Clause getBoxCoverageRequirement(ViewType& view, size_t boxSize, CellType reqCellType) {
Clause req;
for (int row = view.height()-boxSize; row < view.height(); row++) {
for (int col = view.width()-boxSize; col < view.width(); col++) {
req |= view[row][col] == reqCellType;
}
}
return req;
}
KOKKOS_INLINE_FUNCTION
void operator() (const typename ViewType::size_type i) const {
// On CPUs this loop could be vectorized so j should do stride 1
// access on a for optimal performance. I.e. a should be LayoutRight.
// On GPUs threads should do coalesced loads and stores. That means
// that i should be the stride one access for optimal performance.
for (typename ViewType::size_type j = 0; j < a.dimension_1 (); ++j) {
a(i,j) = 1.0*a.dimension_0()*i + 1.0*j;
}
}
string UserInterface::getTaskString(Task* task, int viewType) {
ViewType *taskListType;
switch(viewType) {
case -1:
taskListType = new ViewType();
break;
case 0:
taskListType = new ViewType0();
break;
default:
break;
}
return taskListType->getTaskString(task);
}
ParallelReduce( const FunctorType & arg_functor
, const Policy & arg_policy
, const ViewType & arg_result )
: m_functor( arg_functor )
, m_policy( arg_policy )
, m_result_ptr( arg_result.ptr_on_device() )
, m_shared( arg_policy.scratch_size() + FunctorTeamShmemSize< FunctorType >::value( arg_functor , arg_policy.team_size() ) )
{ }
void UserInterface::printSearchList(string searchTerm, int viewType) {
printNotificationSearchTerm(searchTerm);
ViewType *taskListType;
switch(viewType) {
case -1:
taskListType = new ViewType(_taskList);
break;
case 0:
taskListType = new ViewType0(_taskList);
break;
default:
break;
}
printDisplayList(createDisplayBox(taskListType->createDisplayList()));
delete taskListType;
}
Teuchos::ArrayRCP<typename ViewType::value_type>
persistingView(
const ViewType& view,
typename Teuchos::ArrayRCP<typename ViewType::value_type>::size_type offset,
typename Teuchos::ArrayRCP<typename ViewType::value_type>::size_type size) {
return Teuchos::arcp(view.ptr_on_device()+offset, 0, size,
deallocator(view), false);
}
void UserInterface::printTaskList(int currentDate ,int viewType) {
ViewType *taskListType;
switch(viewType) {
case -1:
taskListType = new ViewType(_taskList , currentDate);
break;
case 0:
taskListType = new ViewType0(_taskList , currentDate);
break;
case 1:
taskListType = new ViewType1(_taskList , currentDate);
break;
default:
break;
}
printDisplayList(createDisplayBox(taskListType->createDisplayList()));
delete taskListType;
}
void draw_v_disparity_line(ViewType &view, PixelType &color, const GroundPlaneEstimator::line_t &v_disparity_line)
{
const int x1 = 0, y1 = v_disparity_line.origin()(0);
const int x2 = view.width(),
y2 = v_disparity_line.origin()(0) + x2*v_disparity_line.direction()(0);
draw_line(view, color, x1, y1, x2, y2);
// printf("draw_v_disparity_line (origin, direction) == (%.2f, %.2f) -> (x2, y2) == (%i, %i)\n",
// v_disparity_line.origin()(0), v_disparity_line.direction()(0),
// x2, y2);
return;
}
Teuchos::ArrayRCP<typename ViewType::value_type>
persistingView(const ViewType& view) {
// mfh (05 Sep 2013) We use a custom deallocator that keeps the
// view, but otherwise does nothing. It doesn't deallocate the
// view; it just maintains the reference count for the lifetime
// of the deallocator object (and therefore, for the lifetime of
// the ArrayRCP). It seems weird that this is a nonowning
// ArrayRCP (the last argument is false). However, that's OK,
// because (a) the ArrayRCP (actually, the underlying
// RCPNodeTmpl; see Teuchos_RCPNode.hpp) keeps the deallocator
// object regardless of whether the ArrayRCP is owning, and (b)
// if we make this an owning ArrayRCP, the RCPNode tracking
// system (in a debug build: Teuchos_ENABLE_DEBUG:BOOL=ON) will
// report problems, because multiple owning ArrayRCPs have the
// same raw pointer but not the same underlying RCPNode. That
// would be a problem if those ArrayRCPs shared memory that was
// allocated (e.g.,) using new or malloc, but it's not a problem
// here, because the custom deallocator does not free anything.
// Nevertheless, it's better not to trouble the tracking system.
return Teuchos::arcp(view.ptr_on_device(), 0, view.capacity(),
deallocator(view), false);
}
ParallelReduce( const FunctorType & functor
, const Policy & policy
, const ViewType & result )
: m_func( functor )
, m_policy( policy )
{
ThreadsExec::resize_shared_scratch( FunctorShmemSize< FunctorType >::value( functor ) );
ThreadsExec::resize_reduce_scratch( Reduce::value_size( m_func ) );
ThreadsExec::start( & ParallelReduce::execute , this , policy.league_size() , policy.team_size() );
const pointer_type data = (pointer_type) ThreadsExec::root_reduce_scratch();
ThreadsExec::fence();
const unsigned n = Reduce::value_count( m_func );
for ( unsigned i = 0 ; i < n ; ++i ) { result.ptr_on_device()[i] = data[i]; }
}
MPVectorAtomicFunctor( const ViewType & v , const scalar_type & s ) : m_v( v ), m_s( s )
{
Kokkos::parallel_for( m_v.dimension_0() , *this );
}
// Kernel launch
static void apply(const ViewType& v, const ScalarType& s) {
const size_type nrow = v.dimension_0();
Kokkos::parallel_for( nrow, ScalarAssignKernel(v,s) );
}
GetMeanValsFunc(const ViewType& vals)
{
mean_vals = ViewType("mean-values", vals.dimension_0());
Kokkos::deep_copy( mean_vals, vals );
}
size_t getStride2DView (ViewType A) {
size_t stride[8];
A.stride (stride);
return A.dimension_1 () > 1 ? stride[1] : A.dimension_0 ();
}
inline bool isInside(const ViewType& v, const CoordType i, const CoordType j)
{
return i>=0 && j>=0 && i<v.width() && j<v.height();
}
Teuchos::ArrayView<const typename ViewType::value_type>
getConstArrayView(const ViewType& a) {
return Teuchos::ArrayView<const typename ViewType::value_type>(
a.ptr_on_device(), a.size());
}