Function tune_sort

compute/perf/perf_sort.cpp:50–82

template<class T>
void tune_sort(const std::vector<T>& data,
               const size_t trials,
               compute::command_queue& queue)
{
    boost::shared_ptr<compute::detail::parameter_cache>
        params = compute::detail::parameter_cache::get_global_cache(queue.get_device());

    const std::string cache_key =
        std::string("__boost_radix_sort_") + compute::type_name<T>();

    const compute::uint_ tpbs[] = { 32, 64, 128, 256, 512, 1024 };

    double min_time = (std::numeric_limits<double>::max)();
    compute::uint_ best_tpb = 0;

    for(size_t i = 0; i < sizeof(tpbs) / sizeof(*tpbs); i++){
        params->set(cache_key, "tpb", tpbs[i]);

        try {
            const double t = perf_sort(data, trials, queue);
            if(t < min_time){
                best_tpb = tpbs[i];
                min_time = t;
            }
        }
        catch(compute::opencl_error&){
            // invalid work group size for this device, skip
        }
    }

    // store optimal parameters
    params->set(cache_key, "tpb", best_tpb);
}

int main(int argc, char *argv[])
{