
1375
        i_nal_type    = NAL_SLICE;
 
1376
        i_nal_ref_idc = NAL_PRIORITY_HIGH; /* Not completely true but for now it is (as all I/P are kept as ref)*/
 
1377
        h->sh.i_type = SLICE_TYPE_I;
 
1378
    }
 
1379
    else if( h->fenc->i_type == X264_TYPE_P )
 
1380
    {
 
1381
        i_nal_type    = NAL_SLICE;
 
1382
        i_nal_ref_idc = NAL_PRIORITY_HIGH; /* Not completely true but for now it is (as all I/P are kept as ref)*/
 
1383
        h->sh.i_type = SLICE_TYPE_P;
 
1384
    }
 
1385
    else if( h->fenc->i_type == X264_TYPE_BREF )
 
1386
    {
 
1387
        i_nal_type    = NAL_SLICE;
 
1388
        i_nal_ref_idc = NAL_PRIORITY_HIGH; /* maybe add MMCO to forget it? -> low */
 
1389
        h->sh.i_type = SLICE_TYPE_B;
 
1390
    }
 
1391
    else    /* B frame */
 
1392
    {
 
1393
        i_nal_type    = NAL_SLICE;
 
1394
        i_nal_ref_idc = NAL_PRIORITY_DISPOSABLE;
 
1395
        h->sh.i_type = SLICE_TYPE_B;
 
1396
    }
 
1397
 
 
1398
    h->fdec->i_poc =
 
1399
    h->fenc->i_poc = 2 * (h->fenc->i_frame - h->frames.i_last_idr);
 
1400
    h->fdec->i_type = h->fenc->i_type;
 
1401
    h->fdec->i_frame = h->fenc->i_frame;
 
1402
    h->fenc->b_kept_as_ref =
 
1403
    h->fdec->b_kept_as_ref = i_nal_ref_idc != NAL_PRIORITY_DISPOSABLE && h->param.i_keyint_max > 1;
 
1404
 
 
1405
 
 
1406
 
 
1407
    /* ------------------- Init                ----------------------------- */
 
1408
    /* build ref list 0/1 */
 
1409
    x264_reference_build_list( h, h->fdec->i_poc );
 
1410
 
 
1411
    /* Init the rate control */
 
1412
    x264_ratecontrol_start( h, h->fenc->i_qpplus1 );
 
1413
    i_global_qp = x264_ratecontrol_qp( h );
 
1414
 
 
1415
    pic_out->i_qpplus1 =
 
1416
    h->fdec->i_qpplus1 = i_global_qp + 1;
 
1417
 
 
1418
    if( h->sh.i_type == SLICE_TYPE_B )
 
1419
        x264_macroblock_bipred_init( h );
 
1420
 
 
1421
    /* ------------------------ Create slice header  ----------------------- */
 
1422
    x264_slice_init( h, i_nal_type, i_global_qp );
 
1423
 
 
1424
    if( i_nal_ref_idc != NAL_PRIORITY_DISPOSABLE )
 
1425
        h->i_frame_num++;
 
1426
 
 
1427
    /* ---------------------- Write the bitstream -------------------------- */
 
1428
    /* Init bitstream context */
 
1429
    h->out.i_nal = 0;
 
1430
    bs_init( &h->out.bs, h->out.p_bitstream, h->out.i_bitstream );
 
1431
 
 
1432
    if(h->param.b_aud){
 
1433
        int pic_type;
 
1434
 
 
1435
        if(h->sh.i_type == SLICE_TYPE_I)
 
1436
            pic_type = 0;
 
1437
        else if(h->sh.i_type == SLICE_TYPE_P)
 
1438
            pic_type = 1;
 
1439
        else if(h->sh.i_type == SLICE_TYPE_B)
 
1440
            pic_type = 2;
 
1441
        else
 
1442
            pic_type = 7;
 
1443
 
 
1444
        x264_nal_start(h, NAL_AUD, NAL_PRIORITY_DISPOSABLE);
 
1445
        bs_write(&h->out.bs, 3, pic_type);
 
1446
        bs_rbsp_trailing(&h->out.bs);
 
1447
        x264_nal_end(h);
 
1448
    }
 
1449
 
 
1450
    h->i_nal_type = i_nal_type;
 
1451
    h->i_nal_ref_idc = i_nal_ref_idc;
 
1452
 
 
1453
    /* Write SPS and PPS */
 
1454
    if( i_nal_type == NAL_SLICE_IDR && h->param.b_repeat_headers )
 
1455
    {
 
1456
        if( h->fenc->i_frame == 0 )
 
1457
        {
 
1458
            /* identify ourself */
 
1459
            x264_nal_start( h, NAL_SEI, NAL_PRIORITY_DISPOSABLE );
 
1460
            x264_sei_version_write( h, &h->out.bs );
 
1461
            x264_nal_end( h );
 
1462
        }
 
1463
 
 
1464
        /* generate sequence parameters */
 
1465
        x264_nal_start( h, NAL_SPS, NAL_PRIORITY_HIGHEST );
 
1466
        x264_sps_write( &h->out.bs, h->sps );
 
1467
        x264_nal_end( h );
 
1468
 
 
1469
        /* generate picture parameters */
 
1470
        x264_nal_start( h, NAL_PPS, NAL_PRIORITY_HIGHEST );
 
1471
        x264_pps_write( &h->out.bs, h->pps );
 
1472
        x264_nal_end( h );
 
1473
    }
 
1474
 
 
1475
    /* Write frame */
 
1476
    if( h->param.i_threads > 1 )
 
1477
    {
 
1478
        x264_pthread_create( &h->thread_handle, NULL, (void*)x264_slices_write, h );
 
1479
        h->b_thread_active = 1;
 
1480
    }
 
1481
    else
 
1482
        x264_slices_write( h );
 
1483
 
 
1484
    /* restore CPU state (before using float again) */
 
1485
    x264_cpu_restore( h->param.cpu );
 
1486
 
 
1487
    if( h->sh.i_type == SLICE_TYPE_P && !h->param.rc.b_stat_read 
 
1488
        && h->param.i_scenecut_threshold >= 0
 
1489
        && !h->param.b_pre_scenecut )
 
1490
    {
 
1491
        const int *mbs = h->stat.frame.i_mb_count;
 
1492
        int i_mb_i = mbs[I_16x16] + mbs[I_8x8] + mbs[I_4x4];
 
1493
        int i_mb_p = mbs[P_L0] + mbs[P_8x8];
 
1494
        int i_mb_s = mbs[P_SKIP];
 
1495
        int i_mb   = h->sps->i_mb_width * h->sps->i_mb_height;
 
1496
        int64_t i_inter_cost = h->stat.frame.i_inter_cost;
 
1497
        int64_t i_intra_cost = h->stat.frame.i_intra_cost;
 
1498
 
 
1499
        float f_bias;
 
1500
        int i_gop_size = h->fenc->i_frame - h->frames.i_last_idr;
 
1501
        float f_thresh_max = h->param.i_scenecut_threshold / 100.0;
 
1502
        /* magic numbers pulled out of thin air */
 
1503
        float f_thresh_min = f_thresh_max * h->param.i_keyint_min
 
1504
                             / ( h->param.i_keyint_max * 4 );
 
1505
        if( h->param.i_keyint_min == h->param.i_keyint_max )
 
1506
             f_thresh_min= f_thresh_max;
 
1507
 
 
1508
        /* macroblock_analyse() doesn't further analyse skipped mbs,
 
1509
         * so we have to guess their cost */
 
1510
        if( h->stat.frame.i_mbs_analysed > 0 )
 
1511
            i_intra_cost = i_intra_cost * i_mb / h->stat.frame.i_mbs_analysed;
 
1512
 
 
1513
        if( i_gop_size < h->param.i_keyint_min / 4 )
 
1514
            f_bias = f_thresh_min / 4;
 
1515
        else if( i_gop_size <= h->param.i_keyint_min )
 
1516
            f_bias = f_thresh_min * i_gop_size / h->param.i_keyint_min;
 
1517
        else
 
1518
        {
 
1519
            f_bias = f_thresh_min
 
1520
                     + ( f_thresh_max - f_thresh_min )
 
1521
                       * ( i_gop_size - h->param.i_keyint_min )
 
1522
                       / ( h->param.i_keyint_max - h->param.i_keyint_min );
 
1523
        }
 
1524
        f_bias = X264_MIN( f_bias, 1.0 );
 
1525
 
 
1526
        /* Bad P will be reencoded as I */
 
1527
        if( h->stat.frame.i_mbs_analysed > 0 &&
 
1528
            i_inter_cost >= (1.0 - f_bias) * i_intra_cost )
 
1529
        {
 
1530
            int b;
 
1531
 
 
1532
            x264_log( h, X264_LOG_DEBUG, "scene cut at %d Icost:%.0f Pcost:%.0f ratio:%.4f bias:%.4f gop:%d (imb:%d pmb:%d smb:%d)\n",
 
1533
                      h->fenc->i_frame,
 
1534
                      (double)i_intra_cost, (double)i_inter_cost,
 
1535
                      1. - (double)i_inter_cost / i_intra_cost,
 
1536
                      f_bias, i_gop_size,
 
1537
                      i_mb_i, i_mb_p, i_mb_s );
 
1538
 
 
1539
            /* Restore frame num */
 
1540
            h->i_frame_num--;
 
1541
 
 
1542
            for( b = 0; h->frames.current[b] && IS_X264_TYPE_B( h->frames.current[b]->i_type ); b++ );
 
1543
            if( b > 0 )
 
1544
            {
 
1545
                /* If using B-frames, force GOP to be closed.
 
1546
                 * Even if this frame is going to be I and not IDR, forcing a
 
1547
                 * P-frame before the scenecut will probably help compression.
 
1548
                 * 
 
1549
                 * We don't yet know exactly which frame is the scene cut, so
 
1550
                 * we can't assign an I-frame. Instead, change the previous
 
1551
                 * B-frame to P, and rearrange coding order. */
 
1552
 
 
1553
                if( h->param.b_bframe_adaptive || b > 1 )
 
1554
                    h->fenc->i_type = X264_TYPE_AUTO;
 
1555
                x264_frame_sort_pts( h->frames.current );
 
1556
                x264_frame_unshift( h->frames.next, h->fenc );
 
1557
                h->fenc = h->frames.current[b-1];
 
1558
                h->frames.current[b-1] = NULL;
 
1559
                h->fenc->i_type = X264_TYPE_P;
 
1560
                x264_frame_sort_dts( h->frames.current );
 
1561
            }
 
1562
            /* Do IDR if needed */
 
1563
            else if( i_gop_size >= h->param.i_keyint_min )
 
1564
            {
 
1565
                /* Reset */
 
1566
                h->i_frame_num = 0;
 
1567
 
 
1568
                /* Reinit field of fenc */
 
1569
                h->fenc->i_type = X264_TYPE_IDR;
 
1570
                h->fenc->i_poc = 0;
 
1571
 
 
1572
                /* Put enqueued frames back in the pool */
 
1573
                while( h->frames.current[0] )
 
1574
                    x264_frame_push( h->frames.next, x264_frame_shift( h->frames.current ) );
 
1575
                x264_frame_sort_pts( h->frames.next );
 
1576
            }
 
1577
            else
 
1578
            {
 
1579
                h->fenc->i_type = X264_TYPE_I;
 
1580
            }
 
1581
            goto do_encode;
 
1582
        }
 
1583
    }
 
1584
 
 
1585
    x264_encoder_frame_end( thread_oldest, thread_current, pp_nal, pi_nal, pic_out );
 
1586
    return 0;
 
1587
}
 
1588
 
 
1589
static void x264_encoder_frame_end( x264_t *h, x264_t *thread_current,
 
1590
                                    x264_nal_t **pp_nal, int *pi_nal,
 
1591
                                    x264_picture_t *pic_out )
 
1592
{
 
1593
    int i;
 
1594
    char psz_message[80];
 
1595
 
 
1596
    if( h->b_thread_active )
 
1597
    {
 
1598
        x264_pthread_join( h->thread_handle, NULL );
 
1599
        h->b_thread_active = 0;
 
1600
    }
 
1601
    if( !h->out.i_nal )
 
1602
    {
 
1603
        pic_out->i_type = X264_TYPE_AUTO;
 
1604
        return;
 
1605
    }
 
1606
 
 
1607
    x264_frame_push_unused( thread_current, h->fenc );
 
1608
 
 
1609
    /* End bitstream, set output  */
 
1610
    *pi_nal = h->out.i_nal;
 
1611
    *pp_nal = h->out.nal;
 
1612
    h->out.i_nal = 0;
 
1613
 
 
1614
    /* Set output picture properties */
 
1615
    if( h->sh.i_type == SLICE_TYPE_I )
 
1616
        pic_out->i_type = h->i_nal_type == NAL_SLICE_IDR ? X264_TYPE_IDR : X264_TYPE_I;
 
1617
    else if( h->sh.i_type == SLICE_TYPE_P )
 
1618
        pic_out->i_type = X264_TYPE_P;
 
1619
    else
 
1620
        pic_out->i_type = X264_TYPE_B;
 
1621
    pic_out->i_pts = h->fenc->i_pts;
 
1622
 
 
1623
    pic_out->img.i_plane = h->fdec->i_plane;
 
1624
    for(i = 0; i < 4; i++){
 
1625
        pic_out->img.i_stride[i] = h->fdec->i_stride[i];
 
1626
        pic_out->img.plane[i] = h->fdec->plane[i];
 
1627
    }
 
1628
 
 
1629
    /* ---------------------- Update encoder state ------------------------- */
 
1630
 
 
1631
    /* update rc */
 
1632
    x264_cpu_restore( h->param.cpu );
 
1633
    x264_ratecontrol_end( h, h->out.i_frame_size * 8 );
 
1634
 
 
1635
    /* restore CPU state (before using float again) */
 
1636
    x264_cpu_restore( h->param.cpu );
 
1637
 
 
1638
    x264_noise_reduction_update( h );
 
1639
 
 
1640
    TIMER_STOP( i_mtime_encode_frame );
 
1641
 
 
1642
    /* ---------------------- Compute/Print statistics --------------------- */
 
1643
    x264_thread_sync_stat( h, h->thread[0] );
 
1644
 
 
1645
    /* Slice stat */
 
1646
    h->stat.i_slice_count[h->sh.i_type]++;
 
1647
    h->stat.i_slice_size[h->sh.i_type] += h->out.i_frame_size + NALU_OVERHEAD;
 
1648
    h->stat.f_slice_qp[h->sh.i_type] += h->fdec->f_qp_avg_aq;
 
1649
 
 
1650
    for( i = 0; i < X264_MBTYPE_MAX; i++ )
 
1651
        h->stat.i_mb_count[h->sh.i_type][i] += h->stat.frame.i_mb_count[i];
 
1652
    for( i = 0; i < 2; i++ )
 
1653
        h->stat.i_mb_count_8x8dct[i] += h->stat.frame.i_mb_count_8x8dct[i];
 
1654
    if( h->sh.i_type != SLICE_TYPE_I )
 
1655
    {
 
1656
        for( i = 0; i < 7; i++ )
 
1657
            h->stat.i_mb_count_size[h->sh.i_type][i] += h->stat.frame.i_mb_count_size[i];
 
1658
        for( i = 0; i < 32; i++ )
 
1659
            h->stat.i_mb_count_ref[h->sh.i_type][i] += h->stat.frame.i_mb_count_ref[i];
 
1660
    }
 
1661
    if( h->sh.i_type == SLICE_TYPE_B )
 
1662
    {
 
1663
        h->stat.i_direct_frames[ h->sh.b_direct_spatial_mv_pred ] ++;
 
1664
        if( h->mb.b_direct_auto_write )
 
1665
        {
 
1666
            //FIXME somewhat arbitrary time constants
 
1667
            if( h->stat.i_direct_score[0] + h->stat.i_direct_score[1] > h->mb.i_mb_count )
 
1668
            {
 
1669
                for( i = 0; i < 2; i++ )
 
1670
                    h->stat.i_direct_score[i] = h->stat.i_direct_score[i] * 9/10;
 
1671
            }
 
1672
            for( i = 0; i < 2; i++ )
 
1673
                h->stat.i_direct_score[i] += h->stat.frame.i_direct_score[i];
 
1674
        }
 
1675
    }
 
1676
 
 
1677
    psz_message[0] = '\0';
 
1678
    if( h->param.analyse.b_psnr )
 
1679
    {
 
1680
        int64_t sqe[3];
 
1681
 
 
1682
        for( i=0; i<3; i++ )
 
1683
        {
 
1684
            sqe[i] = x264_pixel_ssd_wxh( &h->pixf,
 
1685
                         h->fdec->plane[i], h->fdec->i_stride[i],
 
1686
                         h->fenc->plane[i], h->fenc->i_stride[i],
 
1687
                         h->param.i_width >> !!i, h->param.i_height >> !!i );
 
1688
        }
 
1689
        x264_cpu_restore( h->param.cpu );
 
1690
 
 
1691
        h->stat.i_sqe_global[h->sh.i_type] += sqe[0] + sqe[1] + sqe[2];
 
1692
        h->stat.f_psnr_average[h->sh.i_type] += x264_psnr( sqe[0] + sqe[1] + sqe[2], 3 * h->param.i_width * h->param.i_height / 2 );
 
1693
        h->stat.f_psnr_mean_y[h->sh.i_type] += x264_psnr( sqe[0], h->param.i_width * h->param.i_height );
 
1694
        h->stat.f_psnr_mean_u[h->sh.i_type] += x264_psnr( sqe[1], h->param.i_width * h->param.i_height / 4 );
 
1695
        h->stat.f_psnr_mean_v[h->sh.i_type] += x264_psnr( sqe[2], h->param.i_width * h->param.i_height / 4 );
 
1696
 
 
1697
        snprintf( psz_message, 80, " PSNR Y:%5.2f U:%5.2f V:%5.2f",
 
1698
                  x264_psnr( sqe[0], h->param.i_width * h->param.i_height ),
 
1699
                  x264_psnr( sqe[1], h->param.i_width * h->param.i_height / 4),
 
1700
                  x264_psnr( sqe[2], h->param.i_width * h->param.i_height / 4) );
 
1701
    }
 
1702
 
 
1703
    if( h->param.analyse.b_ssim )
 
1704
    {
 
1705
        // offset by 2 pixels to avoid alignment of ssim blocks with dct blocks
 
1706
        float ssim_y = x264_pixel_ssim_wxh( &h->pixf,
 
1707
                         h->fdec->plane[0] + 2+2*h->fdec->i_stride[0], h->fdec->i_stride[0],
 
1708
                         h->fenc->plane[0] + 2+2*h->fenc->i_stride[0], h->fenc->i_stride[0],
 
1709
                         h->param.i_width-2, h->param.i_height-2 );
 
1710
        h->stat.f_ssim_mean_y[h->sh.i_type] += ssim_y;
 
1711
        snprintf( psz_message + strlen(psz_message), 80 - strlen(psz_message),
 
1712
                  " SSIM Y:%.5f", ssim_y );
 
1713
    }
 
1714
    psz_message[79] = '\0';
 
1715
    
 
1716
    x264_log( h, X264_LOG_DEBUG,
 
1717
                  "frame=%4d QP=%.2f NAL=%d Slice:%c Poc:%-3d I:%-4d P:%-4d SKIP:%-4d size=%d bytes%s\n",
 
1718
              h->i_frame,
 
1719
              h->fdec->f_qp_avg_aq,
 
1720
              h->i_nal_ref_idc,
 
1721
              h->sh.i_type == SLICE_TYPE_I ? 'I' : (h->sh.i_type == SLICE_TYPE_P ? 'P' : 'B' ),
 
1722
              h->fdec->i_poc,
 
1723
              h->stat.frame.i_mb_count_i,
 
1724
              h->stat.frame.i_mb_count_p,
 
1725
              h->stat.frame.i_mb_count_skip,
 
1726
              h->out.i_frame_size,
 
1727
              psz_message );
 
1728
 
 
1729
    // keep stats all in one place
 
1730
    x264_thread_sync_stat( h->thread[0], h );
 
1731
    // for the use of the next frame
 
1732
    x264_thread_sync_stat( thread_current, h );
 
1733
 
 
1734
#ifdef DEBUG_MB_TYPE
 
1735
{
 
1736
    static const char mb_chars[] = { 'i', 'i', 'I', 'C', 'P', '8', 'S',
 
1737
        'D', '<', 'X', 'B', 'X', '>', 'B', 'B', 'B', 'B', '8', 'S' };
 
1738
    int mb_xy;
 
1739
    for( mb_xy = 0; mb_xy < h->sps->i_mb_width * h->sps->i_mb_height; mb_xy++ )
 
1740
    {
 
1741
        if( h->mb.type[mb_xy] < X264_MBTYPE_MAX && h->mb.type[mb_xy] >= 0 )
 
1742
            fprintf( stderr, "%c ", mb_chars[ h->mb.type[mb_xy] ] );
 
1743
        else
 
1744
            fprintf( stderr, "? " );
 
1745
 
 
1746
        if( (mb_xy+1) % h->sps->i_mb_width == 0 )
 
1747
            fprintf( stderr, "\n" );
 
1748
    }
 
1749
}
 
1750
#endif
 
1751
 
 
1752
#ifdef DEBUG_DUMP_FRAME
 
1753
    /* Dump reconstructed frame */
 
1754
    x264_frame_dump( h, h->fdec, "fdec.yuv" );
 
1755
#endif
 
1756
}
 
1757
 
 
1758
/****************************************************************************
 
1759
 * x264_encoder_close:
 
1760
 ****************************************************************************/
 
1761
void    x264_encoder_close  ( x264_t *h )
 
1762
{
 
1763
#ifdef DEBUG_BENCHMARK
 
1764
    int64_t i_mtime_total = i_mtime_analyse + i_mtime_encode + i_mtime_write + i_mtime_filter + 1;
 
1765
#endif
 
1766
    int64_t i_yuv_size = 3 * h->param.i_width * h->param.i_height / 2;
 
1767
    int i;
 
1768
 
 
1769
    for( i=0; i<h->param.i_threads; i++ )
 
1770
    {
 
1771
        // don't strictly have to wait for the other threads, but it's simpler than cancelling them
 
1772
        if( h->thread[i]->b_thread_active )
 
1773
            x264_pthread_join( h->thread[i]->thread_handle, NULL );
 
1774
    }
 
1775
 
 
1776
#ifdef DEBUG_BENCHMARK
 
1777
    x264_log( h, X264_LOG_INFO,
 
1778
              "analyse=%d(%lldms) encode=%d(%lldms) write=%d(%lldms) filter=%d(%lldms)\n",
 
1779
              (int)(100*i_mtime_analyse/i_mtime_total), i_mtime_analyse/1000,
 
1780
              (int)(100*i_mtime_encode/i_mtime_total), i_mtime_encode/1000,
 
1781
              (int)(100*i_mtime_write/i_mtime_total), i_mtime_write/1000,
 
1782
              (int)(100*i_mtime_filter/i_mtime_total), i_mtime_filter/1000 );
 
1783
#endif
 
1784
 
 
1785
    /* Slices used and PSNR */
 
1786
    for( i=0; i<5; i++ )
 
1787
    {
 
1788
        static const int slice_order[] = { SLICE_TYPE_I, SLICE_TYPE_SI, SLICE_TYPE_P, SLICE_TYPE_SP, SLICE_TYPE_B };
 
1789
        static const char *slice_name[] = { "P", "B", "I", "SP", "SI" };
 
1790
        int i_slice = slice_order[i];
 
1791
 
 
1792
        if( h->stat.i_slice_count[i_slice] > 0 )
 
1793
        {
 
1794
            const int i_count = h->stat.i_slice_count[i_slice];
 
1795
            if( h->param.analyse.b_psnr )
 
1796
            {
 
1797
                x264_log( h, X264_LOG_INFO,
 
1798
                          "slice %s:%-5d Avg QP:%5.2f  size:%6.0f  PSNR Mean Y:%5.2f U:%5.2f V:%5.2f Avg:%5.2f Global:%5.2f\n",
 
1799
                          slice_name[i_slice],
 
1800
                          i_count,
 
1801
                          h->stat.f_slice_qp[i_slice] / i_count,
 
1802
                          (double)h->stat.i_slice_size[i_slice] / i_count,
 
1803
                          h->stat.f_psnr_mean_y[i_slice] / i_count, h->stat.f_psnr_mean_u[i_slice] / i_count, h->stat.f_psnr_mean_v[i_slice] / i_count,
 
1804
                          h->stat.f_psnr_average[i_slice] / i_count,
 
1805
                          x264_psnr( h->stat.i_sqe_global[i_slice], i_count * i_yuv_size ) );
 
1806
            }
 
1807
            else
 
1808
            {
 
1809
                x264_log( h, X264_LOG_INFO,
 
1810
                          "slice %s:%-5d Avg QP:%5.2f  size:%6.0f\n",
 
1811
                          slice_name[i_slice],
 
1812
                          i_count,
 
1813
                          h->stat.f_slice_qp[i_slice] / i_count,
 
1814
                          (double)h->stat.i_slice_size[i_slice] / i_count );
 
1815
            }
 
1816
        }
 
1817
    }
 
1818
 
 
1819
    /* MB types used */
 
1820
    if( h->stat.i_slice_count[SLICE_TYPE_I] > 0 )
 
1821
    {
 
1822
        const int64_t *i_mb_count = h->stat.i_mb_count[SLICE_TYPE_I];
 
1823
        const double i_count = h->stat.i_slice_count[SLICE_TYPE_I] * h->mb.i_mb_count / 100.0;
 
1824
        x264_log( h, X264_LOG_INFO,
 
1825
                  "mb I  I16..4: %4.1f%% %4.1f%% %4.1f%%\n",
 
1826
                  i_mb_count[I_16x16]/ i_count,
 
1827
                  i_mb_count[I_8x8]  / i_count,
 
1828
                  i_mb_count[I_4x4]  / i_count );
 
1829
    }
 
1830
    if( h->stat.i_slice_count[SLICE_TYPE_P] > 0 )
 
1831
    {
 
1832
        const int64_t *i_mb_count = h->stat.i_mb_count[SLICE_TYPE_P];
 
1833
        const int64_t *i_mb_size = h->stat.i_mb_count_size[SLICE_TYPE_P];
 
1834
        const double i_count = h->stat.i_slice_count[SLICE_TYPE_P] * h->mb.i_mb_count / 100.0;
 
1835
        x264_log( h, X264_LOG_INFO,
 
1836
                  "mb P  I16..4: %4.1f%% %4.1f%% %4.1f%%  P16..4: %4.1f%% %4.1f%% %4.1f%% %4.1f%% %4.1f%%    skip:%4.1f%%\n",
 
1837
                  i_mb_count[I_16x16]/ i_count,
 
1838
                  i_mb_count[I_8x8]  / i_count,
 
1839
                  i_mb_count[I_4x4]  / i_count,
 
1840
                  i_mb_size[PIXEL_16x16] / (i_count*4),
 
1841
                  (i_mb_size[PIXEL_16x8] + i_mb_size[PIXEL_8x16]) / (i_count*4),
 
1842
                  i_mb_size[PIXEL_8x8] / (i_count*4),
 
1843
                  (i_mb_size[PIXEL_8x4] + i_mb_size[PIXEL_4x8]) / (i_count*4),
 
1844
                  i_mb_size[PIXEL_4x4] / (i_count*4),
 
1845
                  i_mb_count[P_SKIP] / i_count );
 
1846
    }
 
1847
    if( h->stat.i_slice_count[SLICE_TYPE_B] > 0 )
 
1848
    {
 
1849
        const int64_t *i_mb_count = h->stat.i_mb_count[SLICE_TYPE_B];
 
1850
        const int64_t *i_mb_size = h->stat.i_mb_count_size[SLICE_TYPE_B];
 
1851
        const double i_count = h->stat.i_slice_count[SLICE_TYPE_B] * h->mb.i_mb_count / 100.0;
 
1852
        x264_log( h, X264_LOG_INFO,
 
1853
                  "mb B  I16..4: %4.1f%% %4.1f%% %4.1f%%  B16..8: %4.1f%% %4.1f%% %4.1f%%  direct:%4.1f%%  skip:%4.1f%%\n",
 
1854
                  i_mb_count[I_16x16]  / i_count,
 
1855
                  i_mb_count[I_8x8]    / i_count,
 
1856
                  i_mb_count[I_4x4]    / i_count,
 
1857
                  i_mb_size[PIXEL_16x16] / (i_count*4),
 
1858
                  (i_mb_size[PIXEL_16x8] + i_mb_size[PIXEL_8x16]) / (i_count*4),
 
1859
                  i_mb_size[PIXEL_8x8] / (i_count*4),
 
1860
                  i_mb_count[B_DIRECT] / i_count,
 
1861
                  i_mb_count[B_SKIP]   / i_count );
 
1862
    }
 
1863
 
 
1864
    x264_ratecontrol_summary( h );
 
1865
 
 
1866
    if( h->stat.i_slice_count[SLICE_TYPE_I] + h->stat.i_slice_count[SLICE_TYPE_P] + h->stat.i_slice_count[SLICE_TYPE_B] > 0 )
 
1867
    {
 
1868
        const int i_count = h->stat.i_slice_count[SLICE_TYPE_I] +
 
1869
                            h->stat.i_slice_count[SLICE_TYPE_P] +
 
1870
                            h->stat.i_slice_count[SLICE_TYPE_B];
 
1871
        float fps = (float) h->param.i_fps_num / h->param.i_fps_den;
 
1872
#define SUM3(p) (p[SLICE_TYPE_I] + p[SLICE_TYPE_P] + p[SLICE_TYPE_B])
 
1873
#define SUM3b(p,o) (p[SLICE_TYPE_I][o] + p[SLICE_TYPE_P][o] + p[SLICE_TYPE_B][o])
 
1874
        float f_bitrate = fps * SUM3(h->stat.i_slice_size) / i_count / 125;
 
1875
 
 
1876
        if( h->pps->b_transform_8x8_mode )
 
1877
        {
 
1878
            int64_t i_i8x8 = SUM3b( h->stat.i_mb_count, I_8x8 );
 
1879
            int64_t i_intra = i_i8x8 + SUM3b( h->stat.i_mb_count, I_4x4 )
 
1880
                                     + SUM3b( h->stat.i_mb_count, I_16x16 );
 
1881
            x264_log( h, X264_LOG_INFO, "8x8 transform  intra:%.1f%%  inter:%.1f%%\n",
 
1882
                      100. * i_i8x8 / i_intra,
 
1883
                      100. * h->stat.i_mb_count_8x8dct[1] / h->stat.i_mb_count_8x8dct[0] );
 
1884
        }
 
1885
 
 
1886
        if( h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO
 
1887
            && h->stat.i_slice_count[SLICE_TYPE_B] )
 
1888
        {
 
1889
            x264_log( h, X264_LOG_INFO, "direct mvs  spatial:%.1f%%  temporal:%.1f%%\n",
 
1890
                      h->stat.i_direct_frames[1] * 100. / h->stat.i_slice_count[SLICE_TYPE_B],
 
1891
                      h->stat.i_direct_frames[0] * 100. / h->stat.i_slice_count[SLICE_TYPE_B] );
 
1892
        }
 
1893
 
 
1894
        if( h->frames.i_max_ref0 > 1 )
 
1895
        {
 
1896
            int i_slice;
 
1897
            for( i_slice = 0; i_slice < 2; i_slice++ )
 
1898
            {
 
1899
                char buf[200];
 
1900
                char *p = buf;
 
1901
                int64_t i_den = 0;
 
1902
                int i_max = 0;
 
1903
                for( i = 0; i < h->frames.i_max_ref0 << h->param.b_interlaced; i++ )
 
1904
                    if( h->stat.i_mb_count_ref[i_slice][i] )
 
1905
                    {
 
1906
                        i_den += h->stat.i_mb_count_ref[i_slice][i];
 
1907
                        i_max = i;
 
1908
                    }
 
1909
                if( i_max == 0 )
 
1910
                    continue;
 
1911
                for( i = 0; i <= i_max; i++ )
 
1912
                    p += sprintf( p, " %4.1f%%", 100. * h->stat.i_mb_count_ref[i_slice][i] / i_den );
 
1913
                x264_log( h, X264_LOG_INFO, "ref %c %s\n", i_slice==SLICE_TYPE_P ? 'P' : 'B', buf );
 
1914
            }
 
1915
        }
 
1916
 
 
1917
        if( h->param.analyse.b_ssim )
 
1918
        {
 
1919
            x264_log( h, X264_LOG_INFO,
 
1920
                      "SSIM Mean Y:%.7f\n",
 
1921
                      SUM3( h->stat.f_ssim_mean_y ) / i_count );
 
1922
        }
 
1923
        if( h->param.analyse.b_psnr )
 
1924
        {
 
1925
            x264_log( h, X264_LOG_INFO,
 
1926
                      "PSNR Mean Y:%6.3f U:%6.3f V:%6.3f Avg:%6.3f Global:%6.3f kb/s:%.2f\n",
 
1927
                      SUM3( h->stat.f_psnr_mean_y ) / i_count,
 
1928
                      SUM3( h->stat.f_psnr_mean_u ) / i_count,
 
1929
                      SUM3( h->stat.f_psnr_mean_v ) / i_count,
 
1930
                      SUM3( h->stat.f_psnr_average ) / i_count,
 
1931
                      x264_psnr( SUM3( h->stat.i_sqe_global ), i_count * i_yuv_size ),
 
1932
                      f_bitrate );
 
1933
        }
 
1934
        else
 
1935
            x264_log( h, X264_LOG_INFO, "kb/s:%.1f\n", f_bitrate );
 
1936
    }
 
1937
 
 
1938
    /* frames */
 
1939
    for( i = 0; h->frames.current[i]; i++ )
 
1940
        x264_frame_delete( h->frames.current[i] );
 
1941
    for( i = 0; h->frames.next[i]; i++ )
 
1942
        x264_frame_delete( h->frames.next[i] );
 
1943
    for( i = 0; h->frames.unused[i]; i++ )
 
1944
        x264_frame_delete( h->frames.unused[i] );
 
1945
    for( i = 0; h->frames.reference[i]; i++ )
 
1946
        x264_frame_delete( h->frames.reference[i] );
 
1947
 
 
1948
    /* rc */
 
1949
    x264_ratecontrol_delete( h );
 
1950
 
 
1951
    /* param */
 
1952
    if( h->param.rc.psz_stat_out )
 
1953
        free( h->param.rc.psz_stat_out );
 
1954
    if( h->param.rc.psz_stat_in )
 
1955
        free( h->param.rc.psz_stat_in );
 
1956
    if( h->param.rc.psz_rc_eq )
 
1957
        free( h->param.rc.psz_rc_eq );
 
1958
 
 
1959
    x264_cqm_delete( h );
 
1960
    for( i = h->param.i_threads - 1; i >= 0; i-- )
 
1961
    {
 
1962
        x264_macroblock_cache_end( h->thread[i] );
 
1963
        x264_free( h->thread[i]->out.p_bitstream );
 
1964
        x264_free( h->thread[i] );
 
1965
    }
 
1966
}