Radix Sort

radix_sort Function Templates

The radix_sort function sorts data using the radix sort algorithm. The sorting is stable, preserving the relative order of equal elements. Both in-place and out-of-place overloads are provided. Out-of-place overloads do not alter the input sequence.

The functions implement a Onesweep* 1 algorithm variant.

Note

radix_sort is currently available for Intel® Arc™ B-Series and Intel® Data Center GPU Max Series. The Intel® oneAPI DPC++/C++ Compiler 2025.1.0 or greater is required, and the Unified Runtime adapter over Level Zero must be used (which is the default adapter for Intel GPUs). For more information, please refer to DPC++ Device Selection.

A synopsis of the radix_sort function is provided below:

// defined in <oneapi/dpl/experimental/kernel_templates>

namespace oneapi::dpl::experimental::kt::gpu {

// Sort in-place
template <bool IsAscending = true, std::uint8_t RadixBits = 8,
          typename KernelParam, typename Iterator>
sycl::event
radix_sort (sycl::queue q, Iterator first, Iterator last,
            KernelParam param = {}); // (1)

template <bool IsAscending = true, std::uint8_t RadixBits = 8,
          typename KernelParam, typename Range>
sycl::event
radix_sort (sycl::queue q, Range&& r, KernelParam param = {}); // (2)

// Sort out-of-place
template <bool IsAscending = true, std::uint8_t RadixBits = 8,
          typename KernelParam, typename Iterator1,
          typename Iterator2>
sycl::event
radix_sort (sycl::queue q, Iterator1 first, Iterator1 last,
            Iterator2 first_out, KernelParam param = {}); // (3)

template <bool IsAscending = true, std::uint8_t RadixBits = 8,
          typename KernelParam, typename Range1, typename Range2>
sycl::event
radix_sort (sycl::queue q, Range1&& r, Range2&& r_out,
            KernelParam param = {}); // (4)
}

Template Parameters

Name	Description
bool IsAscending	The sort order. Ascending: true; Descending: false.
std::uint8_t RadixBits	The number of bits to sort for each radix sort algorithm pass.

Parameters

Name	Description
q	The SYCL* queue where kernels are submitted.
first, last (1), r (2), first, last, first_out (3), r, r_out (4).	The sequences to apply the algorithm to. Supported sequence types: USM pointers (1,3), oneapi::dpl::begin and oneapi::dpl::end (1,3). sycl::buffer (2,4), oneapi::dpl::experimental::ranges::views::all (2,4), oneapi::dpl::experimental::ranges::views::subrange (2,4).
param	A kernel_param object.

Type Requirements:

• The element type of sequence(s) to sort must be a C++ integral or floating-point type other than bool with a width of up to 64 bits.

Note

Current limitations:

• Number of elements to sort must not exceed 2³⁰.

• RadixBits can only be 8.

• param.workgroup_size can be 512 or 1024.

Return Value

A sycl::event object representing the status of the algorithm execution.

Usage Examples

In-Place Example

// possible build and run commands:
//    icpx -fsycl radix_sort.cpp -o radix_sort -I /path/to/oneDPL/include && ./radix_sort

#include <cstdint>
#include <iostream>
#include <sycl/sycl.hpp>

#include <oneapi/dpl/experimental/kernel_templates>

namespace kt = oneapi::dpl::experimental::kt;

int main()
{
   std::size_t n = 6;
   sycl::queue q{sycl::gpu_selector_v};
   std::uint32_t* keys = sycl::malloc_shared<std::uint32_t>(n, q);

   // initialize
   keys[0] = 3, keys[1] = 2, keys[2] = 1, keys[3] = 5, keys[4] = 3, keys[5] = 3;

   // sort
   auto e = kt::gpu::radix_sort<false, 8>(q, keys, keys + n, kt::kernel_param<10, 512>{}); // (1)
   e.wait();

   // print
   for(std::size_t i = 0; i < n; ++i)
      std::cout << keys[i] << ' ';
   std::cout << '\n';

   sycl::free(keys, q);
   return 0;
}

Output:

5 3 3 3 2 1

Out-of-Place Example

// possible build and run commands:
//    icpx -fsycl radix_sort.cpp -o radix_sort -I /path/to/oneDPL/include && ./radix_sort

#include <cstdint>
#include <iostream>
#include <sycl/sycl.hpp>

#include <oneapi/dpl/experimental/kernel_templates>

namespace kt = oneapi::dpl::experimental::kt;

int main()
{
   std::size_t n = 6;
   sycl::queue q{sycl::gpu_selector_v};
   std::uint32_t* keys = sycl::malloc_shared<std::uint32_t>(n, q);
   std::uint32_t* keys_out = sycl::malloc_shared<std::uint32_t>(n, q);

   // initialize
   keys[0] = 3, keys[1] = 2, keys[2] = 1, keys[3] = 5, keys[4] = 3, keys[5] = 3;

   // sort
   auto e = kt::gpu::radix_sort<false, 8>(q, keys, keys + n, keys_out, kt::kernel_param<10, 512>{}); // (3)
   e.wait();

   // print
   for(std::size_t i = 0; i < n; ++i)
      std::cout << keys[i] << ' ';
   std::cout << '\n';
   for(std::size_t i = 0; i < n; ++i)
      std::cout << keys_out[i] << ' ';
   std::cout << '\n';

   sycl::free(keys, q);
   sycl::free(keys_out, q);
   return 0;
}

Output:

3 2 1 5 3 3
5 3 3 3 2 1

Memory Requirements

The algorithm uses global and local device memory (see SYCL 2020 Specification) for intermediate data storage. For the algorithm to operate correctly, there must be enough memory on the device. If there is not enough global device memory, a std::bad_alloc exception is thrown. The behavior is undefined if there is not enough local memory. The amount of memory that is required depends on input data and configuration parameters, as described below.

Global Memory Requirements

Global memory is used for copying the input sequence(s) and storing internal data such as radix value counters. The amount used depends on many parameters; below is an upper bound approximation:

N_keys + C * N_keys

where the sequence with keys takes N_keys space, and the additional space is C * N_keys.

The value of C depends on param.data_per_workitem, param.workgroup_size and RadixBits. For param.data_per_workitem set to 10, param.workgroup_size to 512, and RadixBits to 8, C is typically less than 1. Doubling either param.data_per_workitem or param.workgroup_size leads to a halving of C.

Local Memory Requirements

Local memory is used for reordering keys within a work-group, and for storing internal data such as radix value counters. The amount used depends on many parameters; below is an upper bound approximation:

N_{keys_per_workgroup} + C

where N_{keys_per_workgroup} is the amount of memory to store keys. C is some additional space for storing internal data.

N_{keys_per_workgroup} is equal to sizeof(key_type) * param.data_per_workitem * param.workgroup_size, C does not exceed 4KB.

Recommended Settings for Best Performance

The general advice is to choose kernel parameters based on performance measurements and profiling information. The initial configuration may be selected according to these high-level guidelines:

• When the number of elements to sort N is small (e.g., less than ~1M), utilizing all available compute cores may improve performance. Experiment with creating enough work chunks to feed all X^e-cores 2 on a GPU: param.data_per_workitem * param.workgroup_size ≈ N / xe_core_count in addition to the next configuration which maximizes data per work-group. The optimal settings may differ between hardware depending on the cost of inter-work group synchronization.

• When the number of elements to sort is large (e.g., more than ~1M), maximizing the number of elements processed by a work-group, which equals to param.data_per_workitem * param.workgroup_size, reduces synchronization overheads between work-groups and usually benefits the overall performance.

The following table provides starting points for param.data_per_workitem and param.workgroup_size for large input sizes. This configuration performs well when sorting 2²⁸ std::uint32_t input elements uniformly distributed over the range [0, UINT32_MAX]:

Sample Parameters for std::uint32_t Sort

Platform	data_per_workitem	workgroup_size
Intel Data Center GPU Max	28	512
Intel Arc B-Series	10	512

When tuning your own parameters, these param.data_per_workitem values may serve as a good initial starting point and can be thought of as an upper-bound of values to test with during experimentation. For smaller inputs, a lower param.data_per_workitem generally performs better.

Tip

• Optimal parameters may differ by the data type and the entropy of the underlying data, so it is important that this is reflected in your tuning experiments.

• Avoid setting too large param.data_per_workitem and param.workgroup_size values by ensuring that Memory requirements are satisfied.

• Maximizing param.data_per_workitem generally improves scalable performance as long as private memory usage does not exceed available register capacity.

• Large performance drops with an increase to param.data_per_workitem are indicative of excessive register spillage. Ahead-of-Time (AOT) Compilation may be used to emit warnings when this occurs.

1

Andy Adinets and Duane Merrill (2022). Onesweep: A Faster Least Significant Digit Radix Sort for GPUs. https://arxiv.org/abs/2206.01784.

2

The X^e-core term is described in the oneAPI GPU Optimization Guide. Check the number of cores in the device specification, such as Intel Arc B580 specification.