//
// Copyright (C) 2009-2021 Intel Corporation
//
// SPDX-License-Identifier: MIT
//
//

module radix_sort;

kernel_module radix_kernels ("morton_radix_sort.cl") 
{
    links lsc_intrinsics;
    kernel opencl_build_morton_kernel_sort_bin_items              < kernelFunction="sort_morton_codes_bin_items">;
    kernel opencl_build_morton_kernel_sort_reduce_bins            < kernelFunction="sort_morton_codes_reduce_bins">;
    kernel opencl_build_morton_kernel_sort_scatter_items          < kernelFunction="sort_morton_codes_scatter_items">;

    kernel opencl_build_morton_codes_sort_merged                  < kernelFunction="sort_morton_codes_merged">;

    kernel opencl_build_morton_kernel_sort_reduce_bins_wide_partial_sum   < kernelFunction="sort_morton_codes_reduce_bins_wide_partial_sum">;
    kernel opencl_build_morton_kernel_sort_reduce_bins_wide_add_reduce    < kernelFunction="sort_morton_codes_reduce_bins_wide_add_reduce">;
}

metakernel sort(
    qword build_globals,
    dword shift,
    qword global_histogram,
    qword input0,
    qword input1,
    dword input0_offset,
    dword input1_offset,
    dword iteration,
    dword threads)
{
    dispatch opencl_build_morton_kernel_sort_bin_items (threads, 1, 1) args(
        build_globals,
        shift,
        global_histogram,
        input0,
        input1,
        input0_offset,
        input1_offset,
        iteration);

    control(wait_idle);

    dispatch opencl_build_morton_kernel_sort_reduce_bins (1, 1, 1) args(
        threads,
        global_histogram);

    control(wait_idle);

    dispatch opencl_build_morton_kernel_sort_scatter_items (threads, 1, 1) args(
        build_globals,
        shift,
        global_histogram,
        input0,
        input1,
        input0_offset,
        input1_offset,
        iteration);

        control(wait_idle);

}

metakernel sort_bin_items(
    qword build_globals,
    qword global_histogram,
    qword wg_flags,
    qword input0,
    dword iteration,
    dword threads,
    dword update_wg_flags
    )
{
    dispatch opencl_build_morton_kernel_sort_bin_items (threads, 1, 1) args(
        build_globals,
        global_histogram,
        wg_flags,
        input0,
        iteration,
        threads,
        update_wg_flags
    );
}

metakernel sort_reduce_bins(
    qword build_globals,
    qword global_histogram,
    dword threads,
    dword iteration)
{
    dispatch opencl_build_morton_kernel_sort_reduce_bins (1, 1, 1) args(
        build_globals,
        threads,
        global_histogram,
        iteration);
}

metakernel sort_scatter_items(
    qword build_globals,
    qword global_histogram,
    qword input0,
    qword input1,
    dword iteration,
    dword threads,
    dword update_morton_sort_in_flight )
{
    dispatch opencl_build_morton_kernel_sort_scatter_items( threads, 1, 1 ) args(
        build_globals,
        global_histogram,
        input0,
        input1,
        iteration,
        threads,
        update_morton_sort_in_flight
    );
}

metakernel sort_bin_items_merged(
    qword build_globals,
    qword global_histogram,
    qword input0,
    dword iteration,
    dword threads)
{
    dispatch opencl_build_morton_codes_sort_merged (threads, 1, 1) args(
        build_globals,
        global_histogram,
        input0,
        iteration,
        threads
    );
}

metakernel sort_reduce_bins_wide(
    qword build_globals,
    qword global_histogram,
    qword global_histogram_tmp,
    qword wg_flags,
    dword threads,
    dword threads_groups,
    dword iteration)
{
    dispatch opencl_build_morton_kernel_sort_reduce_bins_wide_partial_sum(threads_groups, 1, 1) args(
        build_globals,
        threads,
        threads_groups,
        global_histogram,
        global_histogram_tmp,
        wg_flags,
        iteration);

    control(wait_idle);

    dispatch opencl_build_morton_kernel_sort_reduce_bins_wide_add_reduce(threads_groups, 1, 1) args(
        build_globals,
        threads,
        threads_groups,
        global_histogram,
        global_histogram_tmp,
        iteration);
}