]> git.sesse.net Git - x264/blob - encoder/lookahead.c
Fix possible non-determinism with mbtree + open-gop + sync-lookahead
[x264] / encoder / lookahead.c
1 /*****************************************************************************
2  * lookahead.c: high-level lookahead functions
3  *****************************************************************************
4  * Copyright (C) 2010-2013 Avail Media and x264 project
5  *
6  * Authors: Michael Kazmier <mkazmier@availmedia.com>
7  *          Alex Giladi <agiladi@availmedia.com>
8  *          Steven Walters <kemuri9@gmail.com>
9  *
10  * This program is free software; you can redistribute it and/or modify
11  * it under the terms of the GNU General Public License as published by
12  * the Free Software Foundation; either version 2 of the License, or
13  * (at your option) any later version.
14  *
15  * This program is distributed in the hope that it will be useful,
16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
18  * GNU General Public License for more details.
19  *
20  * You should have received a copy of the GNU General Public License
21  * along with this program; if not, write to the Free Software
22  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02111, USA.
23  *
24  * This program is also available under a commercial proprietary license.
25  * For more information, contact us at licensing@x264.com.
26  *****************************************************************************/
27
28 /* LOOKAHEAD (threaded and non-threaded mode)
29  *
30  * Lookahead types:
31  *     [1] Slice type / scene cut;
32  *
33  * In non-threaded mode, we run the existing slicetype decision code as it was.
34  * In threaded mode, we run in a separate thread, that lives between the calls
35  * to x264_encoder_open() and x264_encoder_close(), and performs lookahead for
36  * the number of frames specified in rc_lookahead.  Recommended setting is
37  * # of bframes + # of threads.
38  */
39 #include "common/common.h"
40 #include "analyse.h"
41
42 static void x264_lookahead_shift( x264_sync_frame_list_t *dst, x264_sync_frame_list_t *src, int count )
43 {
44     int i = count;
45     while( i-- )
46     {
47         assert( dst->i_size < dst->i_max_size );
48         assert( src->i_size );
49         dst->list[ dst->i_size++ ] = x264_frame_shift( src->list );
50         src->i_size--;
51     }
52     if( count )
53     {
54         x264_pthread_cond_broadcast( &dst->cv_fill );
55         x264_pthread_cond_broadcast( &src->cv_empty );
56     }
57 }
58
59 static void x264_lookahead_update_last_nonb( x264_t *h, x264_frame_t *new_nonb )
60 {
61     if( h->lookahead->last_nonb )
62         x264_frame_push_unused( h, h->lookahead->last_nonb );
63     h->lookahead->last_nonb = new_nonb;
64     new_nonb->i_reference_count++;
65 }
66
67 #if HAVE_THREAD
68 static void x264_lookahead_slicetype_decide( x264_t *h )
69 {
70     x264_stack_align( x264_slicetype_decide, h );
71
72     x264_lookahead_update_last_nonb( h, h->lookahead->next.list[0] );
73     int shift_frames = h->lookahead->next.list[0]->i_bframes + 1;
74
75     x264_pthread_mutex_lock( &h->lookahead->ofbuf.mutex );
76     while( h->lookahead->ofbuf.i_size == h->lookahead->ofbuf.i_max_size )
77         x264_pthread_cond_wait( &h->lookahead->ofbuf.cv_empty, &h->lookahead->ofbuf.mutex );
78
79     x264_pthread_mutex_lock( &h->lookahead->next.mutex );
80     x264_lookahead_shift( &h->lookahead->ofbuf, &h->lookahead->next, shift_frames );
81     x264_pthread_mutex_unlock( &h->lookahead->next.mutex );
82
83     /* For MB-tree and VBV lookahead, we have to perform propagation analysis on I-frames too. */
84     if( h->lookahead->b_analyse_keyframe && IS_X264_TYPE_I( h->lookahead->last_nonb->i_type ) )
85         x264_stack_align( x264_slicetype_analyse, h, shift_frames );
86
87     x264_pthread_mutex_unlock( &h->lookahead->ofbuf.mutex );
88 }
89
90 static void *x264_lookahead_thread( x264_t *h )
91 {
92     int shift;
93 #if HAVE_MMX
94     if( h->param.cpu&X264_CPU_SSE_MISALIGN )
95         x264_cpu_mask_misalign_sse();
96 #endif
97     while( !h->lookahead->b_exit_thread )
98     {
99         x264_pthread_mutex_lock( &h->lookahead->ifbuf.mutex );
100         x264_pthread_mutex_lock( &h->lookahead->next.mutex );
101         shift = X264_MIN( h->lookahead->next.i_max_size - h->lookahead->next.i_size, h->lookahead->ifbuf.i_size );
102         x264_lookahead_shift( &h->lookahead->next, &h->lookahead->ifbuf, shift );
103         x264_pthread_mutex_unlock( &h->lookahead->next.mutex );
104         if( h->lookahead->next.i_size <= h->lookahead->i_slicetype_length + h->param.b_vfr_input )
105         {
106             while( !h->lookahead->ifbuf.i_size && !h->lookahead->b_exit_thread )
107                 x264_pthread_cond_wait( &h->lookahead->ifbuf.cv_fill, &h->lookahead->ifbuf.mutex );
108             x264_pthread_mutex_unlock( &h->lookahead->ifbuf.mutex );
109         }
110         else
111         {
112             x264_pthread_mutex_unlock( &h->lookahead->ifbuf.mutex );
113             x264_lookahead_slicetype_decide( h );
114         }
115     }   /* end of input frames */
116     x264_pthread_mutex_lock( &h->lookahead->ifbuf.mutex );
117     x264_pthread_mutex_lock( &h->lookahead->next.mutex );
118     x264_lookahead_shift( &h->lookahead->next, &h->lookahead->ifbuf, h->lookahead->ifbuf.i_size );
119     x264_pthread_mutex_unlock( &h->lookahead->next.mutex );
120     x264_pthread_mutex_unlock( &h->lookahead->ifbuf.mutex );
121     while( h->lookahead->next.i_size )
122         x264_lookahead_slicetype_decide( h );
123     x264_pthread_mutex_lock( &h->lookahead->ofbuf.mutex );
124     h->lookahead->b_thread_active = 0;
125     x264_pthread_cond_broadcast( &h->lookahead->ofbuf.cv_fill );
126     x264_pthread_mutex_unlock( &h->lookahead->ofbuf.mutex );
127     return NULL;
128 }
129 #endif
130
131 int x264_lookahead_init( x264_t *h, int i_slicetype_length )
132 {
133     x264_lookahead_t *look;
134     CHECKED_MALLOCZERO( look, sizeof(x264_lookahead_t) );
135     for( int i = 0; i < h->param.i_threads; i++ )
136         h->thread[i]->lookahead = look;
137
138     look->i_last_keyframe = - h->param.i_keyint_max;
139     look->b_analyse_keyframe = (h->param.rc.b_mb_tree || (h->param.rc.i_vbv_buffer_size && h->param.rc.i_lookahead))
140                                && !h->param.rc.b_stat_read;
141     look->i_slicetype_length = i_slicetype_length;
142
143     /* init frame lists */
144     if( x264_sync_frame_list_init( &look->ifbuf, h->param.i_sync_lookahead+3 ) ||
145         x264_sync_frame_list_init( &look->next, h->frames.i_delay+3 ) ||
146         x264_sync_frame_list_init( &look->ofbuf, h->frames.i_delay+3 ) )
147         goto fail;
148
149     if( !h->param.i_sync_lookahead )
150         return 0;
151
152     x264_t *look_h = h->thread[h->param.i_threads];
153     *look_h = *h;
154     if( x264_macroblock_cache_allocate( look_h ) )
155         goto fail;
156
157     if( x264_macroblock_thread_allocate( look_h, 1 ) < 0 )
158         goto fail;
159
160     if( x264_pthread_create( &look->thread_handle, NULL, (void*)x264_lookahead_thread, look_h ) )
161         goto fail;
162     look->b_thread_active = 1;
163
164     return 0;
165 fail:
166     x264_free( look );
167     return -1;
168 }
169
170 void x264_lookahead_delete( x264_t *h )
171 {
172     if( h->param.i_sync_lookahead )
173     {
174         x264_pthread_mutex_lock( &h->lookahead->ifbuf.mutex );
175         h->lookahead->b_exit_thread = 1;
176         x264_pthread_cond_broadcast( &h->lookahead->ifbuf.cv_fill );
177         x264_pthread_mutex_unlock( &h->lookahead->ifbuf.mutex );
178         x264_pthread_join( h->lookahead->thread_handle, NULL );
179         x264_macroblock_cache_free( h->thread[h->param.i_threads] );
180         x264_macroblock_thread_free( h->thread[h->param.i_threads], 1 );
181         x264_free( h->thread[h->param.i_threads] );
182     }
183     x264_sync_frame_list_delete( &h->lookahead->ifbuf );
184     x264_sync_frame_list_delete( &h->lookahead->next );
185     if( h->lookahead->last_nonb )
186         x264_frame_push_unused( h, h->lookahead->last_nonb );
187     x264_sync_frame_list_delete( &h->lookahead->ofbuf );
188     x264_free( h->lookahead );
189 }
190
191 void x264_lookahead_put_frame( x264_t *h, x264_frame_t *frame )
192 {
193     if( h->param.i_sync_lookahead )
194         x264_sync_frame_list_push( &h->lookahead->ifbuf, frame );
195     else
196         x264_sync_frame_list_push( &h->lookahead->next, frame );
197 }
198
199 int x264_lookahead_is_empty( x264_t *h )
200 {
201     x264_pthread_mutex_lock( &h->lookahead->ofbuf.mutex );
202     x264_pthread_mutex_lock( &h->lookahead->next.mutex );
203     int b_empty = !h->lookahead->next.i_size && !h->lookahead->ofbuf.i_size;
204     x264_pthread_mutex_unlock( &h->lookahead->next.mutex );
205     x264_pthread_mutex_unlock( &h->lookahead->ofbuf.mutex );
206     return b_empty;
207 }
208
209 static void x264_lookahead_encoder_shift( x264_t *h )
210 {
211     if( !h->lookahead->ofbuf.i_size )
212         return;
213     int i_frames = h->lookahead->ofbuf.list[0]->i_bframes + 1;
214     while( i_frames-- )
215     {
216         x264_frame_push( h->frames.current, x264_frame_shift( h->lookahead->ofbuf.list ) );
217         h->lookahead->ofbuf.i_size--;
218     }
219     x264_pthread_cond_broadcast( &h->lookahead->ofbuf.cv_empty );
220 }
221
222 void x264_lookahead_get_frames( x264_t *h )
223 {
224     if( h->param.i_sync_lookahead )
225     {   /* We have a lookahead thread, so get frames from there */
226         x264_pthread_mutex_lock( &h->lookahead->ofbuf.mutex );
227         while( !h->lookahead->ofbuf.i_size && h->lookahead->b_thread_active )
228             x264_pthread_cond_wait( &h->lookahead->ofbuf.cv_fill, &h->lookahead->ofbuf.mutex );
229         x264_lookahead_encoder_shift( h );
230         x264_pthread_mutex_unlock( &h->lookahead->ofbuf.mutex );
231     }
232     else
233     {   /* We are not running a lookahead thread, so perform all the slicetype decide on the fly */
234
235         if( h->frames.current[0] || !h->lookahead->next.i_size )
236             return;
237
238         x264_stack_align( x264_slicetype_decide, h );
239         x264_lookahead_update_last_nonb( h, h->lookahead->next.list[0] );
240         int shift_frames = h->lookahead->next.list[0]->i_bframes + 1;
241         x264_lookahead_shift( &h->lookahead->ofbuf, &h->lookahead->next, shift_frames );
242
243         /* For MB-tree and VBV lookahead, we have to perform propagation analysis on I-frames too. */
244         if( h->lookahead->b_analyse_keyframe && IS_X264_TYPE_I( h->lookahead->last_nonb->i_type ) )
245             x264_stack_align( x264_slicetype_analyse, h, shift_frames );
246
247         x264_lookahead_encoder_shift( h );
248     }
249 }