static void exec(LHS &lhs, RHS const &rhs)
{
expr::evaluate(rhs);
length_type const size = lhs.size(1, 0);
for (index_type i=0; i<size; ++i)
lhs.put(i, rhs.get(i));
}
static void exec(LHS &lhs, RHS const &rhs)
{
expr::evaluate(rhs);
length_type const rows = lhs.size(2, 0);
length_type const cols = lhs.size(2, 1);
for (index_type j=0; j<cols; ++j)
for (index_type i=0; i<rows; ++i)
lhs.put(i, j, rhs.get(i, j));
}
static void exec(LHS &lhs, RHS const &rhs)
{
expr::evaluate(rhs);
length_type const size0 = lhs.size(3, 0);
length_type const size1 = lhs.size(3, 1);
length_type const size2 = lhs.size(3, 2);
for (index_type i=0; i<size0; ++i)
for (index_type k=0; k<size2; ++k)
for (index_type j=0; j<size1; ++j)
lhs.put(i, j, k, rhs.get(i, j, k));
}
binary_operator<
typename LHS::const_iterator,
typename RHS::const_iterator,
Op,
NA
> operator()(const LHS &lhs, const RHS &rhs, const Op &op) {
return binary_operator<
typename LHS::const_iterator,
typename RHS::const_iterator,
Op,
NA
>(lhs.begin(), lhs.end(), hooks::extract_dims(lhs), rhs.begin(),
rhs.end(), hooks::extract_dims(rhs), op);
}
void operator()() {
LHS lhs;
RHS rhs;
auto lhs_it = lhs.begin();
auto rhs_it = rhs.begin();
const auto lhs_end = lhs.end();
const auto rhs_end = rhs.end();
for( ; lhs_it != lhs_end && rhs_it != rhs_end; ) {
std::tuple<typename LHS::rtype,typename RHS::rtype> t(*lhs_it, *rhs_it);
//std::cout << "lhs_it="<<(*lhs_it)<<","<<(*rhs_it)<<std::endl;
this->yield(t);
++lhs_it;
++rhs_it;
}
}};
static void exec(LHS &lhs, RHS const &rhs)
{
if (parallel::has_same_map<D>(lhs.map(), rhs))
{
// Maps are same, no communication required.
typedef typename distributed_local_block<LHS>::type lhs_local_block_type;
typedef typename distributed_local_block<RHS>::type rhs_local_block_type;
typedef typename block_traits<lhs_local_block_type>::plain_type
lhs_local_storage_type;
typedef typename block_traits<rhs_local_block_type>::plain_type
rhs_local_storage_type;
lhs_local_storage_type lhs_local_block = get_local_block(lhs);
rhs_local_storage_type rhs_local_block = get_local_block(rhs);
Dispatcher<D, lhs_local_block_type, rhs_local_block_type>::exec(lhs_local_block,
rhs_local_block);
}
else
{
// Maps are different, fall out to general expression.
lhs_view_type lhs_view(lhs);
rhs_view_type rhs_view(const_cast<RHS&>(rhs));
parallel::expr(lhs_view, rhs_view);
}
}
boost::logic::tribool evaluate(const typed_map& map) const {
boost::logic::tribool result = lhs.valid(map) && rhs.valid(map);
if (!result)
return boost::logic::indeterminate;
OP op;
return op(lhs(map),rhs(map));
}
static void exec(LHS &lhs, RHS const &rhs)
{
using namespace impl;
typedef typename simd::LValue_access_traits<typename LHS::value_type> WAT;
typedef typename simd::Proxy_factory<RHS, false>::access_traits EAT;
length_type const vec_size =
simd::Simd_traits<typename LHS::value_type>::vec_size;
Ext_data<LHS, layout_type> dda(lhs, SYNC_OUT);
simd::Proxy<WAT,true> lp(dda.data());
simd::Proxy<EAT,false> rp(simd::Proxy_factory<RHS,false>::create(rhs));
length_type const size = dda.size(0);
length_type n = size;
// loop using proxy interface. This generates the best code
// with gcc 3.4 (with gcc 4.1 the difference to the first case
// above is negligible).
while (n >= vec_size)
{
lp.store(rp.load());
n -= vec_size;
lp.increment();
rp.increment();
}
// Process the remainder, using simple loop fusion.
for (index_type i = size - n; i != size; ++i) lhs.put(i, rhs.get(i));
}
static void exec(LHS &lhs, RHS const &rhs)
{
VBlock const& vblock = rhs.get_vblk();
MBlock const& mblock = rhs.get_mblk();
typedef typename get_block_layout<LHS>::order_type order_type;
Matrix<lhs_value_type, LHS> m_dst(lhs);
const_Vector<v_value_type, VBlock> v(const_cast<VBlock&>(vblock));
const_Matrix<m_value_type, MBlock> m(const_cast<MBlock&>(mblock));
if (is_same<order_type, row2_type>::value)
{
if (D == row)
{
for (index_type r = 0; r < lhs.size(2, 0); ++r)
m_dst.row(r) = v * m.row(r);
}
else
{
for (index_type r = 0; r < lhs.size(2, 0); ++r)
m_dst.row(r) = v.get(r) * m.row(r);
}
}
else // col2_type
{
if (D == row)
{
for (index_type c = 0; c < lhs.size(2, 1); ++c)
m_dst.col(c) = v.get(c) * m.col(c);
}
else
{
for (index_type c = 0; c < lhs.size(2, 1); ++c)
m_dst.col(c) = v * m.col(c);
}
}
}
static void exec(LHS &lhs, RHS const &rhs)
{
if (parallel::has_same_map<D>(lhs.map(), rhs))
{
// Maps are same, no communication required.
typedef typename distributed_local_block<LHS>::type lhs_local_block_type;
typedef typename distributed_local_block<RHS>::type rhs_local_block_type;
typedef typename block_traits<lhs_local_block_type>::plain_type
lhs_local_storage_type;
typedef typename block_traits<rhs_local_block_type>::plain_type
rhs_local_storage_type;
lhs_local_storage_type lhs_local_block = get_local_block(lhs);
rhs_local_storage_type rhs_local_block = get_local_block(rhs);
Dispatcher<D, lhs_local_block_type, rhs_local_block_type>::exec(lhs_local_block,
rhs_local_block);
}
else
{
OVXX_DO_THROW(unimplemented("Expression cannot be reorganized"));
}
}
void checkDeleter(LHS& lhs, RHS& rhs, int LHSState, int RHSState) {
assert(lhs.get_deleter().state() == LHSState);
assert(rhs.get_deleter().state() == RHSState);
}
//Standard case: size1 from lhs, size2 from rhs (fits most cases)
static vcl_size_t size1(LHS & lhs, RHS & /*rhs*/) { return lhs.size1(); }
//Standard case: size1 from lhs, size2 from rhs (fits most cases)
static std::size_t size1(LHS & lhs, RHS & /*rhs*/) { return lhs.size1(); }
//Standard case: LHS is the vector type and carries the correct size
static unsigned int size(LHS & lhs, RHS & rhs) { return lhs.size(); }
//Standard case: size1 from lhs, size2 from rhs (fits most cases)
static size_t size1(LHS & lhs, RHS & rhs) { return lhs.size1(); }
