Function tune_accumulate

compute/perf/perf_accumulate.cpp:48–87

template<class T>
void tune_accumulate(const compute::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_reduce_on_gpu_") + compute::type_name<T>();

    const compute::uint_ tpbs[] = { 4, 8, 16, 32, 64, 128, 256, 512, 1024 };
    const compute::uint_ vpts[] = { 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16 };

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

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

            try {
                const double t = perf_accumulate(data, trials, queue);
                if(t < min_time){
                    best_tpb = tpbs[i];
                    best_vpt = vpts[j];
                    min_time = t;
                }
            }
            catch(compute::opencl_error&){
                // invalid parameters for this device, skip
            }
        }
    }

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

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