You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Updated RAJA (v2026.07.1), CHAI (v2026.07.1), Umpire (v2026.07.2), BLT (v0.7.2), Caliper (2.15.0), and added Adiak (0.5.0).
Added explicit Umpire memory management alongside optional CHAI support, including immediate GPU allocation and initialization, and field copies.
Added automatic device-pool sizing based on problem dimensions and decomposition, a --dev_pool_size override, and reporting of pool size and memory high watermark.
Added GPU-aware MPI, and optional direct Umpire allocations for communication planes.
Added asynchronous CUDA/HIP sweep execution and processing of all ready subdomains during sweep scheduling.
Corrected GPU kernel loop ordering, including the LTimes GDZ case, field-copy direction, and host/device synchronization and transfers.
Expanded Caliper instrumentation to cover main() and solver iterations. Added Adiak metadata for configuration, problem size, performance, and device-pool usage.
Adopted C++20, reorganized CMake into a superbuild (to accomodate Adiak), improved bundled/external profiling-library integration, and added CUDA 13 compatibility.
Added H100 CUDA configurations for GCC and Clang, MI300A and Intel 2024 configurations, and updated existing ROCm and compiler configurations.