Trilinos/Kokkos: Core Kernels Package

Version of the Day

Introduction

Kokkos provides two main capabilities for Trilinos. The first is an abstract programming interface (API) for handling memory and specify work for shared-memory parallel nodes (the Kokkos Node API). The second is a collection of local (non-distributed) distributed linear algebra classes and the kernels needed to realized their functionality in parallel (the Kokkos Linear Algebra kernels library).

Kokkos is typically not used in applications, but instead as a significant component in higher-level Trilinos packages, such as Tpetra. The functionality contained here-in has been put in the Kokkos package in order to facilitate code re-use and ease experimentation for kernel researchers.

Overview of Kokkos.

Kokkos Classes

Kokkos contains a number of classes. They can be categorized as follows:

• Kokkos Node API

  • Nodes based on the Kokkos::CUDANodeMemoryModel: GPU nodes using the NVIDIA/CUDA graphic processing units (GPUs)
    • Kokkos::ThrustGPUNode provides parallel compute capability for via the Thrust library.
  • Nodes based on the Kokkos::StandardNodeMemoryModel: Host-bound/CPU nodes using standard memory allocation.
    • Kokkos::SerialNode: provides a simple node, where all kernels are executed in serial.
    • Kokkos::TBBNode: provides support for multi-core CPUs using Intel TBB (requires third-party library TBB).
    • Kokkos::TPINode: provides support for multi-core CPUs using Pthreads, via the ThreadPool package (requires ThreadPool and Pthreads TPL).

• Kokkos Linear Algebra Kernels Library

  • Data structures:

    1. Kokkos::CrsMatrix, Kokkos::CrsGraph class: Classes providing the storage structure for a row-oriented sparse matrix.

    2. Kokkos::MultiVector class: A class providing the structure of a dense multivector. Here, as throughout most of Trilinos, a "multivector" refers to a collection of vectors, typically stored in a contiguous allocation of memory.

  • Arithmetic interfaces:

    1. Kokkos::DefaultArithmetic: Provides basic functionality for perfoming arithmetic of multivectors.

    2. Kokkos::DefaultSparseMultiply, Kokkos::DefaultSparseSolve: Provides basic functionality for computing multiplication and solve of a sparse matrix against a multivector.

tpetra_trilinos

Kokkos can be used as a stand-alone package. However, it also provides functionality for Tpetra and Tifpack.