Skip to content

v2023.06.0
Compare
Choose a tag to compare
@rhornung67 rhornung67 released this 06 Jul 18:15
· 493 commits to main since this release
v2023.06.0
e330b25
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 to improve GPU kernel performance and some bug fixes. It contains one breaking change described below and an execution policy deprecation also described below. The policy deprecation is not a breaking change in this release, but will result in a breaking change in the next release.

Please download the RAJA-v2023.06.0.tar.gz file below. The others, generated by GitHub, may not work for you due to RAJA's dependencies on git submodules.

Notable changes include:

  • New features / API changes:

    • In this release the loop_exec execution policy is deprecated and will be removed in the next release. RAJA has had two sequential execution policies for some time, seq_exec and loop_exec. When using the seq_exec execution policy, RAJA would attach #pragma novector, or similar depending on the compiler, to force strictly sequential execution of a loop; e.g., by preventing a compiler from vectorizing a loop, even if it was correct to do so. When the loop_exec policy was specified, the compiler was allowed to apply any optimizations, including SIMD, that its heuristics determined were appropriate. In this release, seq_exec behaves the same as how loop_exec behaves historically and the loop_exec and associated policies, such as loop_atomic, loop_reduce, etc. are type aliases to the analogous seq_exec policies. This prevents breaking user code with this release. However, users should prepare to switch loop_exec policies to the seq_exec policy variants in the future.
    • GPU global (thread and block) indexing has been refactored to abstract indexing in a given dimension. The result is that users can now specify a block size or a grid size at compile time or get those values at run time. You can also ignore blocks and index only with threads and vice versa. Kernel and launch policies are now shared. Such policies are now multi-part and contain global indexing information, a way to map global indices like direct or strided loops, and have a synchronization requirement. The synchronization allows one to request that all threads complete even if some have no work so you can synchronize a block. Aliases have been added for all of the preexisting policies and some are deprecated in favor of policies named more consistently. One BREAKING CHANGE is that thread loop policies are no longer safe to block synchronize. That feature still exists but can only be accessed with a custom policy. The RAJA User Guide contains descriptions of the new policy mechanics.

  • Build changes/improvements:

    • Update BLT submodule to v0.5.3
    • Update camp submodule to v2023.06.0

  • Bug fixes/improvements:

    • Fixes a Windows build issue due to macro file definition logic in a RAJA header file. Specifically, the macro constant RAJA_COMPILER_MSVC was not getting defined properly when building on a Windows platform using a compiler other than MSVC.
    • Kernels using the RAJA OpenMP target back-end were not properly seg faulting when expected to do so. This has been fixed.
    • Various improvements, compilation and execution, in RAJA SIMD support.
    • Various improvements and additions to RAJA tests to cover more end-user cases.
Assets 3