[Boost-users] [compute] extending meta_kernel functionality

Subject: [Boost-users] [compute] extending meta_kernel functionality
From: Oswin Krause (Oswin.Krause_at_[hidden])
Date: 2016-09-20 05:43:53

Hi,

I am trying to use boost.compute to extend my linear algebra library to 
GPUs. so far it helped me a lot with vectors as i can directly use the 
iterator interface. Matrices are harder.

for matrices i need to overload the operator

matrix(i,j)-> return the element (i,j).

For this i have to return an expression which then offers operator<< 
with meta_kernel, for example

template<class T, class IndexExpr1, class IndexExpr2>
inline meta_kernel& operator<<(boost::compute::detail::meta_kernel 
&kernel,
                                const matrix_element_expr<T, IndexExpr1, 
IndexExpr2> &expr)
{
     return kernel << kernel.get_buffer_identifier<T>(expr.m_buffer, 
expr.m_address_space) <<
                    '[' << expr.m_expr1 <<"*" 
<<expr.m_stride1<<"+"<<expr.m_expr2<< "*"<<expr.m_stride2 << ']';
}

Unfortunately all of this is in namespace boost::compute::detail, making 
it a no-go area. Therefore two questions:

1. How stable is i?/can meta_kernel be moved into a normal namespace?
2. Is there a different way to achieve this? (Note that the code above 
is only an example. I could hack it in this case using the return type 
of buffer_iterator but not in all cases unfortunately...)