oneAPI DPC++ Library Introduction#

Parallel API can be used with the C++ Standard Execution Policies to enable parallelism on the host.

The oneAPI DPC++ Library (oneDPL) is implemented in accordance with the oneDPL Specification.

To support heterogeneity, oneDPL works with the DPC++ API. More information can be found in the oneAPI Specification.

Before You Begin#

Visit the oneDPL Release Notes page for:

  • Where to Find the Release

  • Overview

  • New Features

  • Fixed Issues

  • Deprecation Notice

  • Known Issues and Limitations

  • Previous Release Notes

Install the Intel® oneAPI Base Toolkit (Base Kit) to use oneDPL.

All oneDPL header files are in the oneapi/dpl directory. To use the oneDPL API, include the corresponding header in your source code with the #include <oneapi/dpl/…> directive. oneDPL introduces the namespace oneapi::dpl for most its classes and functions.

To use tested C++ standard APIs, you need to include the corresponding C++ standard header files and use the std namespace.

System Requirements#

Prerequisites#

C++17 is the minimal supported version of the C++ standard. That means, any use of oneDPL may require a C++17 compiler. While some APIs of the library may accidentally work with earlier versions of the C++ standard, it is no more guaranteed.

To call Parallel API with the C++ standard policies, you need to install the following software:

  • A C++ compiler with support for OpenMP* 4.0 (or higher) SIMD constructs

  • Depending on what parallel backend you want to use install either:

    • oneAPI Threading Building Blocks (oneTBB) or Intel® Threading Building Blocks (Intel® TBB) 2019 and later

    • A C++ compiler with support for OpenMP 4.5 (or higher)

For more information about parallel backends, see Execution Policies

To use Parallel API with the device execution policies, you need to install the following software:

  • A C++ compiler with support for SYCL 2020

Difference with Standard C++ Parallel Algorithms#

  • oneDPL execution policies only result in parallel execution if random access iterators are provided, the execution will remain serial for other iterator types.

  • Function objects passed in to algorithms executed with device policies must provide const-qualified operator(). The SYCL specification states that writing to such an object during a SYCL kernel is undefined behavior.

  • For the following algorithms, par_unseq and unseq policies do not result in vectorized execution: includes, inplace_merge, merge, set_difference, set_intersection, set_symmetric_difference, set_union, stable_partition, unique.

  • The following algorithms require additional O(n) memory space for parallel execution: copy_if, inplace_merge, partial_sort, partial_sort_copy, partition_copy, remove, remove_if, rotate, sort, stable_sort, unique, unique_copy.

Restrictions#

When called with DPC++ execution policies, oneDPL algorithms apply the same restrictions as DPC++ does (see the DPC++ specification and the SYCL specification for details), such as:

  • Adding buffers to a lambda capture list is not allowed for lambdas passed to an algorithm.

  • Passing data types, which are not trivially copyable, is only allowed via USM, but not via buffers or host-allocated containers.

  • The definition of lambda functions used with parallel algorithms should not depend on preprocessor macros that makes it different for the host and the device. Otherwise, the behavior is undefined.

  • When used within SYCL kernels or transferred to/from a device, a container class can only hold objects whose type meets SYCL requirements for use in kernels and for data transfer, respectively.

  • Calling the API that throws exception is not allowed within callable objects passed to an algorithm.

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.

  • 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.

  • For transform_exclusive_scan, transform_inclusive_scan algorithms the result of the unary operation should be convertible to the type of the initial value if one is provided, otherwise it is convertible to the type of values in the processed data sequence: std::iterator_traits<IteratorType>::value_type.

  • exclusive_scan and transform_exclusive_scan algorithms may provide wrong results with vector execution policies when building a program with GCC 10 and using -O0 option.

  • Compiling reduce and transform_reduce algorithms with the Intel DPC++ Compiler, versions 2021 and older, may result in a runtime error. To fix this issue, use an Intel DPC++ Compiler version 2022 or newer.

  • When compiling on Windows, add the option /EHsc to the compilation command to avoid errors with oneDPL’s experimental ranges API that uses exceptions.

  • The use of oneDPL together with the GNU C++ standard library (libstdc++) version 9 or 10 may lead to compilation errors (caused by oneTBB API changes). Using libstdc++ version 9 requires TBB version 2020 for the header file. This may result in compilation errors when using C++17 or C++20 and TBB is not found in the environment, even if its use in oneDPL is switched off. To overcome these issues, include oneDPL header files before the standard C++ header files, or disable parallel algorithms support in the standard library. For more information, please see Intel® oneAPI Threading Building Blocks (oneTBB) Release Notes.

  • The using namespace oneapi; directive in a oneDPL program code may result in compilation errors with some compilers including GCC 7 and earlier. Instead of this directive, explicitly use oneapi::dpl namespace, or create a namespace alias.

  • std::array::at member function cannot be used in kernels because it may throw an exception; use std::array::operator[] instead.

  • Due to specifics of Microsoft* Visual C++, some standard floating-point math functions (including std::ldexp, std::frexp, std::sqrt(std::complex<float>)) require device support for double precision.

  • exclusive_scan, inclusive_scan, exclusive_scan_by_segment, inclusive_scan_by_segment, transform_exclusive_scan, transform_inclusive_scan, when used with C++ standard policies, impose limitations on the initial value type if an initial value is provided, and on the value type of the input iterator if an initial value is not provided. Firstly, it must satisfy the DefaultConstructible requirements. Secondly, a default-constructed instance of that type should act as the identity element for the binary scan function.

  • reduce_by_segment, when used with C++ standard policies, imposes limitations on the value type. Firstly, it must satisfy the DefaultConstructible requirements. Secondly, a default-constructed instance of that type should act as the identity element for the binary reduction function.

  • The initial value type for exclusive_scan, inclusive_scan, exclusive_scan_by_segment, inclusive_scan_by_segment, reduce, reduce_by_segment, transform_reduce, transform_exclusive_scan, transform_inclusive_scan should satisfy the MoveAssignable and the CopyConstructible requirements.

  • For max_element, min_element, minmax_element, partial_sort, partial_sort_copy, sort, stable_sort the dereferenced value type of the provided iterators should satisfy the DefaultConstructible requirements.

  • For remove, remove_if, unique the dereferenced value type of the provided iterators should be MoveConstructible.

  • The algorithms that process uninitialized storage: uninitialized_copy, uninitialized_copy_n, uninitialized_fill, uninitialized_fill_n, uninitialized_fill_n, uninitialized_move, uninitialized_move_n, uninitialized_default_construct, uninitialized_default_construct_n, uninitialized_value_construct, uninitialized_value_construct_n should be called with a device policy when using device data and should be called with a host policy when using host data. Otherwise, the result is undefined.

  • The algorithms that destroy data: destroy and destroy_n should be called with a host policy when using host data that was initialized on the host, and should be called with a device policy when using device data that was initialized on the device. Otherwise, the result is undefined.

Build Your Code with oneDPL#

Follow the steps below to build your code with oneDPL:

  1. To build with the Intel® oneAPI DPC++/C++ Compiler, see the Get Started with the Intel® oneAPI DPC++/C++ Compiler for details.

  2. Set the environment variables for oneDPL and oneTBB.

  3. To avoid naming device policy objects explicitly, add the -fsycl-unnamed-lambda option.

Below is an example of a command line used to compile code that contains oneDPL parallel algorithms on Linux* (depending on the code, parameters within [] could be unnecessary):

dpcpp [-fsycl-unnamed-lambda] test.cpp [-ltbb|-fopenmp] -o test