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;
}
}
MPVectorAtomicFunctor( const ViewType & v , const scalar_type & s ) : m_v( v ), m_s( s )
{
Kokkos::parallel_for( m_v.dimension_0() , *this );
}
GetMeanValsFunc(const ViewType& vals)
{
mean_vals = ViewType("mean-values", vals.dimension_0());
Kokkos::deep_copy( mean_vals, vals );
}
// 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) );
}
size_t getStride2DView (ViewType A) {
size_t stride[8];
A.stride (stride);
return A.dimension_1 () > 1 ? stride[1] : A.dimension_0 ();
}