Kokkos Kernels

Kokkos C++ Performance Portability Programming EcoSystem: Math Kernels - Provides BLAS, Sparse BLAS and Graph Kernels

KokkosKernels implements local computational kernels for linear algebra and graph operations, using the Kokkos shared-memory parallel programming model. "Local" means not using MPI, or running within a single MPI process without knowing about MPI. "Computational kernels" are coarse-grained operations; they take a lot of work and make sense to parallelize inside using Kokkos. KokkosKernels can be the building block of a parallel linear algebra library like Tpetra that uses MPI and threads for parallelism, or it can be used stand-alone in your application.

Computational kernels in this subpackage include the following:

(Multi)vector dot products, norms, and updates (AXPY-like operations that add vectors together entry-wise)
Sparse matrix-vector multiply and other sparse matrix / dense vector kernels
Sparse matrix-matrix multiply
Graph coloring
Gauss-Seidel with coloring (generalization of red-black)
Other linear algebra and graph operations

We organize this directory as follows:

Public interfaces to computational kernels live in the src/ subdirectory (kokkos-kernels/src):

Kokkos_Blas1_MV.hpp: (Multi)vector operations that Tpetra::MultiVector uses
Kokkos_Sparse_CrsMatrix.hpp: Declaration and definition of KokkosSparse::CrsMatrix, the sparse matrix data structure used for the computational kernels below
KokkosSparse_spmv.hpp: Sparse matrix-vector multiply with a single vector, stored in a 1-D View + Sparse matrix-vector multiply with multiple vectors at a time (multivectors), stored in a 2-D View

Implementations of computational kernels live in the src/impl/ subdirectory (kokkos-kernels/src/impl)
Correctness tests live in the unit_test/ subdirectory, and performance tests live in the perf_test/ subdirectory
Simple example scripts to build Kokkoskernels are in example/buildlib/

Do NOT use or rely on anything in the KokkosBlas::Impl namespace, or on anything in the impl/ subdirectory.

This separation of interface and implementation lets the interface assign the users' Views to View types with the desired attributes (e.g., read-only, RandomRead). This also makes it easier to provide full specializations of the implementation. "Full specializations" mean that all the template parameters are fixed, so that the compiler can actually compile the code. This technique keeps your library's or application's build times down, since kernels are already precompiled for certain template parameter combinations. It also improves performance, since compilers have an easier time optimizing code in shorter .cpp files.

Building Kokkoskernels

CMake

Following Kokkos style, all CMake options are of the form

KokkosKernels_ENABLE_OPTION

with options capitalized at the end. Almost all Kokkos Kernels options determine whether ETI is used with a particular datatype, e.g.

-DKokkosKernels_INST_DOUBLE=On

which does explicit instantation of all kernels for double type. Kokkos Kernels derives most of its CXXFLAGS, C++ standard, architecture flags, and other options from an installed (or in-tree) Kokkos package. Tuning for a particular device or architecture is generally done through Kokkos while tuning which kernels get instantiated is done through Kokkos Kernels.

Kokkos Kernels does supply flags for asserting properies of the linked Kokkos, for example:

-DKokkosKernels_REQUIRE_DEVICES=CUDA
-DKokkosKernels_REQUIRE_OPTIONS=cuda_relocatable_device_code

This does NOT enable CUDA directly, but rather verifies that the underlying Kokkos supports the desired option. If the underlying Kokkos was not built properly, CMake will crash and tell you to re-build Kokkos. The values (unlike the option names) are not case-sensitive. More details can be found in the build instructions or developer instructions.

Spack

An alternative to manually building with the CMake is to use the Spack package manager. To do so, download the kokkos-spack git repo and add to the package list:

spack repo add $path-to-kokkos-spack

A basic installation would be done as:

spack install kokkos-kernels

Spack allows options and and compilers to be tuned in the install command.

spack install kokkos-kernels@3.0 +double %gcc@7.3.0 +openmp

This example illustrates the three most common parameters to Spack:

Variants: specified with, e.g. +openmp, this activates (or deactivates with, e.g. ~openmp) certain options.
Version: immediately following kokkos-kernels the @version can specify a particular Kokkos Kernels to build
Compiler: a default compiler will be chosen if not specified, but an exact compiler version can be given with the %option.

For a complete list of Kokkos Kernels options, run:

spack info kokkos-kernels

Tuning Kokkos Options

As discussed above in the CMake section, Kokkos Kernels inherits much of its configuration from the installed Kokkos. Spack gives a mechanism for directly specifying Kokkos dependency options:

spack install kokkos-kernels ^kokkos@3.0+cuda+cuda_uvm

The carat ^ sepcifies an exact dependency configuration, which in this case activates CUDA and CUDA_UVM. For a complete list of tunable Kokkos options, run

spack info kokkos

Settuping a development environment with Spack

Spack is generally most useful for installng packages to use. If you want to install all dependencies of Kokkos Kernels first so that you can actively develop a given Kokkos Kernels source this can still be done. Go to the Kokkos Kernels source code folder and run:

spack diy -u cmake kokkos-kernels@{version} ...

specifying the exact version you want to develop and giving any spec options in .... This creates a folder spack-build where you can make.

Trilinos

For Trilinos builds with the Cuda backend and complex double enabled with ETI, the cmake option below may need to be set to avoid Error 127 errors: CMAKE_CXX_USE_RESPONSE_FILE_FOR_OBJECTS:BOOL=ON

If the option above is not set, a warning will be issued during configuration:

"The CMake option CMAKE_CXX_USE_RESPONSE_FILE_FOR_OBJECTS is either undefined or OFF. Please set CMAKE_CXX_USE_RESPONSE_FILE_FOR_OBJECTS:BOOL=ON when building with CUDA and complex double enabled."

Using Kokkoskernels Test Drivers

In perf_test there are test drivers.

KokkosGraph_triangle.exe : Triangle counting driver.
KokkosSparse_spgemm.exe : Sparse Matrix Sparse Matrix Multiply:
*NOTE: KKMEM is outdated. Use default algorithm: KKSPGEMM = KKDEFAULT = DEFAULT
Or within the code:

kh.create_spgemm_handle(KokkosSparse::SPGEMM_KK);

KokkosSparse_spmv.exe : Sparse matvec.
KokkosSparse_pcg.exe : CG method with Gauss Seidel as preconditioner.
KokkosGraph_color.exe : Distance-1 Graph coloring
KokkosKernels_MatrixConverter.exe : given a matrix market format, converts it ".bin" binary format for fast input output readings, which can be read by other test drivers.

Please report bugs or performance issues to: https://github.com/kokkos/kokkos-kernels/issues

LICENSE

KokkosKernels is licensed under standard 3-clause BSD terms of use. For specifics, please refer to the LICENSE file contained in the repository or distribution. Under the terms of Contract DE-NA0003525 with NTESS, the U.S. Government retains certain rights in this software.

ytopt-team/kokkos-kernels