代码分子转自雷神。
264的预处理与avs3的预处理,不能说一模一样,只能说大同小异叭~
//通过lookahead分析帧类型
void x264_lookahead_get_frames( x264_t *h )
{
if( h->param.i_sync_lookahead )
{ /* We have a lookahead thread, so get frames from there */
x264_pthread_mutex_lock( &h->lookahead->ofbuf.mutex );
while( !h->lookahead->ofbuf.i_size && h->lookahead->b_thread_active )
x264_pthread_cond_wait( &h->lookahead->ofbuf.cv_fill, &h->lookahead->ofbuf.mutex );
x264_lookahead_encoder_shift( h );
x264_pthread_mutex_unlock( &h->lookahead->ofbuf.mutex );
}
else
{ /* We are not running a lookahead thread, so perform all the slicetype decide on the fly */
//currect[]必须为空,next不能为空?
if( h->frames.current[0] || !h->lookahead->next.i_size )
return;
//分析lookahead->next->list帧的类型
x264_stack_align( x264_slicetype_decide, h );
//更新lookahead->last_nonb
x264_lookahead_update_last_nonb( h, h->lookahead->next.list[0] );
int shift_frames = h->lookahead->next.list[0]->i_bframes + 1;
//lookahead->next.list移动到lookahead->ofbuf.list
x264_lookahead_shift( &h->lookahead->ofbuf, &h->lookahead->next, shift_frames );
/* For MB-tree and VBV lookahead, we have to perform propagation analysis on I-frames too. */
if( h->lookahead->b_analyse_keyframe && IS_X264_TYPE_I( h->lookahead->last_nonb->i_type ) )
x264_stack_align( x264_slicetype_analyse, h, shift_frames );
//lookahead->ofbuf.list帧移动到frames->current
x264_lookahead_encoder_shift( h );
}
}
x264_lookahead_get_frames()调用了x264_slicetype_decide()用于确定帧类型
//确定帧的类型(I、B、P)
void x264_slicetype_decide( x264_t *h )
{
x264_frame_t *frames[X264_BFRAME_MAX+2];
x264_frame_t *frm;
int bframes;
int brefs;
if( !h->lookahead->next.i_size )
return;
int lookahead_size = h->lookahead->next.i_size;
//遍历next队列
for( int i = 0; i < h->lookahead->next.i_size; i++ )
{
if( h->param.b_vfr_input )
{
if( lookahead_size-- > 1 )
h->lookahead->next.list[i]->i_duration = 2 * (h->lookahead->next.list[i+1]->i_pts - h->lookahead->next.list[i]->i_pts);
else
h->lookahead->next.list[i]->i_duration = h->i_prev_duration;
}
else
h->lookahead->next.list[i]->i_duration = delta_tfi_divisor[h->lookahead->next.list[i]->i_pic_struct];
h->i_prev_duration = h->lookahead->next.list[i]->i_duration;
h->lookahead->next.list[i]->f_duration = (double)h->lookahead->next.list[i]->i_duration
* h->sps->vui.i_num_units_in_tick
/ h->sps->vui.i_time_scale;
if( h->lookahead->next.list[i]->i_frame > h->i_disp_fields_last_frame && lookahead_size > 0 )
{
h->lookahead->next.list[i]->i_field_cnt = h->i_disp_fields;
h->i_disp_fields += h->lookahead->next.list[i]->i_duration;
h->i_disp_fields_last_frame = h->lookahead->next.list[i]->i_frame;
}
else if( lookahead_size == 0 )
{
h->lookahead->next.list[i]->i_field_cnt = h->i_disp_fields;
h->lookahead->next.list[i]->i_duration = h->i_prev_duration;
}
}
if( h->param.rc.b_stat_read )
{
//b_stat_read在2pass模式的第2遍才不为0
/* Use the frame types from the first pass */
for( int i = 0; i < h->lookahead->next.i_size; i++ )
h->lookahead->next.list[i]->i_type =
x264_ratecontrol_slice_type( h, h->lookahead->next.list[i]->i_frame );
}
else if( (h->param.i_bframe && h->param.i_bframe_adaptive)
|| h->param.i_scenecut_threshold
|| h->param.rc.b_mb_tree
|| (h->param.rc.i_vbv_buffer_size && h->param.rc.i_lookahead) )
x264_slicetype_analyse( h, 0 );//分析帧的类型(I、B、P)
//===========================================================================
for( bframes = 0, brefs = 0;; bframes++ )
{
//从next队列取出1个
frm = h->lookahead->next.list[bframes];
//BREF的处理
if( frm->i_type == X264_TYPE_BREF && h->param.i_bframe_pyramid < X264_B_PYRAMID_NORMAL &&
brefs == h->param.i_bframe_pyramid )
{
//BREF改成B
frm->i_type = X264_TYPE_B;
x264_log( h, X264_LOG_WARNING, "B-ref at frame %d incompatible with B-pyramid %s \n",
frm->i_frame, x264_b_pyramid_names[h->param.i_bframe_pyramid] );
}
/* pyramid with multiple B-refs needs a big enough dpb that the preceding P-frame stays available.
smaller dpb could be supported by smart enough use of mmco, but it's easier just to forbid it. */
else if( frm->i_type == X264_TYPE_BREF && h->param.i_bframe_pyramid == X264_B_PYRAMID_NORMAL &&
brefs && h->param.i_frame_reference <= (brefs+3) )
{
frm->i_type = X264_TYPE_B;
x264_log( h, X264_LOG_WARNING, "B-ref at frame %d incompatible with B-pyramid %s and %d reference frames\n",
frm->i_frame, x264_b_pyramid_names[h->param.i_bframe_pyramid], h->param.i_frame_reference );
}
//Keyframe处理
if( frm->i_type == X264_TYPE_KEYFRAME )
frm->i_type = h->param.b_open_gop ? X264_TYPE_I : X264_TYPE_IDR;
/* Limit GOP size */
if( (!h->param.b_intra_refresh || frm->i_frame == 0) && frm->i_frame - h->lookahead->i_last_keyframe >= h->param.i_keyint_max )
{
if( frm->i_type == X264_TYPE_AUTO || frm->i_type == X264_TYPE_I )
frm->i_type = h->param.b_open_gop && h->lookahead->i_last_keyframe >= 0 ? X264_TYPE_I : X264_TYPE_IDR;
int warn = frm->i_type != X264_TYPE_IDR;
if( warn && h->param.b_open_gop )
warn &= frm->i_type != X264_TYPE_I;
if( warn )
{
x264_log( h, X264_LOG_WARNING, "specified frame type (%d) at %d is not compatible with keyframe interval\n", frm->i_type, frm->i_frame );
frm->i_type = h->param.b_open_gop && h->lookahead->i_last_keyframe >= 0 ? X264_TYPE_I : X264_TYPE_IDR;
}
}
if( frm->i_type == X264_TYPE_I && frm->i_frame - h->lookahead->i_last_keyframe >= h->param.i_keyint_min )
{
if( h->param.b_open_gop )
{
h->lookahead->i_last_keyframe = frm->i_frame; // Use display order
if( h->param.b_bluray_compat )
h->lookahead->i_last_keyframe -= bframes; // Use bluray order
frm->b_keyframe = 1;
}
else
frm->i_type = X264_TYPE_IDR;
}
if( frm->i_type == X264_TYPE_IDR )
{
/* Close GOP */
//设置当前帧为“上一个关键帧”
h->lookahead->i_last_keyframe = frm->i_frame;
frm->b_keyframe = 1;
if( bframes > 0 )
{
bframes--;
h->lookahead->next.list[bframes]->i_type = X264_TYPE_P;
}
}
if( bframes == h->param.i_bframe ||
!h->lookahead->next.list[bframes+1] )
{
if( IS_X264_TYPE_B( frm->i_type ) )
x264_log( h, X264_LOG_WARNING, "specified frame type is not compatible with max B-frames\n" );
if( frm->i_type == X264_TYPE_AUTO
|| IS_X264_TYPE_B( frm->i_type ) )
frm->i_type = X264_TYPE_P;
}
if( frm->i_type == X264_TYPE_BREF )
brefs++;
if( frm->i_type == X264_TYPE_AUTO )
frm->i_type = X264_TYPE_B;
else if( !IS_X264_TYPE_B( frm->i_type ) ) break;
}
if( bframes )
h->lookahead->next.list[bframes-1]->b_last_minigop_bframe = 1;
h->lookahead->next.list[bframes]->i_bframes = bframes;
/* insert a bref into the sequence */
if( h->param.i_bframe_pyramid && bframes > 1 && !brefs )
{
h->lookahead->next.list[bframes/2]->i_type = X264_TYPE_BREF;
brefs++;
}
/* calculate the frame costs ahead of time for x264_rc_analyse_slice while we still have lowres */
if( h->param.rc.i_rc_method != X264_RC_CQP )
{
x264_mb_analysis_t a;
int p0, p1, b;
p1 = b = bframes + 1;
x264_lowres_context_init( h, &a );
frames[0] = h->lookahead->last_nonb;
memcpy( &frames[1], h->lookahead->next.list, (bframes+1) * sizeof(x264_frame_t*) );
if( IS_X264_TYPE_I( h->lookahead->next.list[bframes]->i_type ) )
p0 = bframes + 1;
else // P
p0 = 0;
x264_slicetype_frame_cost( h, &a, frames, p0, p1, b, 0 );
if( (p0 != p1 || bframes) && h->param.rc.i_vbv_buffer_size )
{
/* We need the intra costs for row SATDs. */
x264_slicetype_frame_cost( h, &a, frames, b, b, b, 0 );
/* We need B-frame costs for row SATDs. */
p0 = 0;
for( b = 1; b <= bframes; b++ )
{
if( frames[b]->i_type == X264_TYPE_B )
for( p1 = b; frames[p1]->i_type == X264_TYPE_B; )
p1++;
else
p1 = bframes + 1;
x264_slicetype_frame_cost( h, &a, frames, p0, p1, b, 0 );
if( frames[b]->i_type == X264_TYPE_BREF )
p0 = b;
}
}
}
/* Analyse for weighted P frames */
if( !h->param.rc.b_stat_read && h->lookahead->next.list[bframes]->i_type == X264_TYPE_P
&& h->param.analyse.i_weighted_pred >= X264_WEIGHTP_SIMPLE )
{
x264_emms();
x264_weights_analyse( h, h->lookahead->next.list[bframes], h->lookahead->last_nonb, 0 );
}
/* shift sequence to coded order.
use a small temporary list to avoid shifting the entire next buffer around */
int i_coded = h->lookahead->next.list[0]->i_frame;
if( bframes )
{
int idx_list[] = { brefs+1, 1 };
for( int i = 0; i < bframes; i++ )
{
int idx = idx_list[h->lookahead->next.list[i]->i_type == X264_TYPE_BREF]++;
frames[idx] = h->lookahead->next.list[i];
frames[idx]->i_reordered_pts = h->lookahead->next.list[idx]->i_pts;
}
frames[0] = h->lookahead->next.list[bframes];
frames[0]->i_reordered_pts = h->lookahead->next.list[0]->i_pts;
memcpy( h->lookahead->next.list, frames, (bframes+1) * sizeof(x264_frame_t*) );
}
for( int i = 0; i <= bframes; i++ )
{
h->lookahead->next.list[i]->i_coded = i_coded++;
if( i )
{
x264_calculate_durations( h, h->lookahead->next.list[i], h->lookahead->next.list[i-1], &h->i_cpb_delay, &h->i_coded_fields );
h->lookahead->next.list[0]->f_planned_cpb_duration[i-1] = (double)h->lookahead->next.list[i]->i_cpb_duration *
h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
}
else
x264_calculate_durations( h, h->lookahead->next.list[i], NULL, &h->i_cpb_delay, &h->i_coded_fields );
}
}
x264_slicetype_decide()函数中调用了一个比较重要的函数x264_slicetype_analyse(),用来分析帧的类型
//分析帧的类型(I、B、P)
void x264_slicetype_analyse( x264_t *h, int intra_minigop )
{
x264_mb_analysis_t a;
x264_frame_t *frames[X264_LOOKAHEAD_MAX+3] = { NULL, };
int num_frames, orig_num_frames, keyint_limit, framecnt;
int i_mb_count = NUM_MBS;
int cost1p0, cost2p0, cost1b1, cost2p1;
// 确定最大的搜索长度
// 在我的调试当中, h->lookahead->next.i_size = 4
int i_max_search = X264_MIN( h->lookahead->next.i_size, X264_LOOKAHEAD_MAX );
int vbv_lookahead = h->param.rc.i_vbv_buffer_size && h->param.rc.i_lookahead;
/* For determinism we should limit the search to the number of frames lookahead has for sure
* in h->lookahead->next.list buffer, except at the end of stream.
* For normal calls with (intra_minigop == 0) that is h->lookahead->i_slicetype_length + 1 frames.
* And for I-frame calls (intra_minigop != 0) we already removed intra_minigop frames from there. */
if( h->param.b_deterministic )
i_max_search = X264_MIN( i_max_search, h->lookahead->i_slicetype_length + 1 - intra_minigop );
int keyframe = !!intra_minigop;
assert( h->frames.b_have_lowres );
if( !h->lookahead->last_nonb )
return;
//frames[0]指向上一次的非B帧
frames[0] = h->lookahead->last_nonb;
//frames[] 依次指向 lookahead->next链表中的帧
for( framecnt = 0; framecnt < i_max_search && h->lookahead->next.list[framecnt]->i_type == X264_TYPE_AUTO; framecnt++ )
frames[framecnt+1] = h->lookahead->next.list[framecnt];
x264_lowres_context_init( h, &a );
if( !framecnt )
{
if( h->param.rc.b_mb_tree )
x264_macroblock_tree( h, &a, frames, 0, keyframe );
return;
}
keyint_limit = h->param.i_keyint_max - frames[0]->i_frame + h->lookahead->i_last_keyframe - 1;
orig_num_frames = num_frames = h->param.b_intra_refresh ? framecnt : X264_MIN( framecnt, keyint_limit );
/* This is important psy-wise: if we have a non-scenecut keyframe,
* there will be significant visual artifacts if the frames just before
* go down in quality due to being referenced less, despite it being
* more RD-optimal. */
if( (h->param.analyse.b_psy && h->param.rc.b_mb_tree) || vbv_lookahead )
num_frames = framecnt;
else if( h->param.b_open_gop && num_frames < framecnt )
num_frames++;
else if( num_frames == 0 )
{
frames[1]->i_type = X264_TYPE_I;
return;
}
int num_bframes = 0;
int num_analysed_frames = num_frames;
int reset_start;
//通过scenecut()函数判断是否有场景切换,从而确定I帧
if( h->param.i_scenecut_threshold && scenecut( h, &a, frames, 0, 1, 1, orig_num_frames, i_max_search ) )
{
frames[1]->i_type = X264_TYPE_I;
return;
}
#if HAVE_OPENCL
x264_opencl_slicetype_prep( h, frames, num_frames, a.i_lambda );
#endif
//允许有B帧的时候
if( h->param.i_bframe )
{
if( h->param.i_bframe_adaptive == X264_B_ADAPT_TRELLIS )
{
if( num_frames > 1 )
{
char best_paths[X264_BFRAME_MAX+1][X264_LOOKAHEAD_MAX+1] = {"","P"};
int best_path_index = num_frames % (X264_BFRAME_MAX+1);
/* Perform the frametype analysis. */
for( int j = 2; j <= num_frames; j++ )
x264_slicetype_path( h, &a, frames, j, best_paths );
num_bframes = strspn( best_paths[best_path_index], "B" );
/* Load the results of the analysis into the frame types. */
for( int j = 1; j < num_frames; j++ )
frames[j]->i_type = best_paths[best_path_index][j-1] == 'B' ? X264_TYPE_B : X264_TYPE_P;
}
frames[num_frames]->i_type = X264_TYPE_P;
}
else if( h->param.i_bframe_adaptive == X264_B_ADAPT_FAST )
{
for( int i = 0; i <= num_frames-2; )
{
//i+2作为P帧编码的代价
//注:i+2始终为P帧
cost2p1 = x264_slicetype_frame_cost( h, &a, frames, i+0, i+2, i+2, 1 );
if( frames[i+2]->i_intra_mbs[2] > i_mb_count / 2 )
{
frames[i+1]->i_type = X264_TYPE_P;
frames[i+2]->i_type = X264_TYPE_P;
i += 2;
continue;
}
#if HAVE_OPENCL
if( h->param.b_opencl )
{
int b_work_done = 0;
b_work_done |= x264_opencl_precalculate_frame_cost(h, frames, a.i_lambda, i+0, i+2, i+1 );
b_work_done |= x264_opencl_precalculate_frame_cost(h, frames, a.i_lambda, i+0, i+1, i+1 );
b_work_done |= x264_opencl_precalculate_frame_cost(h, frames, a.i_lambda, i+1, i+2, i+2 );
if( b_work_done )
x264_opencl_flush( h );
}
#endif
//计算代价
//x264_slicetype_frame_cost(,,,p0,p1,b,)
//p0 b p1
//p1!=b为B帧,否则为P帧
// i + 1 作为B帧编码的代价
cost1b1 = x264_slicetype_frame_cost( h, &a, frames, i+0, i+2, i+1, 0 );
// i + 1 作为P帧编码的代价
cost1p0 = x264_slicetype_frame_cost( h, &a, frames, i+0, i+1, i+1, 0 );
// i + 2 作为P帧编码的代价
cost2p0 = x264_slicetype_frame_cost( h, &a, frames, i+1, i+2, i+2, 0 );
//如果i+1作为P帧编码的代价 + i+2作为P帧编码的代价
//小于 i+1作为B帧编码的代价 + i+2作为P帧编码的代价
if( cost1p0 + cost2p0 < cost1b1 + cost2p1 )
{
//那么i+1将作为P帧编码
//然后直接continue
frames[i+1]->i_type = X264_TYPE_P;
i += 1;
continue;
}
// arbitrary and untuned
#define INTER_THRESH 300
#define P_SENS_BIAS (50 - h->param.i_bframe_bias)
// i+1 将作为B帧编码
frames[i+1]->i_type = X264_TYPE_B;
int j;
for( j = i+2; j <= X264_MIN( i+h->param.i_bframe, num_frames-1 ); j++ )
{
int pthresh = X264_MAX(INTER_THRESH - P_SENS_BIAS * (j-i-1), INTER_THRESH/10);
// 预测j+1作为P帧编码代价
int pcost = x264_slicetype_frame_cost( h, &a, frames, i+0, j+1, j+1, 1 );
// 如果pcost 满足下述条件, 则确定了一个P帧,跳出循环
if( pcost > pthresh*i_mb_count || frames[j+1]->i_intra_mbs[j-i+1] > i_mb_count/3 )
break;
// 否则就是B帧
frames[j]->i_type = X264_TYPE_B;
}
// 将j帧确定为P帧
frames[j]->i_type = X264_TYPE_P;
i = j;
}
// 最后一帧确定为P帧
frames[num_frames]->i_type = X264_TYPE_P;
num_bframes = 0;
// 确定有多少个B帧
while( num_bframes < num_frames && frames[num_bframes+1]->i_type == X264_TYPE_B )
num_bframes++;
}
else
{
// 确定多少B帧
num_bframes = X264_MIN(num_frames-1, h->param.i_bframe);
// 每num_bframes + 1一个P帧, 其余皆为B帧
for( int j = 1; j < num_frames; j++ )
frames[j]->i_type = (j%(num_bframes+1)) ? X264_TYPE_B : X264_TYPE_P;
// 最后一帧为P帧
frames[num_frames]->i_type = X264_TYPE_P;
}
/* Check scenecut on the first minigop. */
// 如果B帧中, 有帧有场景切换, 则改变其为P帧
for( int j = 1; j < num_bframes+1; j++ )
if( h->param.i_scenecut_threshold && scenecut( h, &a, frames, j, j+1, 0, orig_num_frames, i_max_search ) )
{
frames[j]->i_type = X264_TYPE_P;
num_analysed_frames = j;
break;
}
reset_start = keyframe ? 1 : X264_MIN( num_bframes+2, num_analysed_frames+1 );
}
else
{
//h->param.i_bframe为 0
//则所有的帧皆为P帧
for( int j = 1; j <= num_frames; j++ )
frames[j]->i_type = X264_TYPE_P;
reset_start = !keyframe + 1;
num_bframes = 0;
}
/* Perform the actual macroblock tree analysis.
* Don't go farther than the maximum keyframe interval; this helps in short GOPs. */
if( h->param.rc.b_mb_tree )
x264_macroblock_tree( h, &a, frames, X264_MIN(num_frames, h->param.i_keyint_max), keyframe );
/* Enforce keyframe limit. */
if( !h->param.b_intra_refresh )
for( int i = keyint_limit+1; i <= num_frames; i += h->param.i_keyint_max )
{
//迫使为I帧
frames[i]->i_type = X264_TYPE_I;
reset_start = X264_MIN( reset_start, i+1 );
if( h->param.b_open_gop && h->param.b_bluray_compat )
while( IS_X264_TYPE_B( frames[i-1]->i_type ) )
i--;
}
if( vbv_lookahead )
x264_vbv_lookahead( h, &a, frames, num_frames, keyframe );
/* Restore frametypes for all frames that haven't actually been decided yet. */
for( int j = reset_start; j <= num_frames; j++ )
frames[j]->i_type = X264_TYPE_AUTO;
#if HAVE_OPENCL
x264_opencl_slicetype_end( h );
#endif
}
通过源代码可以看出,x264_slicetype_analyse()分析了frames[]队列中的视频帧的帧类型。简单总结一下该函数的流程:
(1)如果frames[1]通过scenecut()判断为场景切换,设置为I帧,并且直接返回。
(2)如果i_bframe为0,即不使用B帧,则将所有帧都设置为P帧。
(3)如果i_bframe不为0,即使用B帧,则需要进行比较复杂的帧开销计算。这时候需要调用一帧图像开销的计算函数x264_slicetype_frame_cost()。
继续看一下x264_slicetype_frame_cost()函数
//一帧图像的开销
//x264_slicetype_frame_cost(,,,p0,p1,b,)
// p0 b p1
static int x264_slicetype_frame_cost( x264_t *h, x264_mb_analysis_t *a,
x264_frame_t **frames, int p0, int p1, int b,
int b_intra_penalty )
{
int i_score = 0;
int do_search[2];
const x264_weight_t *w = x264_weight_none;
x264_frame_t *fenc = frames[b];
/* Check whether we already evaluated this frame
* If we have tried this frame as P, then we have also tried
* the preceding frames as B. (is this still true?) */
/* Also check that we already calculated the row SATDs for the current frame. */
//如果已经计算过就不用算了
if( fenc->i_cost_est[b-p0][p1-b] >= 0 && (!h->param.rc.i_vbv_buffer_size || fenc->i_row_satds[b-p0][p1-b][0] != -1) )
i_score = fenc->i_cost_est[b-p0][p1-b];
else
{
int dist_scale_factor = 128;
/* For each list, check to see whether we have lowres motion-searched this reference frame before. */
do_search[0] = b != p0 && fenc->lowres_mvs[0][b-p0-1][0][0] == 0x7FFF;
do_search[1] = b != p1 && fenc->lowres_mvs[1][p1-b-1][0][0] == 0x7FFF;
if( do_search[0] )
{
if( h->param.analyse.i_weighted_pred && b == p1 )
{
x264_emms();
x264_weights_analyse( h, fenc, frames[p0], 1 );
w = fenc->weight[0];
}
fenc->lowres_mvs[0][b-p0-1][0][0] = 0;
}
if( do_search[1] ) fenc->lowres_mvs[1][p1-b-1][0][0] = 0;
if( p1 != p0 )
dist_scale_factor = ( ((b-p0) << 8) + ((p1-p0) >> 1) ) / (p1-p0);
int output_buf_size = h->mb.i_mb_height + (NUM_INTS + PAD_SIZE) * h->param.i_lookahead_threads;
int *output_inter[X264_LOOKAHEAD_THREAD_MAX+1];
int *output_intra[X264_LOOKAHEAD_THREAD_MAX+1];
output_inter[0] = h->scratch_buffer2;
output_intra[0] = output_inter[0] + output_buf_size;
#if HAVE_OPENCL
if( h->param.b_opencl )
{
x264_opencl_lowres_init(h, fenc, a->i_lambda );
if( do_search[0] )
{
x264_opencl_lowres_init( h, frames[p0], a->i_lambda );
x264_opencl_motionsearch( h, frames, b, p0, 0, a->i_lambda, w );
}
if( do_search[1] )
{
x264_opencl_lowres_init( h, frames[p1], a->i_lambda );
x264_opencl_motionsearch( h, frames, b, p1, 1, a->i_lambda, NULL );
}
if( b != p0 )
x264_opencl_finalize_cost( h, a->i_lambda, frames, p0, p1, b, dist_scale_factor );
x264_opencl_flush( h );
i_score = fenc->i_cost_est[b-p0][p1-b];
}
else
#endif
{
if( h->param.i_lookahead_threads > 1 )
{
x264_slicetype_slice_t s[X264_LOOKAHEAD_THREAD_MAX];
for( int i = 0; i < h->param.i_lookahead_threads; i++ )
{
x264_t *t = h->lookahead_thread[i];
/* FIXME move this somewhere else */
t->mb.i_me_method = h->mb.i_me_method;
t->mb.i_subpel_refine = h->mb.i_subpel_refine;
t->mb.b_chroma_me = h->mb.b_chroma_me;
s[i] = (x264_slicetype_slice_t){ t, a, frames, p0, p1, b, dist_scale_factor, do_search, w,
output_inter[i], output_intra[i] };
t->i_threadslice_start = ((h->mb.i_mb_height * i + h->param.i_lookahead_threads/2) / h->param.i_lookahead_threads);
t->i_threadslice_end = ((h->mb.i_mb_height * (i+1) + h->param.i_lookahead_threads/2) / h->param.i_lookahead_threads);
int thread_height = t->i_threadslice_end - t->i_threadslice_start;
int thread_output_size = thread_height + NUM_INTS;
memset( output_inter[i], 0, thread_output_size * sizeof(int) );
memset( output_intra[i], 0, thread_output_size * sizeof(int) );
output_inter[i][NUM_ROWS] = output_intra[i][NUM_ROWS] = thread_height;
output_inter[i+1] = output_inter[i] + thread_output_size + PAD_SIZE;
output_intra[i+1] = output_intra[i] + thread_output_size + PAD_SIZE;
x264_threadpool_run( h->lookaheadpool, (void*)x264_slicetype_slice_cost, &s[i] );
}
for( int i = 0; i < h->param.i_lookahead_threads; i++ )
x264_threadpool_wait( h->lookaheadpool, &s[i] );
}
else
{
h->i_threadslice_start = 0;
h->i_threadslice_end = h->mb.i_mb_height;
memset( output_inter[0], 0, (output_buf_size - PAD_SIZE) * sizeof(int) );
memset( output_intra[0], 0, (output_buf_size - PAD_SIZE) * sizeof(int) );
output_inter[0][NUM_ROWS] = output_intra[0][NUM_ROWS] = h->mb.i_mb_height;
//作为参数的结构体
x264_slicetype_slice_t s = (x264_slicetype_slice_t){ h, a, frames, p0, p1, b, dist_scale_factor, do_search, w,
output_inter[0], output_intra[0] };
//一个slice的开销
//输入输出参数都在s结构体中
x264_slicetype_slice_cost( &s );
}
/* Sum up accumulators */
if( b == p1 )
fenc->i_intra_mbs[b-p0] = 0;
if( !fenc->b_intra_calculated )
{
fenc->i_cost_est[0][0] = 0;
fenc->i_cost_est_aq[0][0] = 0;
}
fenc->i_cost_est[b-p0][p1-b] = 0;
fenc->i_cost_est_aq[b-p0][p1-b] = 0;
int *row_satd_inter = fenc->i_row_satds[b-p0][p1-b];
int *row_satd_intra = fenc->i_row_satds[0][0];
for( int i = 0; i < h->param.i_lookahead_threads; i++ )
{
//累加output_inter[]或output_intra[]
//这2个变量中存储了整帧的开销
if( b == p1 )
fenc->i_intra_mbs[b-p0] += output_inter[i][INTRA_MBS];
if( !fenc->b_intra_calculated )
{
//帧内编码的代价
fenc->i_cost_est[0][0] += output_intra[i][COST_EST];
fenc->i_cost_est_aq[0][0] += output_intra[i][COST_EST_AQ];
}
//帧间编码的代价
fenc->i_cost_est[b-p0][p1-b] += output_inter[i][COST_EST];
fenc->i_cost_est_aq[b-p0][p1-b] += output_inter[i][COST_EST_AQ];
if( h->param.rc.i_vbv_buffer_size )
{
int row_count = output_inter[i][NUM_ROWS];
memcpy( row_satd_inter, output_inter[i] + NUM_INTS, row_count * sizeof(int) );
if( !fenc->b_intra_calculated )
memcpy( row_satd_intra, output_intra[i] + NUM_INTS, row_count * sizeof(int) );
row_satd_inter += row_count;
row_satd_intra += row_count;
}
}
//一帧的开销
i_score = fenc->i_cost_est[b-p0][p1-b];
if( b != p1 )//B帧
i_score = (uint64_t)i_score * 100 / (120 + h->param.i_bframe_bias);
else
fenc->b_intra_calculated = 1;
fenc->i_cost_est[b-p0][p1-b] = i_score;
x264_emms();
}
}
if( b_intra_penalty )
{
// arbitrary penalty for I-blocks after B-frames
int nmb = NUM_MBS;
i_score += (uint64_t)i_score * fenc->i_intra_mbs[b-p0] / (nmb * 8);
}
//返回一帧的开销值
return i_score;
}
x264_slicetype_analyse()调用了x264_slicetype_slice_cost()来计算一个slice的开销。
//一个slice的开销
static void x264_slicetype_slice_cost( x264_slicetype_slice_t *s )
{
x264_t *h = s->h;
/* Lowres lookahead goes backwards because the MVs are used as predictors in the main encode.
* This considerably improves MV prediction overall. */
/* The edge mbs seem to reduce the predictive quality of the
* whole frame's score, but are needed for a spatial distribution. */
int do_edges = h->param.rc.b_mb_tree || h->param.rc.i_vbv_buffer_size || h->mb.i_mb_width <= 2 || h->mb.i_mb_height <= 2;
int start_y = X264_MIN( h->i_threadslice_end - 1, h->mb.i_mb_height - 2 + do_edges );
int end_y = X264_MAX( h->i_threadslice_start, 1 - do_edges );
int start_x = h->mb.i_mb_width - 2 + do_edges;
int end_x = 1 - do_edges;
//逐个计算每个MB的开销
for( h->mb.i_mb_y = start_y; h->mb.i_mb_y >= end_y; h->mb.i_mb_y-- )
for( h->mb.i_mb_x = start_x; h->mb.i_mb_x >= end_x; h->mb.i_mb_x-- )
x264_slicetype_mb_cost( h, s->a, s->frames, s->p0, s->p1, s->b, s->dist_scale_factor,
s->do_search, s->w, s->output_inter, s->output_intra );
}
x264_slicetype_slice_cost()通过循环遍历了每一个宏块,针对每一个宏块调用了x264_slicetype_mb_cost()。处理的图像是经过1/2线性差值的“低分辨率(lowres)”图片,下采样。
//一个MB的开销
static void x264_slicetype_mb_cost( x264_t *h, x264_mb_analysis_t *a,
x264_frame_t **frames, int p0, int p1, int b,
int dist_scale_factor, int do_search[2], const x264_weight_t *w,
int *output_inter, int *output_intra )
{
x264_frame_t *fref0 = frames[p0];
x264_frame_t *fref1 = frames[p1];
x264_frame_t *fenc = frames[b];
const int b_bidir = (b < p1);
const int i_mb_x = h->mb.i_mb_x;
const int i_mb_y = h->mb.i_mb_y;
const int i_mb_stride = h->mb.i_mb_width;
const int i_mb_xy = i_mb_x + i_mb_y * i_mb_stride;
const int i_stride = fenc->i_stride_lowres;
const int i_pel_offset = 8 * (i_mb_x + i_mb_y * i_stride);
const int i_bipred_weight = h->param.analyse.b_weighted_bipred ? 64 - (dist_scale_factor>>2) : 32;
int16_t (*fenc_mvs[2])[2] = { &fenc->lowres_mvs[0][b-p0-1][i_mb_xy], &fenc->lowres_mvs[1][p1-b-1][i_mb_xy] };
int (*fenc_costs[2]) = { &fenc->lowres_mv_costs[0][b-p0-1][i_mb_xy], &fenc->lowres_mv_costs[1][p1-b-1][i_mb_xy] };
int b_frame_score_mb = (i_mb_x > 0 && i_mb_x < h->mb.i_mb_width - 1 &&
i_mb_y > 0 && i_mb_y < h->mb.i_mb_height - 1) ||
h->mb.i_mb_width <= 2 || h->mb.i_mb_height <= 2;
ALIGNED_ARRAY_16( pixel, pix1,[9*FDEC_STRIDE] );
pixel *pix2 = pix1+8;
x264_me_t m[2];
int i_bcost = COST_MAX;
int list_used = 0;
/* A small, arbitrary bias to avoid VBV problems caused by zero-residual lookahead blocks. */
int lowres_penalty = 4;
//计算只涉及一个分量
h->mb.pic.p_fenc[0] = h->mb.pic.fenc_buf;
//从低分辨率(1/2线性内插)图像中拷贝数据
h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fenc[0], FENC_STRIDE, &fenc->lowres[0][i_pel_offset], i_stride, 8 );
if( p0 == p1 )
goto lowres_intra_mb;
// no need for h->mb.mv_min[]
h->mb.mv_limit_fpel[0][0] = -8*h->mb.i_mb_x - 4;
h->mb.mv_limit_fpel[1][0] = 8*( h->mb.i_mb_width - h->mb.i_mb_x - 1 ) + 4;
h->mb.mv_min_spel[0] = 4*( h->mb.mv_limit_fpel[0][0] - 8 );
h->mb.mv_max_spel[0] = 4*( h->mb.mv_limit_fpel[1][0] + 8 );
if( h->mb.i_mb_x >= h->mb.i_mb_width - 2 )
{
h->mb.mv_limit_fpel[0][1] = -8*h->mb.i_mb_y - 4;
h->mb.mv_limit_fpel[1][1] = 8*( h->mb.i_mb_height - h->mb.i_mb_y - 1 ) + 4;
h->mb.mv_min_spel[1] = 4*( h->mb.mv_limit_fpel[0][1] - 8 );
h->mb.mv_max_spel[1] = 4*( h->mb.mv_limit_fpel[1][1] + 8 );
}
#define LOAD_HPELS_LUMA(dst, src) \
{ \
(dst)[0] = &(src)[0][i_pel_offset]; \
(dst)[1] = &(src)[1][i_pel_offset]; \
(dst)[2] = &(src)[2][i_pel_offset]; \
(dst)[3] = &(src)[3][i_pel_offset]; \
}
#define LOAD_WPELS_LUMA(dst,src) \
(dst) = &(src)[i_pel_offset];
#define CLIP_MV( mv ) \
{ \
mv[0] = x264_clip3( mv[0], h->mb.mv_min_spel[0], h->mb.mv_max_spel[0] ); \
mv[1] = x264_clip3( mv[1], h->mb.mv_min_spel[1], h->mb.mv_max_spel[1] ); \
}
#define TRY_BIDIR( mv0, mv1, penalty ) \
{ \
int i_cost; \
if( h->param.analyse.i_subpel_refine <= 1 ) \
{ \
int hpel_idx1 = (((mv0)[0]&2)>>1) + ((mv0)[1]&2); \
int hpel_idx2 = (((mv1)[0]&2)>>1) + ((mv1)[1]&2); \
pixel *src1 = m[0].p_fref[hpel_idx1] + ((mv0)[0]>>2) + ((mv0)[1]>>2) * m[0].i_stride[0]; \
pixel *src2 = m[1].p_fref[hpel_idx2] + ((mv1)[0]>>2) + ((mv1)[1]>>2) * m[1].i_stride[0]; \
h->mc.avg[PIXEL_8x8]( pix1, 16, src1, m[0].i_stride[0], src2, m[1].i_stride[0], i_bipred_weight ); \
} \
else \
{ \
intptr_t stride1 = 16, stride2 = 16; \
pixel *src1, *src2; \
src1 = h->mc.get_ref( pix1, &stride1, m[0].p_fref, m[0].i_stride[0], \
(mv0)[0], (mv0)[1], 8, 8, w ); \
src2 = h->mc.get_ref( pix2, &stride2, m[1].p_fref, m[1].i_stride[0], \
(mv1)[0], (mv1)[1], 8, 8, w ); \
h->mc.avg[PIXEL_8x8]( pix1, 16, src1, stride1, src2, stride2, i_bipred_weight ); \
} \
i_cost = penalty * a->i_lambda + h->pixf.mbcmp[PIXEL_8x8]( \
m[0].p_fenc[0], FENC_STRIDE, pix1, 16 ); \
COPY2_IF_LT( i_bcost, i_cost, list_used, 3 ); \
}
//帧间编码(后面还有帧内编码)
//处理m[0]
m[0].i_pixel = PIXEL_8x8;
m[0].p_cost_mv = a->p_cost_mv;
m[0].i_stride[0] = i_stride;
m[0].p_fenc[0] = h->mb.pic.p_fenc[0];
m[0].weight = w;
m[0].i_ref = 0;
//加载1/2插值像素点
LOAD_HPELS_LUMA( m[0].p_fref, fref0->lowres );
m[0].p_fref_w = m[0].p_fref[0];
if( w[0].weightfn )
LOAD_WPELS_LUMA( m[0].p_fref_w, fenc->weighted[0] );
//双线预测,处理m[1]
if( b_bidir )
{
int16_t *mvr = fref1->lowres_mvs[0][p1-p0-1][i_mb_xy];
ALIGNED_ARRAY_8( int16_t, dmv,[2],[2] );
m[1].i_pixel = PIXEL_8x8;
m[1].p_cost_mv = a->p_cost_mv;
m[1].i_stride[0] = i_stride;
m[1].p_fenc[0] = h->mb.pic.p_fenc[0];
m[1].i_ref = 0;
m[1].weight = x264_weight_none;
LOAD_HPELS_LUMA( m[1].p_fref, fref1->lowres );
m[1].p_fref_w = m[1].p_fref[0];
dmv[0][0] = ( mvr[0] * dist_scale_factor + 128 ) >> 8;
dmv[0][1] = ( mvr[1] * dist_scale_factor + 128 ) >> 8;
dmv[1][0] = dmv[0][0] - mvr[0];
dmv[1][1] = dmv[0][1] - mvr[1];
CLIP_MV( dmv[0] );
CLIP_MV( dmv[1] );
if( h->param.analyse.i_subpel_refine <= 1 )
M64( dmv ) &= ~0x0001000100010001ULL; /* mv & ~1 */
//双向预测,其中包含了mc.avg[PIXEL_8x8]()
TRY_BIDIR( dmv[0], dmv[1], 0 );
if( M64( dmv ) )
{
int i_cost;
h->mc.avg[PIXEL_8x8]( pix1, 16, m[0].p_fref[0], m[0].i_stride[0], m[1].p_fref[0], m[1].i_stride[0], i_bipred_weight );
i_cost = h->pixf.mbcmp[PIXEL_8x8]( m[0].p_fenc[0], FENC_STRIDE, pix1, 16 );
COPY2_IF_LT( i_bcost, i_cost, list_used, 3 );
}
}
for( int l = 0; l < 1 + b_bidir; l++ )
{
if( do_search[l] )
{
int i_mvc = 0;
int16_t (*fenc_mv)[2] = fenc_mvs[l];
ALIGNED_4( int16_t mvc[4][2] );
/* Reverse-order MV prediction. */
M32( mvc[0] ) = 0;
M32( mvc[2] ) = 0;
#define MVC(mv) { CP32( mvc[i_mvc], mv ); i_mvc++; }
if( i_mb_x < h->mb.i_mb_width - 1 )
MVC( fenc_mv[1] );
if( i_mb_y < h->i_threadslice_end - 1 )
{
MVC( fenc_mv[i_mb_stride] );
if( i_mb_x > 0 )
MVC( fenc_mv[i_mb_stride-1] );
if( i_mb_x < h->mb.i_mb_width - 1 )
MVC( fenc_mv[i_mb_stride+1] );
}
#undef MVC
if( i_mvc <= 1 )
CP32( m[l].mvp, mvc[0] );
else
x264_median_mv( m[l].mvp, mvc[0], mvc[1], mvc[2] );
/* Fast skip for cases of near-zero residual. Shortcut: don't bother except in the mv0 case,
* since anything else is likely to have enough residual to not trigger the skip. */
if( !M32( m[l].mvp ) )
{
m[l].cost = h->pixf.mbcmp[PIXEL_8x8]( m[l].p_fenc[0], FENC_STRIDE, m[l].p_fref[0], m[l].i_stride[0] );
if( m[l].cost < 64 )
{
M32( m[l].mv ) = 0;
goto skip_motionest;
}
}
//运动搜索,开销存在m[l].cost中
x264_me_search( h, &m[l], mvc, i_mvc );
m[l].cost -= a->p_cost_mv[0]; // remove mvcost from skip mbs
if( M32( m[l].mv ) )
m[l].cost += 5 * a->i_lambda;
skip_motionest:
CP32( fenc_mvs[l], m[l].mv );
*fenc_costs[l] = m[l].cost;
}
else
{
CP32( m[l].mv, fenc_mvs[l] );
m[l].cost = *fenc_costs[l];
}
//如果更小就拷贝
//帧间编码开销,存储于i_bcost
COPY2_IF_LT( i_bcost, m[l].cost, list_used, l+1 );
}
if( b_bidir && ( M32( m[0].mv ) || M32( m[1].mv ) ) )
TRY_BIDIR( m[0].mv, m[1].mv, 5 );
lowres_intra_mb:
//帧内编码
if( !fenc->b_intra_calculated )
{
ALIGNED_ARRAY_16( pixel, edge,[36] );
pixel *pix = &pix1[8+FDEC_STRIDE];
pixel *src = &fenc->lowres[0][i_pel_offset];
const int intra_penalty = 5 * a->i_lambda;
int satds[3];
int pixoff = 4 / sizeof(pixel);
/* Avoid store forwarding stalls by writing larger chunks */
memcpy( pix-FDEC_STRIDE, src-i_stride, 16 * sizeof(pixel) );
for( int i = -1; i < 8; i++ )
M32( &pix[i*FDEC_STRIDE-pixoff] ) = M32( &src[i*i_stride-pixoff] );
//8x8块的SAD/SATD计算
//x3打表计算了V,H,DC三种模式,开销存储在satds[3]数组的3个元素中
h->pixf.intra_mbcmp_x3_8x8c( h->mb.pic.p_fenc[0], pix, satds );
//帧内编码开销,存储于i_icost
int i_icost = X264_MIN3( satds[0], satds[1], satds[2] );
if( h->param.analyse.i_subpel_refine > 1 )
{
h->predict_8x8c[I_PRED_CHROMA_P]( pix );
int satd = h->pixf.mbcmp[PIXEL_8x8]( pix, FDEC_STRIDE, h->mb.pic.p_fenc[0], FENC_STRIDE );
i_icost = X264_MIN( i_icost, satd );
h->predict_8x8_filter( pix, edge, ALL_NEIGHBORS, ALL_NEIGHBORS );
for( int i = 3; i < 9; i++ )
{
h->predict_8x8[i]( pix, edge );
satd = h->pixf.mbcmp[PIXEL_8x8]( pix, FDEC_STRIDE, h->mb.pic.p_fenc[0], FENC_STRIDE );
i_icost = X264_MIN( i_icost, satd );
}
}
i_icost += intra_penalty + lowres_penalty;
//存一下
fenc->i_intra_cost[i_mb_xy] = i_icost;
int i_icost_aq = i_icost;
if( h->param.rc.i_aq_mode )
i_icost_aq = (i_icost_aq * fenc->i_inv_qscale_factor[i_mb_xy] + 128) >> 8;
output_intra[ROW_SATD] += i_icost_aq;
if( b_frame_score_mb )
{
//累加。[COST_EST]用于整帧的开销计算
output_intra[COST_EST] += i_icost;
output_intra[COST_EST_AQ] += i_icost_aq;
}
}
i_bcost += lowres_penalty;
/* forbid intra-mbs in B-frames, because it's rare and not worth checking */
/* FIXME: Should we still forbid them now that we cache intra scores? */
if( !b_bidir )
{
int i_icost = fenc->i_intra_cost[i_mb_xy];
//帧内开销比帧间更小,b_intra就会取1
int b_intra = i_icost < i_bcost;
if( b_intra )
{
//赋值给i_bcost
i_bcost = i_icost;
list_used = 0;
}
if( b_frame_score_mb )
output_inter[INTRA_MBS] += b_intra;//[INTRA_MBS]统计有多少个帧内模式的宏块
}
/* In an I-frame, we've already added the results above in the intra section. */
if( p0 != p1 )
{
int i_bcost_aq = i_bcost;
if( h->param.rc.i_aq_mode )
i_bcost_aq = (i_bcost_aq * fenc->i_inv_qscale_factor[i_mb_xy] + 128) >> 8;
output_inter[ROW_SATD] += i_bcost_aq;
if( b_frame_score_mb )
{
/* Don't use AQ-weighted costs for slicetype decision, only for ratecontrol. */
//累加。[COST_EST]用于整帧的开销计算
output_inter[COST_EST] += i_bcost;
output_inter[COST_EST_AQ] += i_bcost_aq;
}
}
//存储开销i_bcost
fenc->lowres_costs[b-p0][p1-b][i_mb_xy] = X264_MIN( i_bcost, LOWRES_COST_MASK ) + (list_used << LOWRES_COST_SHIFT);
}
#undef TRY_BIDIR