void operator()(std::size_t begin, std::size_t size)
{
extent_type ext = in_.extent();
std::size_t ibound = boost::fusion::at_c<0>(ext);
std::size_t mbound = boost::fusion::at_c<1>(ext);
for( std::size_t i=0, i_ = o_ + begin; i<size; ++i, ++i_)
{
value_type summary = neutral_(meta::as_<value_type>());
summary = details::scan_step( summary
, out_, in_
, bop_
, i_, mbound, ibound
, false);
}
}
void operator()(std::size_t begin, std::size_t size) const
{
extent_type ext = in_.extent();
std::size_t ibound = boost::fusion::at_c<0>(ext);
std::size_t mbound = boost::fusion::at_c<1>(ext);
std::size_t obound = boost::fusion::at_c<2>(ext);
std::size_t iboundxmbound = ibound * mbound;
std::size_t cache_line_size = nt2::config::top_cache_line_size(2)/sizeof(value_type);
std::size_t grain = cache_line_size;
// Compute the lower multiple of grain of mbound
std::size_t mmbound = (mbound/grain)*grain;
if( ibound < grain )
{
// Instanciate the spawner/worker associated to the mbound dimension
nt2::spawner<tag::scan_, BackEnd, value_type> s;
for(std::size_t o = 0, o_ = begin * iboundxmbound;
o < size;
++o, o_+= iboundxmbound)
{
for(std::size_t i = 0, i_ = o_;
i < ibound;
++i, ++i_)
{
nt2::worker<tag::outer_scan_step_outcache_,BackEnd,Site,Out,In,Neutral,Bop>
w(out_,in_,neutral_,bop_,i_);
value_type s_out = neutral_(nt2::meta::as_<value_type>());
if( (size == obound) && (grain < mmbound) )
s_out = s( w, 0, mmbound, grain);
else if (mmbound != 0)
s_out = w(s_out, 0, mmbound, false);
s_out = w(s_out, mmbound, mbound-mmbound, false);
}
}
}
else
{
// Instanciate the spawner/worker associated to the ibound dimension
nt2::spawner<tag::transform_, BackEnd> s;
for(std::size_t o = 0, o_ = begin * iboundxmbound;
o < size;
++o, o_+= iboundxmbound)
{
nt2::worker<tag::outer_scan_step_incache_,BackEnd,Site,Out,In,Neutral,Bop>
w(out_,in_,neutral_,bop_, o_);
if( size == obound )
s(w,0,ibound,grain);
else
w(0,ibound);
}
}
}
void operator()(std::size_t begin, std::size_t size) const
{
extent_type ext = in_.extent();
static const std::size_t N = boost::simd::meta::cardinal_of<target_type>::value;
std::size_t ibound = boost::fusion::at_c<0>(ext);
std::size_t mbound = boost::fusion::at_c<1>(ext);
std::size_t obound = boost::fusion::at_c<2>(ext);
std::size_t iboundxmbound = ibound * mbound;
std::size_t cache_line_size = nt2::config::top_cache_line_size(2) / sizeof(value_type);
std::size_t grain = cache_line_size;
// Compute the lower multiple of cache_line of ibound
std::size_t cache_bound = (cache_line_size / (sizeof(value_type)*N))*N;
std::size_t iibound = boost::simd::align_under(ibound, cache_bound);
// Compute the lower multiple of grain of mbound
std::size_t mmbound = (mbound/grain)*grain;
if( iibound < grain )
{
// Instanciate the spawner/worker associated to the mbound dimension
nt2::spawner<tag::fold_, BackEnd, target_type> s_simd;
nt2::spawner<tag::fold_, BackEnd, value_type> s_scalar;
nt2::worker<tag::outer_fold_step_outcache_,BackEnd,Site,In,Neutral,Bop>
w(in_,neutral_,bop_);
for(std::size_t o = begin, oout_ = begin*ibound, oin_ = begin * iboundxmbound;
o < begin + size;
++o, oout_+=ibound, oin_+= iboundxmbound)
{
// parallelized part
for (std::size_t i = 0, kout_ = oout_, kin_ = oin_;
i < iibound;
i+=N, kout_+=N, kin_+=N)
{
w.update(kin_);
target_type result = neutral_(nt2::meta::as_<target_type>());
if( (size == obound) && (grain < mmbound) )
result = s_simd( w, 0, mmbound, grain);
else if(mmbound != 0)
result = w(result, 0, mmbound);
result = w(result, mmbound, mbound-mmbound);
nt2::run(out_, kout_, result);
}
// scalar part
for(std::size_t i = iibound, kout_ = oout_ + iibound, kin_ = oin_ + iibound;
i < ibound;
++i, ++kout_, ++kin_)
{
w.update(kin_);
value_type result = neutral_(nt2::meta::as_<value_type>());
if( (size == obound) && (grain < mmbound) )
result = s_scalar(w, 0, mmbound, grain);
else if(mmbound != 0)
result = w(result, 0, mmbound);
result = w(result, mmbound, mbound-mmbound);
nt2::run(out_, kout_, result);
}
}
}
else
{
// Instanciate the spawner/worker associated to the ibound dimension
nt2::spawner<tag::transform_, BackEnd> s;
// vectorized worker
nt2::worker<tag::outer_fold_step_incache_,BackEnd,Site,Out,In,Neutral,Bop>
w1(out_,in_,neutral_,bop_);
// scalar worker
nt2::worker<tag::outer_fold_step_incache_,BackEnd,tag::cpu_,Out,In,Neutral,Bop>
w2(out_,in_,neutral_,bop_);
for(std::size_t o = 0, oout_ = begin*ibound, oin_ = begin * iboundxmbound;
o < size;
++o, oout_+=ibound, oin_+= iboundxmbound)
{
w1.update(oout_, oin_);
w2.update(oout_, oin_);
// parallelized part
if((size == obound) && (grain < iibound))
s(w1,0,iibound,grain);
else if(iibound != 0)
w1(0,iibound);
// scalar part
w2(iibound,ibound-iibound);
}
}
}
