FM / detail

/usr/include/boost/compute/algorithm/detail/

/usr/include/boost/compute/algorithm/detail

Name	Size	Mode	Actions
balanced_path.hpp	6267	0644	edit dl rm
binary_find.hpp	4626	0644	edit dl rm
compact.hpp	2188	0644	edit dl rm
copy_on_device.hpp	7382	0644	edit dl rm
copy_to_device.hpp	6731	0644	edit dl rm
copy_to_host.hpp	6658	0644	edit dl rm
count_if_with_ballot.hpp	2605	0644	edit dl rm
count_if_with_reduce.hpp	2648	0644	edit dl rm
count_if_with_threads.hpp	4241	0644	edit dl rm
find_extrema.hpp	2441	0644	edit dl rm
find_extrema_on_cpu.hpp	5333	0644	edit dl rm
find_extrema_with_atomics.hpp	4251	0644	edit dl rm
find_extrema_with_reduce.hpp	18672	0644	edit dl rm
find_if_with_atomics.hpp	8509	0644	edit dl rm
inplace_reduce.hpp	4839	0644	edit dl rm
insertion_sort.hpp	6030	0644	edit dl rm
merge_path.hpp	3874	0644	edit dl rm
merge_sort_on_cpu.hpp	14654	0644	edit dl rm
merge_sort_on_gpu.hpp	22093	0644	edit dl rm
merge_with_merge_path.hpp	7339	0644	edit dl rm
radix_sort.hpp	15502	0644	edit dl rm
random_fill.hpp	1880	0644	edit dl rm
reduce_by_key.hpp	5243	0644	edit dl rm
reduce_by_key_with_scan.hpp	23735	0644	edit dl rm
reduce_on_cpu.hpp	3949	0644	edit dl rm
reduce_on_gpu.hpp	10391	0644	edit dl rm
scan.hpp	1545	0644	edit dl rm
scan_on_cpu.hpp	7056	0644	edit dl rm
scan_on_gpu.hpp	11320	0644	edit dl rm
search_all.hpp	2579	0644	edit dl rm
serial_accumulate.hpp	1990	0644	edit dl rm
serial_count_if.hpp	2243	0644	edit dl rm
serial_find_extrema.hpp	3151	0644	edit dl rm
serial_merge.hpp	3504	0644	edit dl rm
serial_reduce.hpp	2145	0644	edit dl rm
serial_reduce_by_key.hpp	4240	0644	edit dl rm
serial_scan.hpp	3184	0644	edit dl rm

Edit: /usr/include/boost/compute/algorithm/detail/count_if_with_ballot.hpp (2605B)

//---------------------------------------------------------------------------//
// Copyright (c) 2013 Kyle Lutz <kyle.r.lutz@gmail.com>
//
// Distributed under the Boost Software License, Version 1.0
// See accompanying file LICENSE_1_0.txt or copy at
// http://www.boost.org/LICENSE_1_0.txt
//
// See http://boostorg.github.com/compute for more information.
//---------------------------------------------------------------------------//

#ifndef BOOST_COMPUTE_ALGORITHM_DETAIL_COUNT_IF_WITH_BALLOT_HPP
#define BOOST_COMPUTE_ALGORITHM_DETAIL_COUNT_IF_WITH_BALLOT_HPP

#include <boost/compute/context.hpp>
#include <boost/compute/command_queue.hpp>
#include <boost/compute/container/vector.hpp>
#include <boost/compute/algorithm/reduce.hpp>
#include <boost/compute/functional/detail/nvidia_ballot.hpp>
#include <boost/compute/functional/detail/nvidia_popcount.hpp>
#include <boost/compute/detail/meta_kernel.hpp>

namespace boost {
namespace compute {
namespace detail {

template<class InputIterator, class Predicate>
inline size_t count_if_with_ballot(InputIterator first,
                                   InputIterator last,
                                   Predicate predicate,
                                   command_queue &queue)
{
    size_t count = iterator_range_size(first, last);
    size_t block_size = 32;
    size_t block_count = count / block_size;
    if(block_count * block_size != count){
        block_count++;
    }

const ::boost::compute::context &context = queue.get_context();

::boost::compute::vector<uint_> counts(block_count, context);

::boost::compute::detail::nvidia_popcount<uint_> popc;
    ::boost::compute::detail::nvidia_ballot<uint_> ballot;

meta_kernel k("count_if_with_ballot");
    k <<
        "const uint gid = get_global_id(0);\n" <<

"bool value = false;\n" <<
        "if(gid < count)\n" <<
        "    value = " << predicate(first[k.var<const uint_>("gid")]) << ";\n" <<

"uint bits = " << ballot(k.var<const uint_>("value")) << ";\n" <<

"if(get_local_id(0) == 0)\n" <<
            counts.begin()[k.var<uint_>("get_group_id(0)") ]
                << " = " << popc(k.var<uint_>("bits")) << ";\n";

k.add_set_arg<const uint_>("count", count);

k.exec_1d(queue, 0, block_size * block_count, block_size);

uint_ result;
    ::boost::compute::reduce(
        counts.begin(),
        counts.end(),
        &result,
        queue
    );
    return result;
}

} // end detail namespace
} // end compute namespace
} // end boost namespace

#endif // BOOST_COMPUTE_ALGORITHM_DETAIL_COUNT_IF_WITH_BALLOT_HPP

//---------------------------------------------------------------------------// // Copyright (c) 2013 Kyle Lutz // // Distributed under the Boost Software License, Version 1.0 // See accompanying file LICENSE_1_0.txt or copy at // http://www.boost.org/LICENSE_1_0.txt // // See http://boostorg.github.com/compute for more information. //---------------------------------------------------------------------------// #ifndef BOOST_COMPUTE_ALGORITHM_DETAIL_COUNT_IF_WITH_BALLOT_HPP #define BOOST_COMPUTE_ALGORITHM_DETAIL_COUNT_IF_WITH_BALLOT_HPP #include #include #include #include #include #include #include namespace boost { namespace compute { namespace detail { template inline size_t count_if_with_ballot(InputIterator first, InputIterator last, Predicate predicate, command_queue &queue) { size_t count = iterator_range_size(first, last); size_t block_size = 32; size_t block_count = count / block_size; if(block_count * block_size != count){ block_count++; } const ::boost::compute::context &context = queue.get_context(); ::boost::compute::vector counts(block_count, context); ::boost::compute::detail::nvidia_popcount popc; ::boost::compute::detail::nvidia_ballot ballot; meta_kernel k("count_if_with_ballot"); k << "const uint gid = get_global_id(0);\n" << "bool value = false;\n" << "if(gid < count)\n" << " value = " << predicate(first[k.var("gid")]) << ";\n" << "uint bits = " << ballot(k.var("value")) << ";\n" << "if(get_local_id(0) == 0)\n" << counts.begin()[k.var("get_group_id(0)") ] << " = " << popc(k.var("bits")) << ";\n"; k.add_set_arg("count", count); k.exec_1d(queue, 0, block_size * block_count, block_size); uint_ result; ::boost::compute::reduce( counts.begin(), counts.end(), &result, queue ); return result; } } // end detail namespace } // end compute namespace } // end boost namespace #endif // BOOST_COMPUTE_ALGORITHM_DETAIL_COUNT_IF_WITH_BALLOT_HPP