nrm2.test

PREAMBLE = \
    """
    #include <thrust/transform.h>
    #include <thrust/reduce.h>
    #include <thrust/transform_reduce.h>
    #include <thrust/functional.h>
    
    #include <cmath>

    template <typename T>
    struct square
    {
        __host__ __device__
        T operator()(T x) const
        {
            return x * x;
        }
    };
    
    template <typename Vector>
    typename Vector::value_type nrm2_fast(const Vector& x)
    {
        typedef typename Vector::value_type T;
        return std::sqrt( thrust::transform_reduce(x.begin(), x.end(), square<T>(), T(0), thrust::plus<T>()) );
    }
    
    template <typename Vector>
    typename Vector::value_type nrm2_slow(const Vector& x)
    {
        typedef typename Vector::value_type T;
        
        Vector temp(x.size());
        
        // temp <- x * x
        thrust::transform(x.begin(), x.end(), temp.begin(), square<T>());

        return std::sqrt( thrust::reduce(temp.begin(), temp.end()) );
    }

    """

INITIALIZE = \
    """
    thrust::host_vector<$InputType>   h_input = unittest::random_integers<bool>($InputSize);
    thrust::device_vector<$InputType> d_input = h_input;

    $InputType h_result = $Method(h_input);
    $InputType d_result = $Method(d_input);

    ASSERT_EQUAL(std::abs(h_result - d_result) / std::abs(h_result + d_result) < 1e-3, true);
    """

TIME = \
    """
    $Method(d_input);
    """

FINALIZE = \
    """
    RECORD_TIME();
    RECORD_THROUGHPUT(double($InputSize));
    RECORD_BANDWIDTH(sizeof($InputType) *  double($InputSize));
    """

InputTypes = ['float', 'double']
InputSizes = [2**24]
Methods    = ['nrm2_fast', 'nrm2_slow']

TestVariables = [('InputType', InputTypes), ('InputSize', InputSizes), ('Method', Methods)]