zer011b/fdtd3d

fdtd3d is an open source 1D, 2D, 3D FDTD electromagnetics solver with MPI, OpenMP and CUDA support for x64, ARM, ARM64, RISC-V, PowerPC architectures

fdtd3d's Issues

Fix build without parallel grid

Share fails on 3D parallel grid

3D parallel grid fails to share in yz and xyz modes

Check wrong index in cuda kernels

Cuda kernels should return CUDA_STATUS_ERROR if thread block or index are above grid size

Smaller dimension parallel buffers for other directions

I.e. "y" for 1D or "xz", "yz" for 2D, etc.

Grid share syncronization

• Grid in Scheme should be verified to have the same division over nodes
• During calculation process share should be done after usage of the last row

2D modes metamaterial

Add metamaterials for 2D modes

• TEz
• TMz

Build types

Add build flags for debug and release builds

Non-square grid sizes

Add support on non-square 2-dimensional parallel buffers for 2D and 3D grids and non-qubical 3-dimensional parallel buffers for 3D grids.

• 2D grid
• 3D grid

PML for auxiliary grid for TF/SF

Reflection from right border takes place. Add pml for 1D grid

Remove redundancies from send/receive

• Move redundant calculations from send/receive to initialization
• Add nodeGridSizeX*nodeGridSizeY constant to improve runtime perf, etc.

Compile & link flags

Choose compile and link flags for code checks and best performance.

Move actual FieldValue send/receive to it

Allocate memory for copy to cuda kernels only once

Allocate memory for copy to cuda kernels only once before execution of cuda kernels

Dumper/Loader for 3D mode

Add dumper and loader for 3D grids

Concurrency for 3D mode

Add concurrency (MPI) for 3D mode

• Cube concurrency. (Number of processes - n^3)
• Square concurrency.
• Linear concurrency.

PML layers

Investigate other PML layers than UPML

Gather grid from all nodes for output

Implement gatherer of grid from all nodes

3D mode PML

Add UPML for 3D mode solver

Dumper output size

Check dumper printed grid size

Share threshold

• Remove each time step share procedure and check threshold, so share each K steps.
• For left buffer and right buffer of different size, share could be performed not at the minimum of their K1 and K2 steps but separately for each of left and right.

Improve build test scripts

Replace left and right buffer with just one buffer

Replace -DFULL_VALUES with -DVALUES_TYPE

Access to grid elements

Access grid elements with getFieldPointValue by absolute coordinate and check that coordinate corresponds to this process

4th order accuracy

Investigate 4th order accuracy for fdtd method

Cover code with ifdef

Not all code is covered with ifdef. Fix it

2D TF/SF

Add total field/scattering field for 2D mode

Unify parallel grid code

Parallel share could be unified more

• Grid interface class. Heirs - Grid1D - Grid2D - Grid3D. ParallelGrid - heir of specified one during build.
• Unify parallel grid constructor

Border buffers

Do not initialize buffers for border nodes of grid

• 2D grid
• 3D grid

2D Hz mode

Add 2D TEz mode solver

Material on the border of parallel grid nodes

Results of material fully in grid of single node and share are not the same

3D mode

Add simple 3D mode solver

Dumper/Loader with parameter what part to save

Pass grid coordinate to dumper/loader specifying what part of grid to save/load

3D mode TF/SF

Add total field/scattering field for 3D move solver

TMz missing dump for Ez

TMz has missing dump for Ez from time to time

Fix code style

Fix code style using vera++.

Check CI

Concurrency for 1D mode

• Linear concurrency

Merge parallel grid into master

Test parallel grid

• Test all parallel modes
• Add unit tests

Set sm for gpu

sm_50 is hard coded in CMakeLists.txt

Future optimizations

To be updated...

2D PML

Add UPML for 2D modes

3D mode metamaterials

Add 3D mode metamaterials

Remove redundant elements

Every node allocates node size plus size of buffers. Some nodes do not need some buffers, fix this

Wrong processTo and processFrom

Improve processTo and processFrom in parallel grid

Refactor code structure

1. Replace single .cpp file for solver by additional files.
2. Move to C++
3. ...

Concurrency for 2D mode

• Square concurrency. (Number of processes - n^2)
• Linear concurrency

Wrong color chooser in dumper

Dumper chooses wrong colors for grid points. Investigate

2D Ez mode

Add 2D Ez mode solver