Alpaka Versions Save

This release features small additions, changes, and fixes, including:

Warp function support for Shuff-Up, Down, and Xor. Named access to vector components via .x(), .y(), .... CMake's native HIP support is used to improve the compatibility with future HIP updates. CMake presets for alpaka backends simplify the integration into your favorite IDE. alpaka-ls helps you to see all available backends in your system. View the full changelog here.

This release features countless small additions, changes and fixes, including:

• A rework of the pitches, extents and offsets APIs.
• Experimental support for std::mdspan.
• Platforms now need to be instantiated.
• Moving the SYCL backend use USM pointers, SYCL2020 and oneAPI.
• Removal of OpenMP 5, OpenACC and Boost.Fiber backends.
• And much more!

View the full changelog here.

This release features multiple new major additions:

• A new (experimental) SYCL back-end. This adds support for Intel oneAPI hardware targets (CPUs, GPUs, FPGAs) as well as AMD/Xilinx FPGAs.
• Support for complex numbers.
• The code base has been migrated to C++17. C++14 is no longer supported.

View the full changelog here.

This release features the new portable Philox-based random number generator. In addition there are many more smaller features and compatibility changes, the most notable being:

• The kernel language now supports memory fences.
• alpaka now has an experimental namespace in which we will try out unstable features. The first experimental feature is an abstraction for memory access called accessor.
• We added support for clang 12, CUDA 11.4, GCC 11, clang as HIP compiler, XCode 12.5.1 and XCode 13.
• clang < 5.0, CUDA < 9.2, GCC < 7.0, nvcc as HIP compiler, Visual Studio < 2019, Ubuntu 16.04, XCode < 11.3.1 and XCode < 12.4 are no longer supported.
• Linking to the CMake alpaka::alpaka target no longer annoys users with alpaka-internal warnings.

View the full changelog here.

This release features the new alpaka intrinsic warp::shfl (and an accordingly updated cheat sheet). Apart from that we mostly focused on maintenance and convenience changes:

• We removed support for 32bit Windows, Visual Studio versions older than 2019 and clang+CUDA for clang versions older than 9.
• We now support clang 11 and CUDA 11.3.
• We now mandate CMake 3.18 (or newer) so we can make use of CMake's native CUDA support.
• A few CMake flags have been renamed.
• The CUDA and HIP back-ends no longer enable -ffast-math by default.

See the full changelog here.

This release fixes various bugs and changes the interface for configuring the OpenMP scheduler. A critical bug in the shared memory implementation for CPU backends is fixed, therefore version 0.6.0 should no longer be used. An overview of all changes can be found in the Changelog.

This release adds two new backends: OpenMP 5 target offload and OpenACC.

We improved useability of alpaka API by flattening the namespace hierarchies and making some renamings. A full list of API changes is avaiable in the changelog. New features include support for warp voting functions, setting schedule for OpenMP2Blocks backend and simplified interfaces for atomic functions.

For cmake-based builds, we no longer automatically enable all available backends. Now a user has to explicitly enable the backends to be used.

The readthedocs documentation was extended with a cheatsheet and support for Doxygen.

This release is adding compatibility to the latest CUDA releases up to 11.2. The HIP backend is now more stable and supports HIP 3.5+. We recommend using the latest HIP version to benefit from its fast improvements.

We fixed many bugs and improved support for Intel C++ compiler.

Compatibility Changes:

• the minimum required C++ version has been raised from C++11 to C++14 #900
• drop support for CUDA 8.0 (does not support c++14)
• drop support for gcc 4.9 (does not support c++14)
• drop support for CMake versions lower than 3.15 (3.11, 3.12, 3.13 and 3.14)
• raise minimum supported boost version from 1.62.0 to 1.65.1 #906
• require HIP version to 3.3.0 #1006
• drop HIP-hcc support #945

Bug Fixes:

• fix CMake error #941
• fix HIP math includes #947
• fix: missing hipRand and rocRand library #948
• fix VS 2017 CUDA builds #953
• fix uninitialized pitch #963
• fix windows CI builds #965
• fix conversion warning in TinyMT #997

New Features:

• add automated gh-pages deployment for branch develop #916
• unify CUDA/HIP backend #928 #904 #950 #980 #981
• add support for Visual Studio 2019 #949
• simplify vector operator construction #977
• example heat-equation #978
• extend supported compiler combinations gcc-8+nvcc 10.1-10.2 #985
• add support for CMake 3.17 #988
• adds initial files for sphinx/rst and readthedocs. #990 #1017 #1048
• add support for clang 10 #998
• add popcount intrinsic #1004
• emulate hip/cuda-Memcpy3D with a kernel #1014
• simplify alpaka usage #1017

Thanks to Sergei Bastrakov, Simeon Ehrig, Axel Huebl, Jeffrey Kelling, Jakob Krude, Mutsuo Saito, Jan Stephan, Matthias Werner, René Widera, Benjamin Worpitz for contributions to this release!

Compatibility Changes:

• added support for CUDA 10.0, 10.1 and 10.2
• dropped support for CUDA 7.0 and 7.5
• added official support for Visual Studio 2017 on Windows with CUDA 10 (built on Travis CI instead of appveyor now)
• added support for xcode10.2-11.3 (no official CUDA support yet)
• added support for Ubuntu 18.04
• added support for gcc 9
• added support for clang 7.0, 8.0 and 9.0
• dropped support for clang 3.5, 3.6, 3.7, 3.8 and 3.9
• added support for CMake 3.13, 3.14, 3.15 and 3.16
• dropped support for CMake 3.11.3 and lower, 3.11.4 is the lowest supported version
• added support for Boost 1.69, 1.70 and 1.71
• added support for usage of libc++ instead of libstdc++ for clang builds
• removed dependency to Boost.MPL and BOOST_CURRENT_FUNCTION
• replaced Boost.Test with Catch2 using an internal version of Catch2 by default but allowing to use an external one

Bug Fixes:

• fixed some incorrect host/device function attributes
• fixed warning about comparison unsigned < 0
• There is no need to disable all other backends manually when using ALPAKA_ACC_GPU_CUDA_ONLY_MODE anymore
• fixed static block shared memory of types with alignemnt higher than defaultAlignment
• fixed race-condition in HIP/NVCC queue
• fixed data races when a GPU updates host memory by aligning host memory buffers always to 4kib

New Features:

• Added a new alpaka Logo!
• the whole alpaka code has been relicensed to MPL2 and the examples to ISC
• added ALPAKA_CXX_STANDARD CMake option which allows to select the C++ standard to be used
• added ALPAKA_CUDA_NVCC_SEPARABLE_COMPILATION option to enable separable compilation for nvcc
• added ALPAKA_CUDA_NVCC_EXPT_EXTENDED_LAMBDA and ALPAKA_CUDA_NVCC_EXPT_RELAXED_CONSTEXPR CMake options to enable/disable those nvcc options (they were always ON before)
• added headers for standalone usage without CMake (alpaka/standalone/GpuCudaRt.h, ...) which set the backend defines
• added experimental HIP back-end with using nvcc (HIP >= 1.5.1 required, latest rocRand). More on HIP setup: doc/markdown/user/implementation/mapping/HIP.md
• added sincos math function implementations
• allowed to copy and move construct ViewPlainPtr
• added support for CUDA atomics using "unsigned long int"
• added compile-time error for atomic CUDA ops which are not available due to sm restrictions
• added explicit errors for unsupported types/operations for CUDA atomics
• replaced usages of assert with ALPAKA_ASSERT
• replaced BOOST_VERIFY by ALPAKA_CHECK and returned success from all test kernels
• added alpaka::ignore_unused as replacement for boost::ignore_unused

Breaking changes:

• renamed Queue*Async to Queue*NonBlocking and Queue*Sync to Queue*Blocking
• renamed alpaka::size::Size to alpaka::idx::Idx, alpaka::size::SizeType to alpaka::idx::IdxType (and TSize to TIdx internally)
• replaced ALPAKA_FN_ACC_NO_CUDA by ALPAKA_FN_HOST
• replaced ALPAKA_FN_ACC_CUDA_ONLY by direct usage of __device__
• renamed ALPAKA_STATIC_DEV_MEM_CONSTANT to ALPAKA_STATIC_ACC_MEM_CONSTANT and ALPAKA_STATIC_DEV_MEM_GLOBAL to ALPAKA_STATIC_ACC_MEM_GLOBAL
• renamed alpaka::kernel::createTaskExec to alpaka::kernel::createTaskKernel
• QueueCpuSync now correctly blocks when called from multiple threads ** This broke some previous use-cases (e.g. usage within existing OpenMP parallel regions) ** This use case can now be handled with the support for external CPU queues as can bee seen in the example QueueCpuOmp2CollectiveImpl
• previously it was possible to have kernels return values even though they were always ignored. Now kernels are checked to always return void
• renamed all files with *Stl suffix to *StdLib
• renamed BOOST_ARCH_CUDA_DEVICE to BOOST_ARCH_PTX
• executors have been renamed due to the upcoming standard C++ feature with a different meaning. All files within alpaka/exec/ have been moved to alpaka/kernel/ and the files and classes have been renamed from Exec* to TaskKernel*. This should not affect users of alpaka but will affect extensions.

Bug Fixes:

• fix cuda stream race condition #850
• fix: cuda exceptions #844
• math/abs: Added trait specialisation for double. #862
• alpaka/math Overloaded float specialization #837
• Fixes name conflicts in alpaka math functions. #784