reduce.hpp

namespace detail {

template<class InputIterator, class OutputIterator, class BinaryFunction>
inline void dispatch_gpu_reduce(InputIterator first,
                                InputIterator last,
                                OutputIterator result,
                                BinaryFunction function,
                                command_queue &queue)
{
 ... 
}

template<class InputIterator, class OutputIterator, class BinaryFunction>
inline void dispatch_cpu_reduce(InputIterator first,
                                InputIterator last,
                                OutputIterator result,
                                BinaryFunction function,
                                command_queue &queue)
{
 ... 
}

} // detail namespace

// when InputIterator is a HOST iterator (we will need to copy)
template<class InputIterator, class OutputIterator, class BinaryFunction>
inline void reduce(InputIterator first,
                   InputIterator last,
                   OutputIterator result,
                   BinaryFunction function,
                   command_queue &queue = system::default_queue())
{
    if(small number of elements) { // We may add some option to always force copying and using OpenCL-based algorithm
        // run STL algorithm (why not?)
        // if there's no STL equivalent implement it in C++ or run serial version
    }
    else {
        // copy
        // run device algorithm (below)
    }
}

// when InputIterator is a device iterator
template<class InputIterator, class OutputIterator, class BinaryFunction>
inline void reduce(InputIterator first,
                   InputIterator last,
                   OutputIterator result,
                   BinaryFunction function,
                   command_queue &queue = system::default_queue())
{
    if(small number of elements in the input vector) {
        // run serial algorithm
    }
    
    if(device is a CPU){
        // dispatch a CPU algorithm
    }
    else if (device is a GPU) {
        // dispatch a GPU algorithm
    }
    // some kind of accelerator, we don't know
    else { 
        // run serial algorithm (at least it'll work)
        // for SVM memory we can just map the memory and use STL algorithm
    }
}