Parthenon -- a performance portable block-structured adaptive mesh refinement framework

• High performance by
  • device first/device resident approach (work data only in device memory to prevent expensive transfers between host and device)
  • transparent packing of data across blocks (to reduce/hide kernel launch latency)
  • direct device-to-device communication via asynchronous MPI communication
• Intermediate abstraction layer to hide complexity of device kernel launches
• Flexible, plug-in package system
• Abstract variables controlled via metadata flags
• Support for particles
• Support for cell-, node-, face-, and edge-centered fields
• Multi-stage drivers/integrators with support for task-based parallelism

• Chat room on matrix.org

• CMake 3.16 or greater
• C++17 compatible compiler
• Kokkos 4.0.1 or greater

• MPI
• OpenMP
• HDF5 (for outputs)
• Ascent (for in situ visualization and analysis)

• catch2 (for unit tests)
• python3 (for regression tests)
• numpy (for regression tests)
• matplotlib (optional, for plotting results of regression tests)

For detailed instructions for a given system, see our build doc.

mkdir build
cd build
cmake ../
cmake --build . -j 8
ctest

// Imports all of parthenon's public interface
#include <parthenon/parthenon.hpp>

// You can use one of the following headers instead if you want to limit how
// much you import. They import Parthenon's Driver and Package APIs,
// respectively
#include <parthenon/driver.hpp>
#include <parthenon/package.hpp>

// The following namespaces are good short-hands to import commonly used names
// for each set of Parthenon APIs.
using namespace parthenon::driver::prelude;
using namespace parthenon::package::prelude;

Following options are available to configure the default behavior of the par_for wrappers.

• PAR_LOOP_LAYOUT (sets default layout)
  • MANUAL1D_LOOP maps to Kokkos::RangePolicy (default for CUDA backend)
  • MDRANGE maps to Kokkos::MDRangePolicy
  • SIMDFOR_LOOP maps to standard for loops with #pragma omp simd (default for OpenMP backend)
  • TPTTR_LOOP maps to double nested loop with Kokkos::TeamPolicy and Kokkos::ThreadVectorRange
  • TPTVR_LOOP maps to double nested loop with Kokkos::TeamPolicy and Kokkos::ThreadVectorRange
  • TPTTRTVR_LOOP maps to triple nested loop with Kokkos::TeamPolicy, Kokkos::TeamThreadRange and Kokkos::ThreadVectorRange

Similarly, for explicit nested paralellism the par_for_outer and par_for_inner wrappers are available. par_for_outer always maps to a Kokkos::TeamPolicy and the par_for_inner mapping is controlled by the

• PAR_LOOP_INNER_LAYOUT (sets default innermost loop layout for par_for_inner)
  • SIMDFOR_INNER_LOOP maps to standard for loops with #pragma omp simd (default for OpenMP backend)
  • TVR_INNER_LOOP maps to Kokkos::TeamVectorRange (default for CUDA backend)

Kokkos can be configured through cmake options, see https://github.com/kokkos/kokkos/wiki/Compiling

For example to build with the OpenMP backend for Intel Skylake architecture using Intel compilers

mkdir build-omp-skx && cd build-omp-skx
cmake -DKokkos_ENABLE_OPENMP=ON -DCMAKE_CXX_COMPILER=icpc -DKokkos_ARCH_SKX=ON ../

or to build for NVIDIA V100 GPUs (using nvcc compiler for GPU code, which is automatically picked up by Kokkos)

mkdir build-cuda-v100 && cd build-cuda-v100
cmake -DKokkos_ENABLE_CUDA=ON -DKokkos_ARCH_VOLTA70=On ../

or to build for AMD MI100 GPUs (using hipcc compiler)

mkdir build-hip-mi100 && cd build-hip-mi100
cmake -DKokkos_ENABLE_HIP=ON -DCMAKE_CXX_COMPILER=hipcc -DKokkos_ARCH_Vega908=ON ../

Please see the developer guidelines for additional information.

Please see the docs for additional documentation on features and how to use them.