1 /* $Id: encx264.c,v 1.21 2005/11/04 13:09:41 titer Exp $
3 This file is part of the HandBrake source code.
4 Homepage: <http://handbrake.fr/>.
5 It may be used under the terms of the GNU General Public License. */
13 int encx264Init( hb_work_object_t *, hb_job_t * );
14 int encx264Work( hb_work_object_t *, hb_buffer_t **, hb_buffer_t ** );
15 void encx264Close( hb_work_object_t * );
17 hb_work_object_t hb_encx264 =
20 "H.264/AVC encoder (libx264)",
26 #define DTS_BUFFER_SIZE 32
29 * The frame info struct remembers information about each frame across calls
30 * to x264_encoder_encode. Since frames are uniquely identified by their
31 * timestamp, we use some bits of the timestamp as an index. The LSB is
32 * chosen so that two successive frames will have different values in the
33 * bits over any plausible range of frame rates. (Starting with bit 9 allows
34 * any frame rate slower than 175fps.) The MSB determines the size of the array.
35 * It is chosen so that two frames can't use the same slot during the
36 * encoder's max frame delay (set by the standard as 16 frames) and so
37 * that, up to some minimum frame rate, frames are guaranteed to map to
38 * different slots. (An MSB of 16 which is 2^(16-9+1) = 256 slots guarantees
39 * no collisions down to a rate of 1.4 fps).
41 #define FRAME_INFO_MAX2 (9) // 2^9 = 512; 90000/512 = 175 frames/sec
42 #define FRAME_INFO_MIN2 (16) // 2^16 = 65536; 90000/65536 = 1.4 frames/sec
43 #define FRAME_INFO_SIZE (1 << (FRAME_INFO_MIN2 - FRAME_INFO_MAX2 + 1))
44 #define FRAME_INFO_MASK (FRAME_INFO_SIZE - 1)
46 struct hb_work_private_s
50 x264_picture_t pic_in;
51 uint8_t *x264_allocated_pic;
53 int chap_mark; // saved chap mark when we're propagating it
54 int64_t dts_next; // DTS start time value for next output frame
55 int64_t last_stop; // Debugging - stop time of previous input frame
57 int64_t max_delay; // if init_delay too small, delay really needed
62 } frame_info[FRAME_INFO_SIZE];
67 /***********************************************************************
68 * hb_work_encx264_init
69 ***********************************************************************
71 **********************************************************************/
72 int encx264Init( hb_work_object_t * w, hb_job_t * job )
79 hb_work_private_t * pv = calloc( 1, sizeof( hb_work_private_t ) );
84 memset( pv->filename, 0, 1024 );
85 hb_get_tempory_filename( job->h, pv->filename, "x264.log" );
87 x264_param_default( ¶m );
89 param.i_threads = ( hb_get_cpu_count() * 3 / 2 );
90 param.i_width = job->width;
91 param.i_height = job->height;
92 param.i_fps_num = job->vrate;
93 param.i_fps_den = job->vrate_base;
95 if (job->vrate_base != 1080000)
97 /* If the fps isn't 25, adjust the key intervals. Add 1 because
98 we want 24, not 23 with a truncated remainder. */
99 param.i_keyint_min = (job->vrate / job->vrate_base) + 1;
100 param.i_keyint_max = (10 * job->vrate / job->vrate_base) + 1;
101 hb_log("encx264: keyint-min: %i, keyint-max: %i", param.i_keyint_min, param.i_keyint_max);
104 param.i_log_level = X264_LOG_INFO;
105 if( job->h264_level )
108 param.i_level_idc = job->h264_level;
109 hb_log( "encx264: encoding at level %i",
113 /* Slightly faster with minimal quality lost */
114 param.analyse.i_subpel_refine = 4;
117 This section passes the string x264opts to libx264 for parsing into
118 parameter names and values.
120 The string is set up like this:
121 option1=value1:option2=value 2
123 So, you have to iterate through based on the colons, and then put
124 the left side of the equals sign in "name" and the right side into
125 "value." Then you hand those strings off to x264 for interpretation.
127 This is all based on the universal x264 option handling Loren
128 Merritt implemented in the Mplayer/Mencoder project.
131 if( job->x264opts != NULL && *job->x264opts != '\0' )
133 char *x264opts, *x264opts_start;
135 x264opts = x264opts_start = strdup(job->x264opts);
137 while( x264opts_start && *x264opts )
139 char *name = x264opts;
143 x264opts += strcspn( x264opts, ":" );
150 value = strchr( name, '=' );
158 When B-frames are enabled, the max frame count increments
159 by 1 (regardless of the number of B-frames). If you don't
160 change the duration of the video track when you mux, libmp4
161 barfs. So, check if the x264opts are using B-frames, and
162 when they are, set the boolean job->areBframes as true.
165 if( !( strcmp( name, "bframes" ) ) )
167 if( atoi( value ) > 0 )
173 /* Note b-pyramid here, so the initial delay can be doubled */
174 if( !( strcmp( name, "b-pyramid" ) ) )
178 if( atoi( value ) > 0 )
189 /* Here's where the strings are passed to libx264 for parsing. */
190 ret = x264_param_parse( ¶m, name, value );
192 /* Let x264 sanity check the options for us*/
193 if( ret == X264_PARAM_BAD_NAME )
194 hb_log( "x264 options: Unknown suboption %s", name );
195 if( ret == X264_PARAM_BAD_VALUE )
196 hb_log( "x264 options: Bad argument %s=%s", name, value ? value : "(null)" );
198 free(x264opts_start);
201 /* set up the VUI color model & gamma to match what the COLR atom
202 * set in muxmp4.c says. See libhb/muxmp4.c for notes. */
204 if ( job->title->height >= 720 )
206 // we guess that 720p or above is ITU BT.709 HD content
207 param.vui.i_colorprim = 1;
208 param.vui.i_transfer = 1;
209 param.vui.i_colmatrix = 1;
213 // ITU BT.601 DVD or SD TV content
214 param.vui.i_colorprim = 6;
215 param.vui.i_transfer = 1;
216 param.vui.i_colmatrix = 6;
219 if( job->pixel_ratio )
221 param.vui.i_sar_width = job->pixel_aspect_width;
222 param.vui.i_sar_height = job->pixel_aspect_height;
224 hb_log( "encx264: encoding with stored aspect %d/%d",
225 param.vui.i_sar_width, param.vui.i_sar_height );
229 if( job->vquality > 0.0 && job->vquality < 1.0 )
235 param.rc.i_rc_method = X264_RC_CRF;
236 param.rc.f_rf_constant = 51 - job->vquality * 51;
237 hb_log( "encx264: Encoding at constant RF %f",
238 param.rc.f_rf_constant );
243 param.rc.i_rc_method = X264_RC_CQP;
244 param.rc.i_qp_constant = 51 - job->vquality * 51;
245 hb_log( "encx264: encoding at constant QP %d",
246 param.rc.i_qp_constant );
250 else if( job->vquality == 0 || job->vquality >= 1.0 )
252 /* Use the vquality as a raw RF or QP
253 instead of treating it like a percentage. */
258 param.rc.i_rc_method = X264_RC_CRF;
259 param.rc.f_rf_constant = job->vquality;
260 hb_log( "encx264: Encoding at constant RF %f",
261 param.rc.f_rf_constant );
266 param.rc.i_rc_method = X264_RC_CQP;
267 param.rc.i_qp_constant = job->vquality;
268 hb_log( "encx264: encoding at constant QP %d",
269 param.rc.i_qp_constant );
276 param.rc.i_rc_method = X264_RC_ABR;
277 param.rc.i_bitrate = job->vbitrate;
281 param.rc.b_stat_write = 1;
282 param.rc.psz_stat_out = pv->filename;
285 param.rc.b_stat_read = 1;
286 param.rc.psz_stat_in = pv->filename;
291 hb_log( "encx264: opening libx264 (pass %d)", job->pass );
292 pv->x264 = x264_encoder_open( ¶m );
294 x264_encoder_headers( pv->x264, &nal, &nal_count );
296 /* Sequence Parameter Set */
297 x264_nal_encode( w->config->h264.sps, &nal_size, 0, &nal[1] );
298 w->config->h264.sps_length = nal_size;
300 /* Picture Parameter Set */
301 x264_nal_encode( w->config->h264.pps, &nal_size, 0, &nal[2] );
302 w->config->h264.pps_length = nal_size;
304 x264_picture_alloc( &pv->pic_in, X264_CSP_I420,
305 job->width, job->height );
307 pv->x264_allocated_pic = pv->pic_in.img.plane[0];
314 /* Basic initDelay value is the clockrate divided by the FPS
315 -- the length of one frame in clockticks. */
316 pv->init_delay = 90000. / ((double)job->vrate / (double)job->vrate_base);
318 /* 23.976-length frames are 3753.75 ticks long on average but the DVD
319 creates that average rate by repeating 59.95 fields so the max
320 frame size is actually 4504.5 (3 field times) */
321 if (pv->init_delay == 3753)
322 pv->init_delay = 4505;
324 /* frame rates are not exact in the DVD 90KHz PTS clock (they are
325 exact in the DVD 27MHz system clock but we never see that) so the
326 rates computed above are all +-1 due to quantization. Worst case
327 is when a clock-rounded-down frame is adjacent to a rounded-up frame
328 which makes one of the frames 2 ticks longer than the nominal
332 /* For VFR, libhb sees the FPS as 29.97, but the longest frames
333 will use the duration of frames running at 23.976fps instead.. */
336 pv->init_delay = 7506;
339 /* The delay is 1 frames for regular b-frames, 2 for b-pyramid. */
340 pv->init_delay *= job->areBframes;
346 void encx264Close( hb_work_object_t * w )
348 hb_work_private_t * pv = w->private_data;
350 * Patch the x264 allocated data back in so that x264 can free it
351 * we have been using our own buffers during the encode to avoid copying.
353 pv->pic_in.img.plane[0] = pv->x264_allocated_pic;
354 x264_picture_clean( &pv->pic_in );
355 x264_encoder_close( pv->x264 );
357 w->private_data = NULL;
363 * see comments in definition of 'frame_info' in pv struct for description
364 * of what these routines are doing.
366 static void save_frame_info( hb_work_private_t * pv, hb_buffer_t * in )
368 int i = (in->start >> FRAME_INFO_MAX2) & FRAME_INFO_MASK;
369 pv->frame_info[i].duration = in->stop - in->start;
372 static int64_t get_frame_duration( hb_work_private_t * pv, int64_t pts )
374 int i = (pts >> FRAME_INFO_MAX2) & FRAME_INFO_MASK;
375 return pv->frame_info[i].duration;
378 int encx264Work( hb_work_object_t * w, hb_buffer_t ** buf_in,
379 hb_buffer_t ** buf_out )
381 hb_work_private_t * pv = w->private_data;
382 hb_job_t * job = pv->job;
383 hb_buffer_t * in = *buf_in, * buf;
384 x264_picture_t pic_out;
392 * Point x264 at our current buffers Y(UV) data.
394 pv->pic_in.img.plane[0] = in->data;
398 /* XXX x264 has currently no option for grayscale encoding */
399 memset( pv->pic_in.img.plane[1], 0x80, job->width * job->height / 4 );
400 memset( pv->pic_in.img.plane[2], 0x80, job->width * job->height / 4 );
405 * Point x264 at our buffers (Y)UV data
407 pv->pic_in.img.plane[1] = in->data + job->width * job->height;
408 pv->pic_in.img.plane[2] = in->data + 5 * job->width *
412 if( pv->dts_next == -1 )
414 /* we don't have a start time yet so use the first frame's
415 * start. All other frame times will be determined by the
416 * sum of the prior output frame durations in *DTS* order
417 * (not by the order they arrive here). This timing change is
418 * essential for VFR with b-frames but a complete nop otherwise.
420 pv->dts_next = in->start;
422 if( in->new_chap && job->chapter_markers )
424 /* chapters have to start with an IDR frame so request that this
425 frame be coded as IDR. Since there may be up to 16 frames
426 currently buffered in the encoder remember the timestamp so
427 when this frame finally pops out of the encoder we'll mark
428 its buffer as the start of a chapter. */
429 pv->pic_in.i_type = X264_TYPE_IDR;
430 if( pv->next_chap == 0 )
432 pv->next_chap = in->start;
433 pv->chap_mark = in->new_chap;
435 /* don't let 'work_loop' put a chapter mark on the wrong buffer */
440 pv->pic_in.i_type = X264_TYPE_AUTO;
442 pv->pic_in.i_qpplus1 = 0;
444 /* XXX this is temporary debugging code to check that the upstream
445 * modules (render & sync) have generated a continuous, self-consistent
446 * frame stream with the current frame's start time equal to the
447 * previous frame's stop time.
449 if( pv->last_stop != in->start )
451 hb_log("encx264 input continuity err: last stop %lld start %lld",
452 pv->last_stop, in->start);
454 pv->last_stop = in->stop;
456 // Remember info about this frame that we need to pass across
457 // the x264_encoder_encode call (since it reorders frames).
458 save_frame_info( pv, in );
460 /* Feed the input DTS to x264 so it can figure out proper output PTS */
461 pv->pic_in.i_pts = in->start;
463 x264_encoder_encode( pv->x264, &nal, &i_nal,
464 &pv->pic_in, &pic_out );
468 x264_encoder_encode( pv->x264, &nal, &i_nal,
470 /* No more delayed B frames */
478 /* Since we output at least one more frame, drop another empty
479 one onto our input fifo. We'll keep doing this automatically
480 until we stop getting frames out of the encoder. */
481 hb_fifo_push(w->fifo_in, hb_buffer_init(0));
487 /* Should be way too large */
488 buf = hb_buffer_init( 3 * job->width * job->height / 2 );
492 /* Get next DTS value to use */
493 int64_t dts_start = pv->dts_next;
495 /* compute the stop time based on the original frame's duration */
496 int64_t dts_stop = dts_start + get_frame_duration( pv, pic_out.i_pts );
497 pv->dts_next = dts_stop;
499 for( i = 0; i < i_nal; i++ )
503 data = buf->alloc - buf->size;
504 if( ( size = x264_nal_encode( buf->data + buf->size, &data,
510 if( job->mux & HB_MUX_AVI )
512 if( nal[i].i_ref_idc == NAL_PRIORITY_HIGHEST )
514 buf->frametype = HB_FRAME_KEY;
521 switch( buf->data[buf->size+4] & 0x1f )
529 /* H.264 in mp4 (stolen from mp4creator) */
530 buf->data[buf->size+0] = ( ( size - 4 ) >> 24 ) & 0xFF;
531 buf->data[buf->size+1] = ( ( size - 4 ) >> 16 ) & 0xFF;
532 buf->data[buf->size+2] = ( ( size - 4 ) >> 8 ) & 0xFF;
533 buf->data[buf->size+3] = ( ( size - 4 ) >> 0 ) & 0xFF;
534 switch( pic_out.i_type )
536 /* Decide what type of frame we have. */
538 buf->frametype = HB_FRAME_IDR;
539 /* if we have a chapter marker pending and this
540 frame's presentation time stamp is at or after
541 the marker's time stamp, use this as the
543 if( pv->next_chap != 0 && pv->next_chap <= pic_out.i_pts )
546 buf->new_chap = pv->chap_mark;
550 buf->frametype = HB_FRAME_I;
553 buf->frametype = HB_FRAME_P;
556 buf->frametype = HB_FRAME_B;
558 /* This is for b-pyramid, which has reference b-frames
559 However, it doesn't seem to ever be used... */
561 buf->frametype = HB_FRAME_BREF;
563 /* If it isn't the above, what type of frame is it?? */
568 /* Since libx264 doesn't tell us when b-frames are
569 themselves reference frames, figure it out on our own. */
570 if( (buf->frametype == HB_FRAME_B) && (nal[i].i_ref_idc != NAL_PRIORITY_DISPOSABLE) )
571 buf->frametype = HB_FRAME_BREF;
573 /* Store the output presentation time stamp
574 from x264 for use by muxmp4 in off-setting
575 b-frames with the CTTS atom. */
576 buf->renderOffset = pic_out.i_pts - dts_start + pv->init_delay;
577 if ( buf->renderOffset < 0 )
579 if ( dts_start - pic_out.i_pts > pv->max_delay )
581 pv->max_delay = dts_start - pic_out.i_pts;
582 hb_log( "encx264: init_delay too small: "
583 "is %lld need %lld", pv->init_delay,
586 buf->renderOffset = 0;
591 /* Send out the next dts values */
592 buf->start = dts_start;
593 buf->stop = dts_stop;