C++ command_queue::get_device方法代码示例

inline size_t count_if(InputIterator first,
                       InputIterator last,
                       Predicate predicate,
                       command_queue &queue = system::default_queue())
{
    const device &device = queue.get_device();

    size_t input_size = detail::iterator_range_size(first, last);
    if(input_size == 0){
        return 0;
    }

    if(device.type() & device::cpu){
        if(input_size < 1024){
            return detail::serial_count_if(first, last, predicate, queue);
        }
        else {
            return detail::count_if_with_threads(first, last, predicate, queue);
        }
    }
    else {
        if(input_size < 32){
            return detail::serial_count_if(first, last, predicate, queue);
        }
        else {
            return detail::count_if_with_reduce(first, last, predicate, queue);
        }
    }
}

inline InputIterator binary_find(InputIterator first,
                                 InputIterator last,
                                 UnaryPredicate predicate,
                                 command_queue &queue = system::default_queue())
{
    const device &device = queue.get_device();

    boost::shared_ptr<parameter_cache> parameters =
        detail::parameter_cache::get_global_cache(device);

    const std::string cache_key = "__boost_binary_find";

    size_t find_if_limit = 128;
    size_t threads = parameters->get(cache_key, "tpb", 128);
    size_t count = iterator_range_size(first, last);

    InputIterator search_first = first;
    InputIterator search_last = last;

    scalar<uint_> index(queue.get_context());

    // construct and compile binary_find kernel
    binary_find_kernel<InputIterator, UnaryPredicate>
        binary_find_kernel(search_first, search_last, predicate);
    ::boost::compute::kernel kernel = binary_find_kernel.compile(queue.get_context());

    // set buffer for index
    kernel.set_arg(binary_find_kernel.m_index_arg, index.get_buffer());

    while(count > find_if_limit) {
        index.write(static_cast<uint_>(count), queue);

        // set block and run binary_find kernel
        uint_ block = static_cast<uint_>((count - 1)/(threads - 1));
        kernel.set_arg(binary_find_kernel.m_block_arg, block);
        queue.enqueue_1d_range_kernel(kernel, 0, threads, 0);

        size_t i = index.read(queue);

        if(i == count) {
            search_first = search_last - ((count - 1)%(threads - 1));
            break;
        } else {
            search_last = search_first + i;
            search_first = search_last - ((count - 1)/(threads - 1));
        }

        // Make sure that first and last stay within the input range
        search_last = (std::min)(search_last, last);
        search_last = (std::max)(search_last, first);

        search_first = (std::max)(search_first, first);
        search_first = (std::min)(search_first, last);

        count = iterator_range_size(search_first, search_last);
    }

    return find_if(search_first, search_last, predicate, queue);
}

inline void merge_sort_by_key_on_cpu(KeyIterator keys_first,
                                     KeyIterator keys_last,
                                     ValueIterator values_first,
                                     Compare compare,
                                     command_queue &queue)
{
    typedef typename std::iterator_traits<KeyIterator>::value_type key_type;
    typedef typename std::iterator_traits<ValueIterator>::value_type value_type;

    size_t count = iterator_range_size(keys_first, keys_last);
    if(count < 2){
        return;
    }
    // for small input size only insertion sort is performed
    else if(count <= 512){
        block_insertion_sort(keys_first, values_first, compare,
                             count, count, true, queue);
        return;
    }

    const context &context = queue.get_context();
    const device &device = queue.get_device();

    // loading parameters
    std::string cache_key =
        std::string("__boost_merge_sort_by_key_on_cpu_") + type_name<value_type>()
        + "_with_" + type_name<key_type>();
    boost::shared_ptr<parameter_cache> parameters =
        detail::parameter_cache::get_global_cache(device);

    const size_t block_size =
        parameters->get(cache_key, "insertion_sort_by_key_block_size", 64);
    block_insertion_sort(keys_first, values_first, compare,
                         count, block_size, true, queue);

    // temporary buffer for merge results
    vector<value_type> values_temp(count, context);
    vector<key_type> keys_temp(count, context);
    bool result_in_temporary_buffer = false;

    for(size_t i = block_size; i < count; i *= 2){
        result_in_temporary_buffer = !result_in_temporary_buffer;
        if(result_in_temporary_buffer) {
            merge_blocks(keys_first, values_first,
                         keys_temp.begin(), values_temp.begin(),
                         compare, count, i, true, queue);
        } else {
            merge_blocks(keys_temp.begin(), values_temp.begin(),
                         keys_first, values_first,
                         compare, count, i, true, queue);
        }
    }

    if(result_in_temporary_buffer) {
        copy(keys_temp.begin(), keys_temp.end(), keys_first, queue);
        copy(values_temp.begin(), values_temp.end(), values_first, queue);
    }
}

inline void dispatch_sort(Iterator first,
                          Iterator last,
                          Compare compare,
                          command_queue &queue,
                          typename boost::enable_if<
                              is_device_iterator<Iterator>
                          >::type* = 0)
{
    if(queue.get_device().type() & device::gpu) {
        dispatch_gpu_sort(first, last, compare, queue);
        return;
    }
    ::boost::compute::detail::merge_sort_on_cpu(first, last, compare, queue);
}

inline void dispatch_sort_by_key(KeyIterator keys_first,
                                 KeyIterator keys_last,
                                 ValueIterator values_first,
                                 Compare compare,
                                 command_queue &queue)
{
    if(queue.get_device().type() & device::gpu) {
        dispatch_gpu_sort_by_key(keys_first, keys_last, values_first, compare, queue);
        return;
    }
    ::boost::compute::detail::merge_sort_by_key_on_cpu(
        keys_first, keys_last, values_first, compare, queue
    );
}

inline OutputIterator copy_on_device(InputIterator first,
                                     InputIterator last,
                                     OutputIterator result,
                                     command_queue &queue)
{
    const device &device = queue.get_device();

    copy_kernel<InputIterator, OutputIterator> kernel(device);

    kernel.set_range(first, last, result);
    kernel.exec(queue);

    return result + std::distance(first, last);
}

inline future<OutputIterator> copy_on_device_async(InputIterator first,
                                                   InputIterator last,
                                                   OutputIterator result,
                                                   command_queue &queue)
{
    const device &device = queue.get_device();

    copy_kernel<InputIterator, OutputIterator> kernel(device);

    kernel.set_range(first, last, result);
    event event_ = kernel.exec(queue);

    return make_future(result + std::distance(first, last), event_);
}

inline OutputIterator scan(InputIterator first,
                           InputIterator last,
                           OutputIterator result,
                           bool exclusive,
                           command_queue &queue)
{
    const device &device = queue.get_device();

    if(device.type() & device::cpu){
        return scan_on_cpu(first, last, result, exclusive, queue);
    }
    else {
        return scan_on_gpu(first, last, result, exclusive, queue);
    }
}

inline OutputIterator merge(InputIterator1 first1,
                            InputIterator1 last1,
                            InputIterator2 first2,
                            InputIterator2 last2,
                            OutputIterator result,
                            Compare comp,
                            command_queue &queue = system::default_queue())
{
    BOOST_STATIC_ASSERT(is_device_iterator<InputIterator1>::value);
    BOOST_STATIC_ASSERT(is_device_iterator<InputIterator2>::value);
    BOOST_STATIC_ASSERT(is_device_iterator<OutputIterator>::value);
    typedef typename std::iterator_traits<InputIterator1>::value_type input1_type;
    typedef typename std::iterator_traits<InputIterator2>::value_type input2_type;
    typedef typename std::iterator_traits<OutputIterator>::value_type output_type;

    const device &device = queue.get_device();

    std::string cache_key =
        std::string("__boost_merge_") + type_name<input1_type>() + "_"
        + type_name<input2_type>() + "_" + type_name<output_type>();
    boost::shared_ptr<detail::parameter_cache> parameters =
        detail::parameter_cache::get_global_cache(device);

    // default serial merge threshold depends on device type
    size_t default_serial_merge_threshold = 32768;
    if(device.type() & device::gpu) {
        default_serial_merge_threshold = 2048;
    }

    // loading serial merge threshold parameter
    const size_t serial_merge_threshold =
        parameters->get(cache_key, "serial_merge_threshold",
                        static_cast<uint_>(default_serial_merge_threshold));

    // choosing merge algorithm
    const size_t total_count =
        detail::iterator_range_size(first1, last1)
        + detail::iterator_range_size(first2, last2);
    // for small inputs serial merge turns out to outperform
    // merge with merge path algorithm
    if(total_count <= serial_merge_threshold){
       return detail::serial_merge(first1, last1, first2, last2, result, comp, queue);
    }
    return detail::merge_with_merge_path(first1, last1, first2, last2, result, comp, queue);
}

inline void generic_reduce(InputIterator first,
                           InputIterator last,
                           OutputIterator result,
                           BinaryFunction function,
                           command_queue &queue)
{
    typedef typename
        std::iterator_traits<InputIterator>::value_type
        input_type;
    typedef typename
        boost::compute::result_of<BinaryFunction(input_type, input_type)>::type
        result_type;

    const device &device = queue.get_device();
    const context &context = queue.get_context();

    size_t count = detail::iterator_range_size(first, last);

    if(device.type() & device::cpu){
        boost::compute::vector<result_type> value(1, context);
        detail::serial_reduce(first, last, value.begin(), function, queue);
        boost::compute::copy_n(value.begin(), 1, result, queue);
    }
    else {
        size_t block_size = 256;

        // first pass
        vector<result_type> results = detail::block_reduce(first,
                                                           count,
                                                           block_size,
                                                           function,
                                                           queue);

        if(results.size() > 1){
            detail::inplace_reduce(results.begin(),
                                   results.end(),
                                   function,
                                   queue);
        }

        boost::compute::copy_n(results.begin(), 1, result, queue);
    }
}

bool find_extrema_with_reduce_requirements_met(InputIterator first,
                                               InputIterator last,
                                               command_queue &queue)
{
    typedef typename std::iterator_traits<InputIterator>::value_type input_type;

    const device &device = queue.get_device();

    // device must have dedicated local memory storage
    // otherwise reduction would be highly inefficient
    if(device.get_info<CL_DEVICE_LOCAL_MEM_TYPE>() != CL_LOCAL)
    {
        return false;
    }

    const size_t max_work_group_size = device.get_info<CL_DEVICE_MAX_WORK_GROUP_SIZE>();
    // local memory size in bytes (per compute unit)
    const size_t local_mem_size = device.get_info<CL_DEVICE_LOCAL_MEM_SIZE>();

    std::string cache_key = std::string("__boost_find_extrema_reduce_")
        + type_name<input_type>();
    // load parameters
    boost::shared_ptr<parameter_cache> parameters =
        detail::parameter_cache::get_global_cache(device);

    // Get preferred work group size
    size_t work_group_size = parameters->get(cache_key, "wgsize", 256);

    work_group_size = (std::min)(max_work_group_size, work_group_size);

    // local memory size needed to perform parallel reduction
    size_t required_local_mem_size = 0;
    // indices size
    required_local_mem_size += sizeof(uint_) * work_group_size;
    // values size
    required_local_mem_size += sizeof(input_type) * work_group_size;

    // at least 4 work groups per compute unit otherwise reduction
    // would be highly inefficient
    return ((required_local_mem_size * 4) <= local_mem_size);
}

inline void dispatch_reduce(InputIterator first,
                            InputIterator last,
                            OutputIterator result,
                            const plus<T> &function,
                            command_queue &queue)
{
    const context &context = queue.get_context();
    const device &device = queue.get_device();

    // reduce to temporary buffer on device
    array<T, 1> value(context);
    if(device.type() & device::cpu){
        detail::reduce_on_cpu(first, last, value.begin(), function, queue);
    }
    else {
        reduce_on_gpu(first, last, value.begin(), function, queue);
    }

    // copy to result iterator
    copy_n(value.begin(), 1, result, queue);
}

inline InputIterator binary_find(InputIterator first,
                                 InputIterator last,
                                 UnaryPredicate predicate,
                                 command_queue &queue = system::default_queue())
{
    const device &device = queue.get_device();

    boost::shared_ptr<parameter_cache> parameters =
        detail::parameter_cache::get_global_cache(device);

    const std::string cache_key = "__boost_binary_find";

    size_t find_if_limit = 128;
    size_t threads = parameters->get(cache_key, "tpb", 128);
    size_t count = iterator_range_size(first, last);

    while(count > find_if_limit) {

        scalar<uint_> index(queue.get_context());
        index.write(static_cast<uint_>(count), queue);

        binary_find_kernel kernel(threads);
        kernel.set_range(first, last, predicate);
        kernel.exec(queue, index);

        size_t i = index.read(queue);

        if(i == count) {
            first = last - count%threads;
            break;
        } else {
            last = first + i;
            first = last - count/threads;
        }

        count = iterator_range_size(first, last);
    }

    return find_if(first, last, predicate, queue);
}

/// Checks if the compute device is CPU.
inline bool is_cpu(const command_queue &q) {
    return q.get_device().get_info<cl_device_type>(CL_DEVICE_TYPE) & CL_DEVICE_TYPE_CPU;
}

inline void merge_sort_on_cpu(Iterator first,
                              Iterator last,
                              Compare compare,
                              command_queue &queue)
{
    typedef typename std::iterator_traits<Iterator>::value_type value_type;

    size_t count = iterator_range_size(first, last);
    if(count < 2){
        return;
    }
    // for small input size only insertion sort is performed
    else if(count <= 512){
        block_insertion_sort(first, compare, count, count, queue);
        return;
    }

    const context &context = queue.get_context();
    const device &device = queue.get_device();

    // loading parameters
    std::string cache_key =
        std::string("__boost_merge_sort_on_cpu_") + type_name<value_type>();
    boost::shared_ptr<parameter_cache> parameters =
        detail::parameter_cache::get_global_cache(device);

    // When there is merge_with_path_blocks_no_threshold or less blocks left to
    // merge AND input size is merge_with_merge_path_input_size_threshold or more
    // merge_with_merge_path() algorithm is used to merge sorted blocks;
    // otherwise merge_blocks() is used.
    const size_t merge_with_path_blocks_no_threshold =
        parameters->get(cache_key, "merge_with_merge_path_blocks_no_threshold", 8);
    const size_t merge_with_path_input_size_threshold =
        parameters->get(cache_key, "merge_with_merge_path_input_size_threshold", 2097152);

    const size_t block_size =
        parameters->get(cache_key, "insertion_sort_block_size", 64);
    block_insertion_sort(first, compare, count, block_size, queue);

    // temporary buffer for merge result
    vector<value_type> temp(count, context);
    bool result_in_temporary_buffer = false;

    for(size_t i = block_size; i < count; i *= 2){
        result_in_temporary_buffer = !result_in_temporary_buffer;
        if(result_in_temporary_buffer) {
            dispatch_merge_blocks(first, temp.begin(), compare, count, i,
                                  merge_with_path_input_size_threshold,
                                  merge_with_path_blocks_no_threshold,
                                  queue);
        } else {
            dispatch_merge_blocks(temp.begin(), first, compare, count, i,
                                  merge_with_path_input_size_threshold,
                                  merge_with_path_blocks_no_threshold,
                                  queue);
        }
    }

    if(result_in_temporary_buffer) {
        copy(temp.begin(), temp.end(), first, queue);
    }
}