Parallel Advection

This code implements a 1D advection operator inside a multidimensionnal space. It implements a semi-Lagrangian scheme using C++ parallel progamming models (SYCL, Kokkos).

General algorithm

For one time step, this is the algorithm's structure:

SYCL Implementations

Tools:

• local_accessor: to allocate shared memory between work items of a same work group
• groups barriers on work groups
• simple mapping between physical and logical iteration space

Impls (folder sycl/src/core/.../impl):

• Basic Range
  • 1D
  • 2D
• ND-Range
• Hierarchical
• Scoped

Kokkos Implementations

Tools:

• ScratchSpace memory
• Barriers

Implementations (folder kokkos/src/core/.../impl):

• MDRange
• TeamPolicy

Build the project (see instructions on branches):

You can use the compile.sh script to compile for various hardware and sycl-implementations. Try compile.sh --help to see available options.

mkdir build
cd build
cmake .. -DCMAKE_CXX_COMPILER=/path/to/sycl/compiler
make

Run the executable

1. Set the runtime parameters in build/.../advection.ini

[run]
# Total number of iterations
maxIter = 200
# Wheter to run on the GPU or CPU #only for SYCL
gpu     = true
# The kernel type to use for Xadvection
kernelImpl  = Hierarchical  
# Size of work groups use in the kernels
workGroupSize = 32
# Outputs a solution.log file to be read with the python notebook
outputSolution = false

[geometry]
nx  = 1024 # nb of spatial points
nvx = 2 # nb of speed points
n_fict_dim = 1  # nb of points in the fictive dimension

[discretization]
dt  = 0.001

minRealX  = 0
maxRealX  = 1
minRealVx = -1
maxRealVx = 1

Deltas $d_{vx}$ and $d_x$ are deduced by the number of points and min/max values.

2. Run the executable build/src/.../advection

Credits

This code is largely inspired by the vlp4D code