Intel® oneAPI DPC++ Library (oneDPL) Release Notes

New Features

• Improved performance of the following algorithms with device policies: adjacent_find, all_of, any_of, copy_if, exclusive_scan, equal, find, find_if, find_end, find_first_of, find_if_not, inclusive_scan, includes, is_heap, is_heap_until, is_partitioned, is_sorted, is_sorted_until, lexicographical_compare, max_element, min_element, minmax_element, mismatch, none_of, partition, partition_copy, reduce, remove, remove_copy, remove_copy_if, remove_if, search, search_n, stable_partition, transform_exclusive_scan, transform_inclusive_scan, unique, and unique_copy.

• Improved performance of sort, stable_sort, and sort_by_key algorithms with device policies when using Merge sort [#fnote1]_.

• Added stable_sort_by_key algorithm in namespace oneapi::dpl.

• Introduced parallel range algorithms in namespace oneapi::dpl::ranges: all_of, any_of, none_of, for_each, find, find_if, find_if_not, adjacent_find, search, search_n, transform, sort, stable_sort, is_sorted, merge, count, count_if, equal, copy, copy_if, min_element, max_element. These algorithms work with C++20 random access ranges and views while also taking an execution policy, similar to other oneDPL algorithms.

• Added support for operators ==, !=, <<, and >> for RNG engines and distributions.

• Introduced experimental support for the Philox RNG engine in namespace oneapi::dpl::experimental.

• Added the <oneapi/dpl/version> header containing oneDPL version macros and new feature testing macros.

Fixed Issues

• Fixed unused variable and unused type warnings.
• Resolved memory leaks when using sort and stable_sort algorithms with the oneTBB backend.
• Fixed a build error for oneapi::dpl::begin and oneapi::dpl::end functions used with the Microsoft* Visual C++ standard library and with C++20.
• Reordered template parameters of the histogram algorithm to match its function parameter order. We recommend removing explicit template parameter specifications and instead adding explicit type conversions for function arguments.
• The gpu::esimd::radix_sort and gpu::esimd::radix_sort_by_key kernel templates now throw std::bad_alloc if they fail to allocate global memory.
• Fixed a potential hang with gpu::esimd::radix_sort and gpu::esimd::radix_sort_by_key kernel templates.
• Corrected the documentation for the sort_by_key algorithm, previously described as stable, which can be unstable for some execution policies. Use stable_sort_by_key for guaranteed stability.
• Fixed an error when calling sort with device execution policies on CUDA devices.
• Allowed passing C++20 random access iterators to oneDPL algorithms.
• Fixed issues with the initialization of SYCL queues in predefined device execution policies, which have been updated to be immutable (const) objects.

Known Issues and Limitations

New in This Release

• histogram may provide incorrect results with device policies when built with the -O0 option.
• Including <oneapi/dpl/dynamic_selection> before <oneapi/dpl/random> may result in compilation errors. Include <oneapi/dpl/random> first as a workaround.
• Incorrect results may occur when using oneapi::dpl::experimental::philox_engine with undefined template parameters and with word_size values other than 64 and 32.
• Some algorithms (e.g., exclusive_scan, inclusive_scan, transform_exclusive_scan, copy_if) may produce incorrect results or trigger synchronous SYCL exceptions when built with -O0 on GPU devices.
• For certain algorithms (transform_inclusive_scan, inclusive_scan), the value type of the input sequence must be convertible to the type of the initial element.
• Some algorithms (copy_if, remove, partition_copy, unique) with device policies may exceed the C++ standard's requirements for predicate applications, applying the predicate or equality operator O(n) times.
• Some algorithms (adjacent_find, find_if, search, etc.) may cause segmentation faults when used with device execution policies on a CPU device, particularly when built on Linux with Intel® oneAPI DPC++/C++ Compiler 2025.0.0 and -O0 -g compiler options.

Existing Issues

• histogram requires the output value type to be an integral type no larger than 4 bytes when used with an FPGA policy.
• Compilation issues may arise when passing zip iterators to exclusive_scan_by_segment on Windows.
• Algorithms such as sort, stable_sort, and sort_by_key may work incorrectly or cause segmentation faults when used with device policies on a CPU device and built with -O0 -g. To avoid the issue, pass the -fsycl-device-code-split=per_kernel option to the compiler.
• Some scan algorithms (exclusive_scan, reduce_by_segment) may produce incorrect results when compiled with certain OpenMP options on Linux.
• 64-bit reductions with reduce and transform_reduce may produce incorrect results on GPU devices. Workaround: define the ONEDPL_WORKAROUND_FOR_IGPU_64BIT_REDUCTION macro.
• std::tuple and std::pair cannot be used with SYCL buffers for data transfer.
• The oneapi::dpl::experimental::ranges::reverse algorithm is unavailable with the -fno-sycl-unnamed-lambda option.
• STL algorithms (e.g., std::for_each) do not compile with the debug version of the Microsoft* Visual C++ standard library.

News

• oneAPI DPC++ Library Manual Migration Guide to simplify the migration of Thrust* and CUB* APIs from CUDA*.
• radix_sort and radix_sort_by_key kernel templates were moved into oneapi::dpl::experimental::kt::gpu::esimd namespace. The former oneapi::dpl::experimental::kt::esimd namespace is deprecated and will be removed in a future release.
• The for_loop, for_loop_strided, for_loop_n, for_loop_n_strided algorithms in namespace oneapi::dpl::experimental are enforced to fail with device execution policies.

New Features

• Added experimental inclusive_scan kernel template algorithm residing in the oneapi::dpl::experimental::kt::gpu namespace.
• radix_sort and radix_sort_by_key kernel templates are extended with overloads for out-of-place sorting. These overloads preserve the input sequence and sort data into the user provided output sequence.
• Improved performance of the reduce, min_element, max_element, minmax_element, is_partitioned, lexicographical_compare, binary_search, lower_bound, and upper_bound algorithms with device policies.
• sort, stable_sort, sort_by_key algorithms now use Radix sort1 for sorting sycl::half elements compared with std::less or std::greater.

Fixed Issues

• Fixed compilation errors when using reduce, min_element, max_element, minmax_element, is_partitioned, and lexicographical_compare with Intel oneAPI DPC++/C++ compiler 2023.0 and earlier.
• Fixed possible data races in the following algorithms used with device execution policies: remove_if, unique, inplace_merge, stable_partition, partial_sort_copy, rotate.
• Fixed excessive copying of data in std::vector allocated with a USM allocator for standard library implementations which have allocator information in the std::vector::iterator type.
• Fixed an issue where checking std::is_default_constructible for transform_iterator with a functor that is not default-constructible could cause a build error or an incorrect result.
• Fixed handling of sycl device copyable for internal and public oneDPL types.
• Fixed handling of std::reverse_iterator as input to oneDPL algorithms using a device policy.
• Fixed set_intersection to always copy from the first input sequence to the output, where previously some calls would copy from the second input sequence.
• Fixed compilation errors when using oneapi::dpl::zip_iterator with the oneTBB backend and C++20.

Known Issues and Limitations

New in This Release

• histogram algorithm requires the output value type to be an integral type no larger than 4 bytes when used with an FPGA policy.

Existing Issues

See oneDPL Guide for other restrictions and known limitations.

• When compiled with -fsycl-pstl-offload option of Intel oneAPI DPC++/C++ compiler and with libstdc++ version 8 or libc++, oneapi::dpl::execution::par_unseq offloads standard parallel algorithms to the SYCL device similarly to std::execution::par_unseq in accordance with the -fsycl-pstl-offload option value.
• When using the dpl modulefile to initialize the user's environment and compiling with -fsycl-pstl-offload option of Intel® oneAPI DPC++/C++ compiler, a linking issue or program crash may be encountered due to the directory containing libpstloffload.so not being included in the search path. Use the env/vars.sh to configure the working environment to avoid the issue.
• Compilation issues may be encountered when passing zip iterators to exclusive_scan_by_segment on Windows.
• For transform_exclusive_scan and exclusive_scan to run in-place (that is, with the same data used for both input and destination) and with an execution policy of unseq or par_unseq, it is required that the provided input and destination iterators are equality comparable. Furthermore, the equality comparison of the input and destination iterator must evaluate to true. If these conditions are not met, the result of these algorithm calls is undefined.
• sort, stable_sort, sort_by_key, partial_sort_copy algorithms may work incorrectly or cause a segmentation fault when used a DPC++ execution policy for CPU device, and built on Linux with Intel® oneAPI DPC++/C++ Compiler and -O0 -g compiler options. To avoid the issue, pass -fsycl-device-code-split=per_kernel option to the compiler.
• Incorrect results may be produced by exclusive_scan, inclusive_scan, transform_exclusive_scan, transform_inclusive_scan, exclusive_scan_by_segment, inclusive_scan_by_segment, reduce_by_segment with unseq or par_unseq policy when compiled by Intel® oneAPI DPC++/C++ Compiler with -fiopenmp, -fiopenmp-simd, -qopenmp, -qopenmp-simd options on Linux. To avoid the issue, pass -fopenmp or -fopenmp-simd option instead.
• Incorrect results may be produced by reduce, reduce_by_segment, and transform_reduce with 64-bit data types when compiled by Intel® oneAPI DPC++/C++ Compiler versions 2021.3 and newer and executed on GPU devices. For a workaround, define the ONEDPL_WORKAROUND_FOR_IGPU_64BIT_REDUCTION macro to 1 before including oneDPL header files.
• std::tuple, std::pair cannot be used with SYCL buffers to transfer data between host and device.
• std::array cannot be swapped in DPC++ kernels with std::swap function or swap member function in the Microsoft* Visual C++ standard library.
• The oneapi::dpl::experimental::ranges::reverse algorithm is not available with -fno-sycl-unnamed-lambda option.
• STL algorithm functions (such as std::for_each) used in DPC++ kernels do not compile with the debug version of the Microsoft* Visual C++ standard library.

The sorting algorithms in oneDPL use Radix sort for arithmetic data types and sycl::half (since oneDPL 2022.6) compared with std::less or std::greater, otherwise Merge sort.

New Features

• Added new histogram algorithms for generating a histogram from an input sequence into an output sequence representing either equally spaced or user-defined bins.
  These algorithms are currently only available for device execution policies.
• Supported zip_iterator for the transform algorithm.

Fixed Issues

• Fixed handling of permutation_iterator as input to oneDPL algorithms for a variety of source iterator and permutation types which caused issues.
• Fixed zip_iterator to be SYCL device copyable for trivially copyable source iterator types.
• Added a workaround for reduction algorithm failures with 64-bit data types. Define the ONEDPL_WORKAROUND_FOR_IGPU_64BIT_REDUCTION macro to 1 before including oneDPL header files.

Known Issues and Limitations

New in This Release

• Crashes or incorrect results may occur when using oneapi::dpl::reverse_iterator or std::reverse_iterator as input to oneDPL algorithms with device execution policies.

Existing Issues

See oneDPL Guide for other restrictions and known limitations.

• When compiled with the -fsycl-pstl-offload option of the Intel oneAPI DPC++/C++ compiler and with libstdc++ version 8 or libc++, oneapi::dpl::execution::par_unseq offloads standard parallel algorithms to the SYCL device similarly to std::execution::par_unseq in accordance with the -fsycl-pstl-offload option value.
• When using the dpl modulefile to initialize the user's environment and compiling with the -fsycl-pstl-offload option of the Intel® oneAPI DPC++/C++ compiler, a linking issue or program crash may be encountered due to the directory containing libpstloffload.so not being included in the search path. Use env/vars.sh to configure the working environment to avoid the issue.
• Compilation issues may be encountered when passing zip_iterators to exclusive_scan_by_segment on Windows.
• Incorrect results may be produced by set_intersection with a DPC++ execution policy, where elements are copied from the second input range rather than the first input range.
• For transform_exclusive_scan and exclusive_scan to run in-place (with the same data used for both input and destination) and with an execution policy of unseq or par_unseq, it is required that the provided input and destination iterators are equality comparable. Furthermore, the equality comparison of the input and destination iterator must evaluate to true. If these conditions are not met, the result of these algorithm calls is undefined.
• sort, stable_sort, sort_by_key, and partial_sort_copy algorithms may work incorrectly or cause a segmentation fault when used with a DPC++ execution policy for a CPU device, and built on Linux with the Intel® oneAPI DPC++/C++ Compiler and -O0 -g compiler options. To avoid the issue, pass the -fsycl-device-code-split=per_kernel option to the compiler.
• Incorrect results may be produced by exclusive_scan, inclusive_scan, transform_exclusive_scan, transform_inclusive_scan, exclusive_scan_by_segment, inclusive_scan_by_segment, reduce_by_segment with unseq or par_unseq policy when compiled by Intel® oneAPI DPC++/C++ Compiler with -fiopenmp, -fiopenmp-simd, -qopenmp, -qopenmp-simd options on Linux. To avoid the issue, pass the -fopenmp or -fopenmp-simd option instead.
• Incorrect results may be produced by reduce, reduce_by_segment, and transform_reduce with 64-bit data types when compiled by Intel® oneAPI DPC++/C++ Compiler versions 2021.3 and newer and executed on GPU devices. For a workaround, define the ONEDPL_WORKAROUND_FOR_IGPU_64BIT_REDUCTION macro to 1 before including oneDPL header files.
• std::tuple and std::pair cannot be used with SYCL buffers to transfer data between host and device.
• std::array cannot be swapped in DPC++ kernels with std::swap function or swap member function in the Microsoft* Visual C++ standard library.
• The oneapi::dpl::experimental::ranges::reverse algorithm is not available with the -fno-sycl-unnamed-lambda option.
• STL algorithm functions (such as std::for_each) used in DPC++ kernels do not compile with the debug version of the Microsoft* Visual C++ standard library.

New Features

• Added experimental radix_sort and radix_sort_by_key algorithms residing in the oneapi::dpl::experimental::kt::esimd namespace. These algorithms are first in the family of _kernel templates_ that allow configuring a variety of parameter including the number of elements to process by a work item, and the size of a workgroup. The algorithms only work with Intel® Data Center GPU Max Series.

• Added new transform_if algorithm for applying a transform function conditionally based on a predicate, with overloads provided for one and two input sequences that use correspondingly unary and binary operations and predicates.

• Optimizations used with Intel® oneAPI DPC++/C++ Compiler are expanded to the open source oneAPI DPC++ compiler.

Known Issues and Limitations

New in This Release

• esimd::radix_sort and esimd::radix_sort_by_key kernel templates fail to compile when a program is built with -g, -O0, -O1 compiler options.

• esimd::radix_sort_by_key kernel template produces wrong results with the following combinations of kernel_param and types of keys and values:
  sizeof(key_type) + sizeof(val_type) == 12, kernel_param::workgroup_size == 64, and kernel_param::data_per_workitem == 96
  sizeof(key_type) + sizeof(val_type) == 16, kernel_param::workgroup_size == 64, and kernel_param::data_per_workitem == 64

New Features

• Added an experimental feature to dynamically select an execution context, e.g., a SYCL queue. The feature provides selection functions such as such as select, submit and submit_and_wait,

  and several selection policies: fixed_resource_policy, round_robin_policy, dynamic_load_policy, and auto_tune_policy.

• unseq and par_unseq policies now enable vectorization for Intel oneAPI DPC++/C++ Compiler.

• Added support for passing zip iterators as segment value data in reduce_by_segment, exclusive_scan_by_segment, and inclusive_scan_by_segment.

• Improved performance of merge, sort, stable_sort, and sort_by_key, reduce, min_element, max_element, minmax_element, is_partitioned, and lexicographical_compare algorithms with DPC++ execution policies.

Fixed Issues

• Fixed the reduce_async function to not ignore the provided binary operation.

Known Issues and Limitations

New in This Release

• When compiled with -fsycl-pstl-offload option of Intel oneAPI DPC++/C++ compiler and with libstdc++ version 8 or libc++, oneapi::dpl::execution::par_unseq offloads standard parallel algorithms to the SYCL device similarly to std::execution::par_unseq in accordance with the -fsycl-pstl-offload option value.

• When using the dpl modulefile to initialize the user's environment and compiling with -fsycl-pstl-offload option of Intel oneAPI DPC++/C++ compiler, a linking issue or program crash may be encountered due to the directory containing libpstloffload.so not being included in the search path. Users need to use the env/vars.sh to configure the working environment to avoid the issue.

• Compilation issues may be encountered when passing zip iterators to exclusive_scan_by_segment on Windows.

• Incorrect results may be produced by set_intersection with a DPC++ execution policy, where elements are copied from the second input range rather than the first input range.

• For transform_exclusive_scan and exclusive_scan to run in-place (that is, with the same data used for both input and destination) and with an execution policy of unseq or par_unseq, it is required that the provided input and destination iterators are equality comparable. Furthermore, the equality comparison of the input and destination iterator must evaluate to true. If these conditions are not met, the result of these algorithm calls is undefined.

• sort, stable_sort, stable_sort_by_key, partial_sort_copy algorithms may work incorrectly or cause a segmentation fault when used a DPC++ execution policy for CPU device, and built on Linux with Intel® oneAPI DPC++/C++ Compiler and -O0 -g compiler options. To avoid the issue, pass -fsycl-device-code-split=per_kernel option to the compiler.

• Incorrect results may be produced by exclusive_scan,inclusive_scan, transform_exclusive_scan, transform_inclusive_scan, exclusive_scan_by_segment, inclusive_scan_by_segment, reduce_by_segment with unseq or par_unseq policy when compiled by Intel® oneAPI DPC++/C++ Compiler with -fiopenmp, -fiopenmp-simd, -qopenmp, -qopenmp-simd options on Linux. To avoid the issue, pass -fopenmp or -fopenmp-simd option instead.

• Incorrect results may be produced by reduce and transform_reduce with 64-bit types and std::multiplies, sycl::multiplies operations when compiled by Intel® C++ Compiler 2021.3 and newer and executed on GPU devices.

Existing Issues

See oneDPL Guide for other restrictions and known limitations. 

• std::tuple, std::pair cannot be used with SYCL buffers to transfer data between host and device.

• std::array cannot be swapped in DPC++ kernels with std::swap function or swap member function in the Microsoft* Visual C++ standard library.

• The oneapi::dpl::experimental::ranges::reverse algorithm is not available with -fno-sycl-unnamed-lambda option.

• STL algorithm functions (such as std::for_each) used in DPC++ kernels do not compile with the debug version of the Microsoft* Visual C++ standard library.

New Features

• Added sort_by_key algorithm for key-value sorting.
• Improved performance of the reduce, min_element, max_element, minmax_element, is_partitioned, and lexicographical_compare algorithms with DPC++ execution policies.
• Improved performance of the reduce_by_segment, inclusive_scan_by_segment, and exclusive_scan_by_segment algorithms for binary operators with known identities when using DPC++ execution policies.
• Improved sort algorithm performance for the arithmetic data types with std::less or std::greater comparison operator and DPC++ execution policies.
• Added value_type to all views in oneapi::dpl::__ranges. 
• Extended oneapi::dpl::experimental::ranges::sort to support projections applied to the range elements prior to comparison.   

Fixed Issues

• The minimally required CMake version is raised to 3.11 on Linux and 3.20 on Windows.
• Added new CMake package oneDPLIntelLLVMConfig.cmake to resolve issues using CMake 3.20+ on Windows for icx and icx-cl.
• Fixed an error in the sort and stable_sort algorithms when performing a descending sort on signed numeric types with negative values.
• Fixed an error in reduce_by_segment algorithm when a non-commutative predicate is used.
• Fixed an error in sort and stable_sort algorithms for integral types wider than 4 bytes.
• Fixed an error for some compilers where OpenMP or SYCL backend was selected by CMake scripts without full compiler support.
• Fixed an error that caused segmentation faults in transform_reduce, minmax_element, and related algorithms when ran on CPU devices. 
• Fixed a compilation error in transform_reduce, minmax_element, and related algorithms on FPGAs.
• Fixed a radix sort issue with 64-bit signed integer types.

Known Issues and Limitations

New in This Release

• Incorrect results may be produced with in-place scans using unseq and par_unseq policies on CPUs with the Intel® C++ Compiler 2021.8.

Existing Issues

See oneDPL Guide for other restrictions and known limitations. 

• std::tuple, std::pair cannot be used with SYCL buffers to transfer data between host and device.

• std::array cannot be swapped in DPC++ kernels with std::swap function or swap member function in the Microsoft* Visual C++ standard library.

• The oneapi::dpl::experimental::ranges::reverse algorithm is not available with -fno-sycl-unnamed-lambda option.

• STL algorithm functions (such as std::for_each) used in DPC++ kernels do not compile with the debug version of the Microsoft* Visual C++ standard library.

New Features

• Added generate, generate_n, transform algorithms to Tested Standard C++ API.
• Improved performance of inclusive_scan, exclusive_scan, reduce,  and max_element algorithms with DPC++ execution policies.    

Fixed Issues

• Added a workaround for the "TBB headers not found" issue occurring with libstdc++ version 9 when oneTBB headers are not present in the environment. The workaround requires inclusion of the oneDPL headers before the libstdc++ headers.
• When possible, oneDPL CMake scripts now enforce C++17 as the minimally required language version.

• Fixed an error in the ``exclusive_scan`` algorithm when the output iterator is equal to the input iterator (in-place scan).

Known Issues and Limitations

New in This Release

• None in this release.

Existing Issues

• std::tuple, std::pair cannot be used with SYCL buffers to transfer data between host and device.
• std::array cannot be swapped in DPC++ kernels with std::swap function or swap member function in the Microsoft* Visual C++ standard library.
• The oneapi::dpl::experimental::ranges::reverse algorithm is not available with -fno-sycl-unnamed-lambda option.
• STL algorithm functions (such as std::for_each) used in DPC++ kernels do not compile with the debug version of the Microsoft* Visual C++ standard library.

NOTE: See oneDPL Guide for other restrictions and known limitations. 

New Features

• ​Added ``<complex>`` header functionality as Tested Standard C++ API.
• Improved performance of ``sort`` and ``stable_sort``  algorithms on GPU devices when using Radix sort. The sorting algorithms in oneDPL use Radix sort for arithmetic data types compared with ``std::less`` or ``std::greater``, otherwise Merge sort.   

Fixed Issues

• Fixed error in ``oneapi::dpl::experimental::ranges::guard_view`` and ``oneapi::dpl::experimental::ranges::zip_view`` when using  ``operator[]`` with a very large index. 
• Fixed permutation_iterator to work with C++ lambda functions for index permutation. 
• Fixed an error in ``oneapi::dpl::experimental::ranges::guard_view`` and ``oneapi::dpl::experimental::ranges::zip_view`` when using ``operator[]`` with an index exceeding the limits of a 32 bit integer type. 
• Fixed errors when data size is 0 in ``upper_bound``, ``lower_bound`` and ``binary_search`` algorithms.

Changes Affecting Backward Compatibility

• Removed support of C++11 and C++14.
• Changed the size and the layout of the ``discard_block_engine`` class template.
  
  For further details, please refer to 2022.0 Changes.

Known Issues and Limitations

New in This Release

• None in this release.

Existing Issues

• std::tuple, std::pair cannot be used with SYCL buffers to transfer data between host and device.
• std::array cannot be swapped in DPC++ kernels with std::swap function or swap member function in the Microsoft* Visual C++ standard library.
• The oneapi::dpl::experimental::ranges::reverse algorithm is not available with -fno-sycl-unnamed-lambda option.
• STL algorithm functions (such as std::for_each) used in DPC++ kernels do not compile with the debug version of the Microsoft* Visual C++ standard library.

NOTE: See oneDPL Guide for other restrictions and known limitations. 

​​​​​