RAJA Performance Portability: LLNL's Loop Abstraction — separating algorithm specification from execution policy

RAJA is LLNL's C++ performance portability framework emphasizing explicit loop structure and backend-agnostic execution policies. Unlike Kokkos' view abstraction, RAJA focuses on kernel specification via forall and kernel constructs.

RAJA Execution Policies

forall(RangeSegment(0, N), policy, [](int i) { ... }) encapsulates parallelism. Policies include: cuda_exec (GPU blocks/threads), omp_parallel_for_exec (OpenMP parallel), seq_exec (serial). Policy hierarchy enables composition: nested_loop_exec for loop tiling and blocking. Backend selection is compile-time; same code compiles to different policies for CUDA/HIP/OpenMP/serial.

RAJA::View Layout Abstraction

RAJA::View provides layout-agnostic array access. Template parameters specify: dimensionality, element type, layout (ColMajor/RowMajor), index ordering. STRIDE1 macro converts multi-dimensional indices to linear addresses via layout abstraction. Performance portability: optimal layout differs across architectures (NVIDIA prefers LayoutRight for coalescing, CPUs prefer LayoutLeft for cache).

Atomic Operations and Team Patterns

RAJA atomics (AtomicAdd, AtomicMin) port across backends. TeamExecute enables team-level parallelism: outer loop threads (blocks), inner parallel loop (warp/team), team reduction, and barrier synchronization. Synchronous/asynchronous kernels enable flexible execution synchronization.

CHAI Memory Management

CHAI (Collected Abstraction of Host and Innovations for Execution) manages data movement between execution spaces. CHAI::ManagedArray automatically migrates data to execution space on access, eliminating explicit host/device transfers. MOVE semantics enable explicit prefetching; COPY semantics enable synchronization. Integration with RAJA is seamless: access CHAI arrays via RAJA kernels, automatic data management.

Production Adoption

ARDRA (LLNL radiation hydrodynamics), MFEM (modular finite element methods), and Ares (multiphysics) leverage RAJA. Lawrence Livermore National Laboratory emphasizes RAJA for production codes requiring multi-architecture support (NVIDIA/AMD/CPU).