Mercurial > cgi-bin > hgwebdir.cgi > PR > Applications > VSs > VSs__H264__App

annotate libavcodec/h264.c @ 4:96e628866d41

find changesets by author, revision, files, or words in the commit message
naming some tasks to help debugging
author Nina Engelhardt <nengel@mailbox.tu-berlin.de>
date Wed, 19 Dec 2012 15:40:26 +0100
parents
children
rev   line source
nengel@2 1 #include "config.h"
nengel@2 2 #include "h264.h"
nengel@2 3 #include "h264_misc.h"
nengel@2 4 #include <math.h>
nengel@2 5
nengel@2 6 H264Context *get_h264dec_context(const char *file_name, int ifile, int ofile, int width, int height, h264_options *opts){
nengel@2 7 int i;
nengel@2 8 const int mb_height = (height + 15) / 16;
nengel@2 9 const int mb_width = (width + 15) / 16;
nengel@2 10 const int mb_stride = ((mb_width+1)/16 + 1) *16; //align mb_stride to 16
nengel@2 11
nengel@2 12 ff_init_cabac_states();
nengel@2 13
nengel@2 14 H264Context *h= av_mallocz(sizeof(H264Context));
nengel@2 15
nengel@2 16 start_timer(h, TOTAL);
nengel@2 17 h->file_name = file_name;
nengel@2 18 h->profile = opts->profile;
nengel@2 19 for (i=0; i<PROFILE_STAGES; i++)
nengel@2 20 h->total_time[i]=0;
nengel@2 21
nengel@2 22 h->ifile=ifile;
nengel@2 23 h->ofile =ofile;
nengel@2 24
nengel@2 25 h->verbose =opts->verbose;
nengel@2 26 h->no_mbd =opts->no_mbd;
nengel@2 27 h->static_3d =opts->static_3d;
nengel@2 28 h->pipe_bufs = opts->pipe_bufs;
nengel@2 29 h->slice_bufs = opts->slice_bufs;
nengel@2 30
nengel@2 31 h->ed_ppe_threads =0;
nengel@2 32 if (opts->ppe_ed){
nengel@2 33 h->ed_ppe_threads = (opts->threads >opts->ppe_ed)? opts->ppe_ed :opts->threads;
nengel@2 34 }
nengel@2 35
nengel@2 36 h->threads = opts->threads - h->ed_ppe_threads;
nengel@2 37 h->smt = opts->smt;
nengel@2 38 if (h->smt){
nengel@2 39 h->threads *= 2;
nengel@2 40 }
nengel@2 41
nengel@2 42 h->num_frames = opts->numframes;
nengel@2 43
nengel@2 44 h->frame_width = width;
nengel@2 45 h->frame_height = height;
nengel@2 46
nengel@2 47 while ((width/2) %STRIDE_ALIGN)
nengel@2 48 width+=STRIDE_ALIGN;
nengel@2 49 h->width = width;
nengel@2 50 h->height = mb_height*16;
nengel@2 51
nengel@2 52 h->mb_height = mb_height;
nengel@2 53 h->mb_width = mb_width;
nengel@2 54 h->mb_stride = mb_stride;
nengel@2 55 h->b4_stride = mb_width*4 + 1;
nengel@2 56 h->b_stride = mb_width*4;
nengel@2 57
nengel@2 58 h->smb_width = opts->smb_size[0];
nengel@2 59 h->smb_height = opts->smb_size[1] < h->smb_width ? opts->smb_size[1] : h->smb_width;
nengel@2 60 h->smbc = getSuperMBContext(h, h->smb_width, h->smb_height);
nengel@2 61
nengel@2 62 h->wave_order = opts->wave_order;
nengel@2 63
nengel@2 64 h->pipe_bufs = opts->pipe_bufs;
nengel@2 65
nengel@2 66 h->max_dpb_cnt = DPB_SIZE + opts->pipe_bufs;
nengel@2 67 h->free_dpb_cnt = h->max_dpb_cnt;
nengel@2 68 h->dpb = av_mallocz (h->max_dpb_cnt* sizeof (DecodedPicture));
nengel@2 69
nengel@2 70
nengel@2 71 h->free_sb_cnt = h->threads*opts->slice_bufs + (h->no_mbd != 0) ; //one extra to overlap some latency of signaling/freeing slicebuffers in entropy only mode
nengel@2 72 h->sb_size = h->free_sb_cnt;
nengel@2 73 h->sb = av_mallocz(h->sb_size* sizeof(SliceBufferEntry));
nengel@2 74
nengel@2 75 h->rl_q.size = FFMAX(1, FFMIN( (h->height-3 - 512)/16, h->mb_width/2)) +1;
nengel@2 76 h->rl_q.free = h->rl_q.size -1;
nengel@2 77 h->rl_q.ready=0;
nengel@2 78 h->rl_q.fi = h->rl_q.fo= 0;
nengel@2 79 h->rl_q.queue = av_malloc(h->rl_q.size* sizeof(RingLineEntry*));
nengel@2 80 for (i=0; i<h->rl_q.size; i++){
nengel@2 81 if( posix_memalign((void**)&h->rl_q.queue[i],64,sizeof(RingLineEntry)))
nengel@2 82 h->rl_q.queue[i]=NULL;
nengel@2 83 h->rl_q.queue[i]->top = av_malloc(h->mb_width*sizeof(TopBorder));
nengel@2 84 }
nengel@2 85
nengel@2 86 h->rl_q.queue[0]->prev_line = h->rl_q.queue[h->rl_q.size-1];
nengel@2 87 for (i=1; i<h->rl_q.size; i++){
nengel@2 88 h->rl_q.queue[i]->prev_line = h->rl_q.queue[i-1];
nengel@2 89 }
nengel@2 90
nengel@2 91 if( HAVE_MMX | HAVE_ALTIVEC| HAVE_NEON ){
nengel@2 92 for(i=0; i<16; i++){
nengel@2 93 #define T(x) (x>>2) | ((x<<2) & 0xF)
nengel@2 94 h->zigzag_scan[i] = T(zigzag_scan[i]);
nengel@2 95 #undef T
nengel@2 96 }
nengel@2 97 for(i=0; i<64; i++){
nengel@2 98 #define T(x) (x>>3) | ((x&7)<<3)
nengel@2 99 h->zigzag_scan8x8[i] = T(ff_zigzag_direct[i]);
nengel@2 100 #undef T
nengel@2 101 }
nengel@2 102 }else{
nengel@2 103 memcpy(h->zigzag_scan, zigzag_scan, 16*sizeof(uint8_t));
nengel@2 104 memcpy(h->zigzag_scan8x8, ff_zigzag_direct, 64*sizeof(uint8_t));
nengel@2 105 }
nengel@2 106
nengel@2 107 pthread_mutex_init(&h->smb_lock, NULL);
nengel@2 108 pthread_mutex_init(&h->sdl_lock, NULL);
nengel@2 109 pthread_cond_init(&h->sdl_cond, NULL);
nengel@2 110
nengel@2 111 ///pthread initialization
nengel@2 112 pthread_mutex_init(&h->ilock, NULL);
nengel@2 113 pthread_cond_init(&h->icond, NULL);
nengel@2 114 pthread_mutex_init(&h->slock, NULL);
nengel@2 115 pthread_cond_init(&h->scond, NULL);
nengel@2 116 pthread_mutex_init(&h->tlock, NULL);
nengel@2 117 pthread_cond_init(&h->tcond, NULL);
nengel@2 118 pthread_mutex_init(&h->tdlock, NULL);
nengel@2 119 pthread_cond_init(&h->tdcond, NULL);
nengel@2 120 h->start =!opts->numamap; //default dont wait for start signal
nengel@2 121 h->statmbd = opts->statmbd;
nengel@2 122 h->rl_side_touch= opts->numamap;
nengel@2 123 h->touch_start=0;
nengel@2 124 h->setaff =opts->statsched;
nengel@2 125 h->init_threads=0;
nengel@2 126
nengel@2 127 pthread_mutex_init(&h->task_lock, NULL);
nengel@2 128 pthread_cond_init(&h->task_cond, NULL);
nengel@2 129 for (i=0; i<STAGES; i++){
nengel@2 130 pthread_mutex_init (&h->lock[i], NULL);
nengel@2 131 pthread_cond_init (&h->cond[i], NULL);
nengel@2 132
nengel@2 133 pthread_mutex_init (&h->sb_q[i].lock, NULL);
nengel@2 134 pthread_cond_init (&h->sb_q[i].cond, NULL);
nengel@2 135 h->sb_q[i].size = h->free_sb_cnt; //change to num threads later
nengel@2 136 h->sb_q[i].queue = av_malloc(h->free_sb_cnt* sizeof(SliceBufferEntry*));
nengel@2 137 h->sb_q[i].cnt = h->sb_q[i].fi = h->sb_q[i].fo =0;
nengel@2 138 }
nengel@2 139
nengel@2 140 #if HAVE_LIBSDL2
nengel@2 141 h->sdlq.size=2;
nengel@2 142 h->sdlq.ready=2;
nengel@2 143 h->sdlq.queue = av_malloc(2* sizeof(SDL_Texture*));
nengel@2 144 pthread_mutex_init (&h->sdlq.sdl_lock, NULL);
nengel@2 145 pthread_cond_init (&h->sdlq.sdl_cond, NULL);
nengel@2 146 #endif
nengel@2 147
nengel@2 148 h->display=opts->display;
nengel@2 149 h->fullscreen=opts->fullscreen;
nengel@2 150
nengel@2 151 return h;
nengel@2 152 }
nengel@2 153
nengel@2 154
nengel@2 155 void free_h264dec_context(H264Context *h) {
nengel@2 156 int i;
nengel@2 157
nengel@2 158 for(i=0; i<h->max_dpb_cnt; i++)
nengel@2 159 free_dp(&h->dpb[i]);
nengel@2 160 av_free (h->dpb);
nengel@2 161
nengel@2 162 for(i=0; i<h->sb_size; i++){
nengel@2 163 if (h->sb[i].initialized){
nengel@2 164 free_sb_entry(&h->sb[i]);
nengel@2 165 }
nengel@2 166 }
nengel@2 167 av_freep(&h->sb);
nengel@2 168
nengel@2 169 for (i=0; i<h->rl_q.size; i++){
nengel@2 170 av_freep(&h->rl_q.queue[i]->top);
nengel@2 171 av_freep(&h->rl_q.queue[i]);
nengel@2 172 }
nengel@2 173 av_freep(&h->rl_q.queue);
nengel@2 174
nengel@2 175 ///pthread cleanup
nengel@2 176 pthread_mutex_destroy (&h->task_lock);
nengel@2 177 pthread_cond_destroy (&h->task_cond);
nengel@2 178 for (i=0; i<STAGES; i++){
nengel@2 179 pthread_mutex_destroy (&h->lock[i]);
nengel@2 180 pthread_cond_destroy (&h->cond[i]);
nengel@2 181
nengel@2 182 pthread_mutex_destroy (&h->sb_q[i].lock);
nengel@2 183 pthread_cond_destroy (&h->sb_q[i].cond);
nengel@2 184 av_freep( &h->sb_q[i].queue);
nengel@2 185 }
nengel@2 186 pthread_mutex_destroy (&h->slock);
nengel@2 187 pthread_cond_destroy (&h->scond);
nengel@2 188 pthread_mutex_destroy (&h->ilock);
nengel@2 189 pthread_cond_destroy (&h->icond);
nengel@2 190
nengel@2 191 pthread_mutex_destroy(&h->smb_lock);
nengel@2 192 pthread_mutex_destroy (&h->sdl_lock);
nengel@2 193 pthread_cond_destroy (&h->sdl_cond);
nengel@2 194 #if HAVE_LIBSDL2
nengel@2 195 av_free(h->sdlq.queue);
nengel@2 196 pthread_mutex_destroy (&h->sdlq.sdl_lock);
nengel@2 197 pthread_cond_destroy (&h->sdlq.sdl_cond);
nengel@2 198 #endif
nengel@2 199
nengel@2 200 stop_timer(h, TOTAL);
nengel@2 201 if (h->threads==0){
nengel@2 202 for (i=0; i<PROFILE_STAGES; i++)
nengel@2 203 h->total_time[i] /= h->num_frames;
nengel@2 204 double others = h->total_time[TOTAL];
nengel@2 205 for (i=1; i<PROFILE_STAGES; i++)
nengel@2 206 others-=h->total_time[i];
nengel@2 207 if (h->profile == 1){
nengel@2 208 printf("\n[FRAME %.3fms] [FRONT %.3fms] [ENTROPY %.3fms] [MBREC %.3fms] [OTHERS %.3fms]\n", h->total_time[TOTAL], h->total_time[FRONT], h->total_time[ED], h->total_time[REC], others);
nengel@2 209 }else if (h->profile ==2){
nengel@2 210 printf("\n[FRAME %.3fms] [FRONT %.3fms] [ENTROPY %.3fms] [PRED %.3fms] [OTHERS %.3fms]\n", h->total_time[TOTAL], h->total_time[FRONT], h->total_time[ED],h->total_time[REC], others);
nengel@2 211 }
nengel@2 212 }
nengel@2 213
nengel@2 214 av_free(h);
nengel@2 215 }