void PictureCompressor::ModeDecisionME( EncQueue& my_buffer, int pnum )
{
MEData& me_data = my_buffer.GetPicture(pnum).GetMEData();
PictureParams& pparams = my_buffer.GetPicture(pnum).GetPparams();
PicturePredParams& predparams = me_data.GetPicPredParams();
ModeDecider my_mode_dec( m_encparams );
my_mode_dec.DoModeDecn( my_buffer , pnum );
const int num_refs = pparams.NumRefs();
if (m_orig_prec == MV_PRECISION_PIXEL)
{
// FIXME: HACK HACK
// Divide the motion vectors by 2 to convert back to pixel
// accurate motion vectors and reset MV precision to
// PIXEL accuracy
MvArray &mv_arr1 = me_data.Vectors(1);
for (int j = 0; j < mv_arr1.LengthY(); ++j)
{
for (int i = 0; i < mv_arr1.LengthX(); ++i)
mv_arr1[j][i] = mv_arr1[j][i] >> 1;
}
if (num_refs > 1)
{
MvArray &mv_arr2 = me_data.Vectors(2);
for (int j = 0; j < mv_arr2.LengthY(); ++j)
{
for (int i = 0; i < mv_arr2.LengthX(); ++i)
mv_arr2[j][i] = mv_arr2[j][i]>>1;
}
}
void PictureCompressor::NormaliseComplexity( EncQueue& my_buffer, int pnum )
{
EncPicture& my_picture = my_buffer.GetPicture( pnum );
if ( (my_picture.GetStatus()&DONE_PIC_COMPLEXITY) != 0 ){
std::vector<int> queue_members = my_buffer.Members();
double mean_complexity = 0.0;
int count = 0;
for (size_t i=0; i<queue_members.size(); ++ i){
int n = queue_members[i];
EncPicture& enc_pic = my_buffer.GetPicture( n );
if ( (enc_pic.GetStatus()&DONE_PIC_COMPLEXITY) != 0
&& enc_pic.GetPparams().PicSort().IsInter()
&& n >= pnum - 10
&& n <= pnum + 10){
mean_complexity += enc_pic.GetComplexity();
count++;
}
}
mean_complexity /= count;
my_picture.SetNormComplexity( my_picture.GetComplexity() / mean_complexity );
}
}
void PictureCompressor::SubPixelME( EncQueue& my_buffer , int pnum )
{
const std::vector<int>& refs = my_buffer.GetPicture(pnum).GetPparams().Refs();
const int num_refs = refs.size();
PictureParams& pparams = my_buffer.GetPicture(pnum).GetPparams();
MEData& me_data = my_buffer.GetPicture(pnum).GetMEData();
PicturePredParams& predparams = me_data.GetPicPredParams();
float lambda;
if ( pparams.IsBPicture())
lambda = m_encparams.L2MELambda();
else
lambda = m_encparams.L1MELambda();
//lambda *= my_buffer.GetPicture(pnum).GetNormComplexity();
// Set up the lambda to be used
me_data.SetLambdaMap( num_refs , lambda );
m_orig_prec = predparams.MVPrecision();
// Step 2.
// Pixel accurate vectors are then refined to sub-pixel accuracy
if (m_orig_prec != MV_PRECISION_PIXEL)
{
SubpelRefine pelrefine( m_encparams );
pelrefine.DoSubpel( my_buffer , pnum );
}
else
{
// FIXME: HACK HACK
// Mutiplying the motion vectors by 2 and setting MV precision to
// HALF_PIXEL to implement pixel accurate motion estimate
MvArray &mv_arr1 = me_data.Vectors(1);
for (int j = 0; j < mv_arr1.LengthY(); ++j)
{
for (int i = 0; i < mv_arr1.LengthX(); ++i)
mv_arr1[j][i] = mv_arr1[j][i] << 1;
}
if (num_refs > 1)
{
MvArray &mv_arr2 = me_data.Vectors(2);
for (int j = 0; j < mv_arr2.LengthY(); ++j)
{
for (int i = 0; i < mv_arr2.LengthX(); ++i)
mv_arr2[j][i] = mv_arr2[j][i] << 1;
}
}
predparams.SetMVPrecision(MV_PRECISION_HALF_PIXEL);
}
}
void PictureCompressor::CalcComplexity2( EncQueue& my_buffer, int pnum )
{
// to be used after doing motion compensation
EncPicture& my_picture = my_buffer.GetPicture( pnum );
const PicArray& pic_data = my_picture.Data( Y_COMP );
if ( (my_picture.GetStatus()&DONE_MC) != 0 ){
float cost;
double total_sq_cost = 0.0;
double total_cost = 0.0;
for (int j=0; j<pic_data.LengthY(); ++j){
for (int i=0; i<pic_data.LengthX(); ++i){
cost = float( pic_data[j][i] );
total_cost += cost;
total_sq_cost += cost*cost;
}
}
total_cost /= ( pic_data.LengthX()*pic_data.LengthY() );
total_sq_cost /= ( pic_data.LengthX()*pic_data.LengthY() );
my_picture.SetComplexity( total_sq_cost - total_cost*total_cost );
}
}
void SubpelRefine::DoSubpel( EncQueue& my_buffer,int pic_num )
{
m_predparams = &(my_buffer.GetPicture(pic_num).GetMEData().GetPicPredParams() );
//main loop for the subpel refinement
int ref1,ref2;
const PictureSort psort = my_buffer.GetPicture(pic_num).GetPparams().PicSort();
if (psort.IsInter())
{
// Get the references
const vector<int>& refs = my_buffer.GetPicture(pic_num).GetPparams().Refs();
int num_refs = refs.size();
ref1 = refs[0];
if (num_refs>1)
ref2 = refs[1];
else
ref2 = ref1;
const PicArray& pic_data = my_buffer.GetPicture(pic_num).DataForME(m_encparams.CombinedME());
const PicArray& refup1_data = my_buffer.GetPicture(ref1).UpDataForME(m_encparams.CombinedME());
const PicArray& refup2_data = my_buffer.GetPicture(ref2).UpDataForME(m_encparams.CombinedME());
MEData& me_data = my_buffer.GetPicture(pic_num).GetMEData();
// Now match the pictures
MatchPic( pic_data , refup1_data , me_data ,1 );
if (ref1 != ref2 )
MatchPic( pic_data , refup2_data , me_data ,2 );
}
}
void ModeDecider::DoModeDecn( EncQueue& my_buffer, int pic_num )
{
// We've got 'raw' block motion vectors for up to two reference pictures. Now we want
// to make a decision as to mode. In this initial implementation, this is bottom-up
// i.e. find mvs for MBs and sub-MBs and see whether it's worthwhile merging.
int ref1,ref2;
// Initialise //
////////////////
m_psort = my_buffer.GetPicture(pic_num).GetPparams().PicSort();
if (m_psort.IsInter())
{
// Extract the references
const vector<int>& refs = my_buffer.GetPicture(pic_num).GetPparams().Refs();
num_refs = refs.size();
ref1 = refs[0];
// The picture we're doing estimation from
m_pic_data = &(my_buffer.GetPicture( pic_num ).OrigData(Y_COMP));
// Set up the hierarchy of motion vector data objects
PicturePredParams predparams0 = m_predparams;
predparams0.SetXNumBlocks( m_predparams.XNumBlocks()/4 );
predparams0.SetYNumBlocks( m_predparams.YNumBlocks()/4 );
PicturePredParams predparams1 = m_predparams;
predparams1.SetXNumBlocks( m_predparams.XNumBlocks()/2 );
predparams1.SetYNumBlocks( m_predparams.YNumBlocks()/2 );
m_me_data_set[0] = new MEData( predparams0, num_refs );
m_me_data_set[1] = new MEData( predparams1, num_refs );
m_me_data_set[2] = &my_buffer.GetPicture(pic_num).GetMEData();
// Set up the lambdas to use per block
m_me_data_set[0]->SetLambdaMap( 0 , m_me_data_set[2]->LambdaMap() , 1.0/m_level_factor[0] );
m_me_data_set[1]->SetLambdaMap( 1 , m_me_data_set[2]->LambdaMap() , 1.0/m_level_factor[1] );
// Set up the reference pictures
m_ref1_updata = &(my_buffer.GetPicture( ref1 ).UpOrigData(Y_COMP));
if (num_refs>1)
{
ref2 = refs[1];
m_ref2_updata = &(my_buffer.GetPicture( ref2).UpOrigData(Y_COMP));
// Create an object for computing bi-directional prediction calculations
if ( m_predparams.MVPrecision()==MV_PRECISION_EIGHTH_PIXEL )
m_bicheckdiff = new BiBlockEighthPel( *m_ref1_updata ,
*m_ref2_updata ,
*m_pic_data );
else if ( m_predparams.MVPrecision()==MV_PRECISION_QUARTER_PIXEL )
m_bicheckdiff = new BiBlockQuarterPel( *m_ref1_updata ,
*m_ref2_updata ,
*m_pic_data );
else
m_bicheckdiff = new BiBlockHalfPel( *m_ref1_updata ,
*m_ref2_updata ,
*m_pic_data );
}
else
{
ref2 = ref1;
}
// Create an object for doing intra calculations
m_intradiff = new IntraBlockDiff( *m_pic_data );
// Loop over all the macroblocks, doing the work //
///////////////////////////////////////////////////
for (m_ymb_loc=0 ; m_ymb_loc<m_predparams.YNumMB() ; ++m_ymb_loc )
{
for (m_xmb_loc=0 ; m_xmb_loc<m_predparams.XNumMB(); ++m_xmb_loc )
{
DoMBDecn();
}//m_xmb_loc
}//m_ymb_loc
delete m_intradiff;
if (num_refs>1)
delete m_bicheckdiff;
}
// Finally, although not strictly part of motion estimation,
// we have to assign DC values for chroma components for
// blocks we're decided are intra.
SetChromaDC( my_buffer , pic_num );
}
void ModeDecider::DoModeDecn( EncQueue& my_buffer, int pic_num )
{
m_predparams = &(my_buffer.GetPicture(pic_num).GetMEData().GetPicPredParams() );
// The following factors normalise costs for sub-SBs and SBs to those of
// blocks, so that the overlap is take into account (e.g. a sub-SB has
// length XBLEN+XBSEP and YBLEN+YBSEP). The SB costs for a 1x1
// decomposition are not directly comprable to those for other decompositions
// because of the block overlaps. These factors remove these effects, so that
// all SAD costs are normalised to the area corresponding to non-overlapping
// 16 blocks of size XBLEN*YBLEN.
m_level_factor[0] = float( 16 * m_predparams->LumaBParams(2).Xblen() * m_predparams->LumaBParams(2).Yblen() )/
float( m_predparams->LumaBParams(0).Xblen() * m_predparams->LumaBParams(0).Yblen() );
m_level_factor[1] = float( 4 * m_predparams->LumaBParams(2).Xblen() * m_predparams->LumaBParams(2).Yblen() )/
float( m_predparams->LumaBParams(1).Xblen() * m_predparams->LumaBParams(1).Yblen() );
m_level_factor[2] = 1.0f;
for (int i=0 ; i<=2 ; ++i)
m_mode_factor[i] = 80.0*std::pow(0.8 , 2-i);
// We've got 'raw' block motion vectors for up to two reference pictures. Now we want
// to make a decision as to mode. In this initial implementation, this is bottom-up
// i.e. find mvs for SBs and sub-SBs and see whether it's worthwhile merging.
int ref1,ref2;
// Initialise //
////////////////
m_psort = my_buffer.GetPicture(pic_num).GetPparams().PicSort();
if (m_psort.IsInter())
{
// Extract the references
const vector<int>& refs = my_buffer.GetPicture(pic_num).GetPparams().Refs();
num_refs = refs.size();
ref1 = refs[0];
// The picture we're doing estimation from
m_pic_data = &(my_buffer.GetPicture( pic_num ).DataForME(m_encparams.CombinedME()) );
// Set up the hierarchy of motion vector data objects
PicturePredParams predparams0 = *m_predparams;
predparams0.SetXNumBlocks( m_predparams->XNumBlocks()/4 );
predparams0.SetYNumBlocks( m_predparams->YNumBlocks()/4 );
PicturePredParams predparams1 = *m_predparams;
predparams1.SetXNumBlocks( m_predparams->XNumBlocks()/2 );
predparams1.SetYNumBlocks( m_predparams->YNumBlocks()/2 );
m_me_data_set[0] = new MEData( predparams0, num_refs );
m_me_data_set[1] = new MEData( predparams1, num_refs );
m_me_data_set[2] = &my_buffer.GetPicture(pic_num).GetMEData();
// Set up the lambdas to use per block
m_me_data_set[0]->SetLambdaMap( 0 , m_me_data_set[2]->LambdaMap() , 1.0/m_level_factor[0] );
m_me_data_set[1]->SetLambdaMap( 1 , m_me_data_set[2]->LambdaMap() , 1.0/m_level_factor[1] );
// Set up the reference pictures
m_ref1_updata = &(my_buffer.GetPicture( ref1 ).UpDataForME(m_encparams.CombinedME()) );
if (num_refs>1)
{
ref2 = refs[1];
m_ref2_updata = &(my_buffer.GetPicture( ref2).UpDataForME(m_encparams.CombinedME()) );
// Create an object for computing bi-directional prediction calculations
if ( m_predparams->MVPrecision()==MV_PRECISION_EIGHTH_PIXEL )
m_bicheckdiff = new BiBlockEighthPel( *m_ref1_updata ,
*m_ref2_updata ,
*m_pic_data );
else if ( m_predparams->MVPrecision()==MV_PRECISION_QUARTER_PIXEL )
m_bicheckdiff = new BiBlockQuarterPel( *m_ref1_updata ,
*m_ref2_updata ,
*m_pic_data );
else
m_bicheckdiff = new BiBlockHalfPel( *m_ref1_updata ,
*m_ref2_updata ,
*m_pic_data );
}
else
{
ref2 = ref1;
}
// Create an object for doing intra calculations
m_intradiff = new IntraBlockDiff( *m_pic_data );
// Loop over all the superblocks, doing the work //
///////////////////////////////////////////////////
for (m_ysb_loc=0 ; m_ysb_loc<m_predparams->YNumSB() ; ++m_ysb_loc ){
for (m_xsb_loc=0 ; m_xsb_loc<m_predparams->XNumSB(); ++m_xsb_loc ){
DoSBDecn();
}//m_xsb_loc
}//m_ysb_loc
delete m_intradiff;
if (num_refs>1)
delete m_bicheckdiff;
}
// Finally, although not strictly part of motion estimation,
// we have to assign DC values for
// blocks we're decided are intra.
SetDC( my_buffer , pic_num );
}
void PictureCompressor::CalcComplexity( EncQueue& my_buffer, int pnum , const OLBParams& olbparams )
{
EncPicture& my_picture = my_buffer.GetPicture( pnum );
PictureParams& pparams = my_picture.GetPparams();
if ( (my_picture.GetStatus()&DONE_PEL_ME) != 0 ){
MEData& me_data = my_picture.GetMEData();
TwoDArray<MvCostData>* pcosts1;
TwoDArray<MvCostData>* pcosts2;
pcosts1 = &me_data.PredCosts(1);
if (pparams.NumRefs()>1)
pcosts2 = &me_data.PredCosts(2);
else
pcosts2 = pcosts1;
float cost1, cost2, cost;
double total_cost1 = 0.0;
double total_cost2 = 0.0;
double total_cost = 0.0;
int count1=0;int count=0;
float cost_threshold = float(olbparams.Xblen()*olbparams.Yblen()*10);
for (int j=4; j<pcosts1->LengthY()-4; ++j){
for (int i=4; i<pcosts1->LengthX()-4; ++i){
cost1 = (*pcosts1)[j][i].SAD;
cost2 = (*pcosts2)[j][i].SAD;
cost = std::min(cost1, cost2);
total_cost1 += cost1;
total_cost2 += cost2;
total_cost += cost;
if (pparams.NumRefs()>1 && cost<=cost_threshold){
++count;
if (cost1<=cost2)
++count1;
}
}
}
total_cost1 *= olbparams.Xbsep()*olbparams.Ybsep();
total_cost1 /= olbparams.Xblen()*olbparams.Yblen();
total_cost2 *= olbparams.Xbsep()*olbparams.Ybsep();
total_cost2 /= olbparams.Xblen()*olbparams.Yblen();
if (pparams.NumRefs()>1){
my_picture.SetPredBias(float(count1)/float(count));
}
else
my_picture.SetPredBias(0.5);
total_cost *= olbparams.Xbsep()*olbparams.Ybsep();
total_cost /= olbparams.Xblen()*olbparams.Yblen();
// my_picture.SetComplexity( total_cost );
my_picture.SetComplexity( total_cost*total_cost );
}
}