########################################################################
# How to create custom kernel dispatchers
########################################################################
A kernel dispatcher is kernel implementation that calls other kernel implementations.
By default, a dispatcher is generated by the build system for every kernel such that:
  * the best aligned implemention is called when all pointer arguments are aligned,
  * and otherwise the best unaligned implementation is called.

The author of a VOLK kernel may create a custom dispatcher,
to be called in place of the automatically generated one.
A custom dispatcher may be useful to handle head and tail cases,
or to implement different alignment and bounds checking logic.

########################################################################
# Code for an example dispatcher w/ tail case
########################################################################
#include <volk/volk_common.h>

#ifdef LV_HAVE_DISPATCHER

static inline void volk_32f_x2_add_32f_dispatcher(float* cVector, const float* aVector, const float* bVector, unsigned int num_points)
{
    const unsigned int num_points_r = num_points%4;
    const unsigned int num_points_x = num_points - num_points_r;

    if (volk_is_aligned(VOLK_OR_PTR(cVector, VOLK_OR_PTR(aVector, bVector))))
    {
        volk_32f_x2_add_32f_a(cVector, aVector, bVector, num_points_x);
    }
    else
    {
        volk_32f_x2_add_32f_u(cVector, aVector, bVector, num_points_x);
    }

    volk_32f_x2_add_32f_g(cVector+num_points_x, aVector+num_points_x, bVector+num_points_x, num_points_r);
}

#endif //LV_HAVE_DISPATCHER

########################################################################
# Code for an example dispatcher w/ tail case and accumulator
########################################################################
#include <volk/volk_common.h>

#ifdef LV_HAVE_DISPATCHER

static inline void volk_32f_x2_dot_prod_32f_dispatcher(float * result, const float * input, const float * taps, unsigned int num_points)
{
    const unsigned int num_points_r = num_points%16;
    const unsigned int num_points_x = num_points - num_points_r;

    if (volk_is_aligned(VOLK_OR_PTR(input, taps)))
    {
        volk_32f_x2_dot_prod_32f_a(result, input, taps, num_points_x);
    }
    else
    {
        volk_32f_x2_dot_prod_32f_u(result, input, taps, num_points_x);
    }

    float result_tail = 0;
    volk_32f_x2_dot_prod_32f_g(&result_tail, input+num_points_x, taps+num_points_x, num_points_r);

    *result += result_tail;
}

#endif //LV_HAVE_DISPATCHER