The finite element method (FEM) is a well established numerical technique for solving partial differential equations (PDEs) in a wide range of complex science and engineering applications.
This method has two costly operation that are the construction of global matrices and vectors to form the system of linear or nonlinear equations (assemblage), and their solution (solver).
Many efforts have been directed to accelerate the solver.
However, the assembly stage has been less investigated although it may represent a serious bottleneck in iterative processes such as non-linear and time-dependent phenomena, and in optimization procedures involving FEM with unstructured meshes.
Thus, a fast technique for the global FEM matrices construction is proposed herein by using parallel computing on graphics processing units (GPUs).
This work focuses on matrices that arise by solving elliptic PDEs, what is commonly known as stiffness matrix.
For performance tests, a scalar problem typically represented by the thermal conduction phenomenon and a vector problem represented by the structural elasticity are considered in a three-dimensional (3D) domain.
Unstructured meshes with 8-node hexahedral elements are used to discretize the domain.
The MATLAB Parallel Computing Toolbox (PCT) is used to program the CUDA code.
The stiffness matrix are built with three GPU kernels that are the indices computation, the numerical integration and the global assembly.
Symmetry and adequate data precision are used to save memory and runtime.
This proposed methodology allows generating global stiffness matrices from meshes with more than 16.3 millions elements in less than 3 seconds for the scalar problem and up to 3.1 millions for the vector one in 6 seconds using an Nvidia Tesla V100 GPU with 16 GB of memory.
Large speedups are obtained compared with a non-optimized CPU code.
- Finite element method
- Unstructured mesh
- Sparse matrix generation
- Parallel computing
- Graphics processing unit
- Option 1: Clone the GitHub repository and use the code directly on MATLAB.
- Option 2: Search in the
releases
folder and download the latest MATLAB package, e.g.StiffMa1.6.mltb
. After the file is downloaded, double clic on it and it automatically will be installed on MATLAB as a Toolbox.
%% Inputs
nel = 90; % Number of elements at each direction
sets.sf = 1; % Safety factor. Positive integer to add more partitions
sets.prob_type = 'Vector'; % 'Scalar' or 'Vector'
sets.dTE = 'uint32'; % Data precision for "elements"
sets.dTN = 'double'; % Data precision for "nodes"
MP.c = 384.1; % Thermal conductivity (only for scalar problem)
MP.E = 200e9; % Young's modulus (only for vector problem)
MP.nu = 0.3; % Poisson's ratio (only for vector problem)
%% Mesh generation
[Mesh.elements, Mesh.nodes] = CreateMesh2(nel, nel, nel, sets.dTE, sets.dTN);
Settings
[sets.nel, sets.nxe] = size(Mesh.elements); % Number of elements in the mesh & Number of nodes per element
[sets.nnod, sets.dim] = size(Mesh.nodes); % Number of nodes in the mesh & Space dimension
if strcmp(sets.prob_type,'Scalar')
sets.dxn = 1; % Number of DOFs per node for the scalar problem
elseif strcmp(sets.prob_type,'Vector')
sets.dxn = 3; % Number of DOFs per node for the vector problem
else
error('Problem not defined!');
end
sets.edof = sets.dxn * sets.nxe; % Number of DOFs per element
sets.sz = (sets.edof * (sets.edof + 1) )/2; % Number of NNZ values for each Ke using simmetry
sets.tdofs = sets.nnod * sets.dxn; % Number of total DOFs in the mesh
%% GPU setup
dev = gpuDevice; % Selects the GPU device
sets.tbs = dev.MaxThreadsPerBlock; % Thread block size
sets.numSMs = dev.MultiprocessorCount; % Number of GPU multiprocessors
sets.WarpSize = dev.SIMDWidth; % Warp size
%% Stiffness Matrix generation
K = StiffMa(Mesh, MP, dev, sets);
Written by Francisco Javier Ramírez-Gil, [email protected]
Code site:
- https://github.com/fjramireg/StiffMa
- https://www.mathworks.com/matlabcentral/fileexchange/76947-stiffma
Sponsors: