Skip to content

v0.14.0
Compare
Choose a tag to compare
@rhornung67 rhornung67 released this 19 Aug 16:40
· 3669 commits to main since this release
v0.14.0
357933a
This commit was created on GitHub.com and signed with GitHub’s verified signature. The key has expired.
GPG key ID: 4AEE18F83AFDEB23
Expired
Learn about vigilant mode.

This release contains new features, bug fixes, and build improvements. Please see the RAJA user guide for more information about items in this release.

Please download the RAJA-v0.14.0.tar.gz file below. The others will not work due to the way RAJA uses git submodules.

Notable changes include:

  • New features / API changes:

    • Initial release of some SYCL execution back-end features for supporting Intel GPUs. Users should be able to exercise RAJA::forall, basic RAJA::kernel, and reductions. Future releases will contain additional RAJA feature support for SYCL.
    • Various enhancements to the experimental RAJA "teams" capability, including documentation and complete code examples illustrating usage.
    • The RAJA "teams" interface was expanded to initial support for RAJA/camp resources.
    • The RAJA "teams" interface was expanded to allow users to label kernels with name strings to easily attribute execution timings and other details to specific kernels with NVIDIA profiling tools, for example. Usage information is available in the RAJA User Guide. Kernel naming will be available for all other RAJA kernel execution methods in a future release.
    • Deprecated sort and scan methods taking iterators have been removed, Now, these methods take RAJA span arguments. For example, (begin, end) args are replaced with RAJA::make_span(begin, N), where N = end - begin. Please see the RAJA User Guide documentation for scan and sort operations for details and usage examples.
    • Sort and scan methods now accept an optional resource argument.
    • Methods were added to the RAJA::kernel API to accept a resource argument; specifically 'kernel_resource' and 'kernel_param_resource'. These kernel methods return an Event object similar to the RAJA::forall interface.
    • RAJA resource support added to RAJA workgroup and worksite constructs.
    • OpenMP CPU multithreading policies have been reworked so that usage involving OpenMP scheduling are consistent. Specification of a chunk size for scheduling policies is optional, which is consistent with native OpenMP usage. In addition, no-wait policies are more constrained to prevent potentially non-conforming (to the OpenMP spec) usage. Finally, additional policy type aliases have been added to make common use cases less verbose. Please see the RAJA policy documentation in the User Guide for policy descriptions.
    • Host implementation of HIP atomics added.
    • Add ability to specify atomic to use on the host in CUDA and HIP atomic policies (i.e., added host atomic template parameter), This is useful for host-device decorated lambda expressions that may be used for either host or device execution. It also fixes compilationissues with Hip atomic compilation in host-device contexts.
    • The RAJA Registry API has been changed to return raw pointers to registry objects rather than shared_ptr type. This is better for performance.
    • New content has been added to the RAJA Developer Guide available in the Read The Docs Sphinx documentation. This should help folks align their work with RAJA processes when making contributions to RAJA.
    • Basic doxygen source code documentation is now available via a link in our Read The Docs Sphinx documentation.
    • Unified memory implementation for storing indices in TypedListSegment, which was marked deprecated in v0.12.0 release has been removed. Now, TypedListSegment constructor requires a camp resource object to be passed which indicates the memory space where the indices will live. Specifically, the array of indices passed to the constructor by a user (assumed to live in host memory for the "owned" case) will be copied to an internally owned allocation in the memory space defined by the resource object.
    • The ListSegment constructor takes a resource by value now, previously taken by reference, which allows more resource argument types to be passed more seamlessly to the List Segment constructor.
    • 'CudaKernelFixedSM' and 'CudaKernelFixedSMAsync' methods were added which allow users to specify the minimum number of thread blocks to launch per SM. This resulted in a performance improvement for an application use case. Future work will expand this concept to other GPU kernel execution methods in RAJA.

  • Build changes/improvements:

    • Update BLT submodule to latest release, v0.4.1.
    • Update camp submodule to latest tagged release, v0.2.2
    • The RAJA_CXX_STANDARD_FLAG CMake variable was removed. The BLT_CXX_STD variable is now used instead.
    • Support for building RAJA as a shared library on Windows has been added.
    • A build system adjustment was made to address an issue when RAJA is built with an external version of camp (e.g., through Spack).
    • The build default has been changed to use the version of CUB that is installed in the specified version of the CUDA toolkit, if available, when CUDA is enabled. Similarly, for the analogous functionality in HIP. Specific versions of these libraries can still be specified for a RAJA build. Please see the RAJA User Guide for details.
    • The build system now uses the BLT cmake_dependent_options support for options defined by BLT. This avoids shadowing of BLT options by options defined in RAJA and in the cases where RAJA is used as a sub-module in another BLT project. For example, it provides the ability to disable RAJA tests and examples at a more fine granularity.
    • Checks were added to the RAJA CMake build system to check for minimum equired versions of CUDA (9.2) and HIP (3.5).
    • A build system bug was fixed so that targets for third-party dependencies provided by BLT (e.g., CUDA and HIP) are exported properly. This allows non-BLT projects to use the imported RAJA target.
    • An issue was fixed to appease the MSVC 2019 compiler.
    • Improvements to build system to address Hip linking issues.

  • Bug fixes/improvements:

    • Hip and CUDA block reductions were tweaked to fix the number of steps in the final wavefront/warp reduction. This saves a couple rounds of warp shfls.
    • A runtime bug resulting from defaulted View constructors not being implemented correctly in CUDA 10.1 is fixed. This fixes an issue with CHAI managed arrays not having their copy constructor being triggered properly.
    • Fix bug that caused a CUDA or HIP synchronization error when a zero length loop was enqueued in a workgroup.
    • Added missing HIP workgroup unordered execution policy, so HIP version is consistent with CUDA version.
    • Fixed issue where the RAJA non-resource API returns an EventProxy object with a dangling resource pointer, by getting a reference to the default resource for the execution context.
    • IndexSet utility methods for collecting indices into a separate container now work with any index type.
    • The volatile qualifier was removed from a type conversion function used in RAJA atomics. This fixes a performance issue with HIP where the value was written to stack memory during type conversion.
    • Numerous improvements, updates, and fixes (formatting, typos, etc.) in RAJA User Guide.
Assets 3
Loading