Coverage for pySDC/implementations/problem_classes/generic_MPIFFT_Laplacian.py: 100%
89 statements
« prev ^ index » next coverage.py v7.6.9, created at 2024-12-20 14:51 +0000
1import numpy as np
2from mpi4py import MPI
3from mpi4py_fft import PFFT, newDistArray
5from pySDC.core.errors import ProblemError
6from pySDC.core.problem import Problem, WorkCounter
7from pySDC.implementations.datatype_classes.mesh import mesh, imex_mesh
10class IMEX_Laplacian_MPIFFT(Problem):
11 r"""
12 Generic base class for IMEX problems using a spectral method to solve the Laplacian implicitly and a possible rest
13 explicitly. The FFTs are done with``mpi4py-fft`` [1]_.
14 Works in two and three dimensions.
16 Parameters
17 ----------
18 nvars : tuple, optional
19 Spatial resolution
20 spectral : bool, optional
21 If True, the solution is computed in spectral space.
22 L : float, optional
23 Denotes the period of the function to be approximated for the Fourier transform.
24 alpha : float, optional
25 Multiplicative factor before the Laplacian
26 comm : MPI.COMM_World
27 Communicator for parallelisation.
29 Attributes
30 ----------
31 fft : PFFT
32 Object for parallel FFT transforms.
33 X : mesh-grid
34 Grid coordinates in real space.
35 K2 : matrix
36 Laplace operator in spectral space.
38 References
39 ----------
40 .. [1] Lisandro Dalcin, Mikael Mortensen, David E. Keyes. Fast parallel multidimensional FFT using advanced MPI.
41 Journal of Parallel and Distributed Computing (2019).
42 """
44 dtype_u = mesh
45 dtype_f = imex_mesh
47 xp = np
48 fft_backend = 'fftw'
49 fft_comm_backend = 'MPI'
51 @classmethod
52 def setup_GPU(cls):
53 """switch to GPU modules"""
54 import cupy as cp
55 from pySDC.implementations.datatype_classes.cupy_mesh import cupy_mesh, imex_cupy_mesh
57 cls.xp = cp
59 cls.dtype_u = cupy_mesh
60 cls.dtype_f = imex_cupy_mesh
62 cls.fft_backend = 'cupy'
63 cls.fft_comm_backend = 'NCCL'
65 def __init__(
66 self, nvars=None, spectral=False, L=2 * np.pi, alpha=1.0, comm=MPI.COMM_WORLD, dtype='d', useGPU=False, x0=0.0
67 ):
68 if useGPU:
69 self.setup_GPU()
71 if nvars is None:
72 nvars = (128, 128)
74 if not (isinstance(nvars, tuple) and len(nvars) > 1):
75 raise ProblemError('Need at least two dimensions for distributed FFTs')
77 # Creating FFT structure
78 self.ndim = len(nvars)
79 axes = tuple(range(self.ndim))
80 self.fft = PFFT(
81 comm,
82 list(nvars),
83 axes=axes,
84 dtype=dtype,
85 collapse=True,
86 backend=self.fft_backend,
87 comm_backend=self.fft_comm_backend,
88 )
90 # get test data to figure out type and dimensions
91 tmp_u = newDistArray(self.fft, spectral)
93 L = np.array([L] * self.ndim, dtype=float)
95 # invoke super init, passing the communicator and the local dimensions as init
96 super().__init__(init=(tmp_u.shape, comm, tmp_u.dtype))
97 self._makeAttributeAndRegister(
98 'nvars', 'spectral', 'L', 'alpha', 'comm', 'x0', 'useGPU', localVars=locals(), readOnly=True
99 )
101 self.getLocalGrid()
102 self.getLaplacian()
104 # Need this for diagnostics
105 self.dx = self.L[0] / nvars[0]
106 self.dy = self.L[1] / nvars[1]
108 # work counters
109 self.work_counters['rhs'] = WorkCounter()
111 def getLocalGrid(self):
112 X = list(self.xp.ogrid[self.fft.local_slice(False)])
113 N = self.fft.global_shape()
114 for i in range(len(N)):
115 X[i] = self.x0 + (X[i] * self.L[i] / N[i])
116 self.X = [self.xp.broadcast_to(x, self.fft.shape(False)) for x in X]
118 def getLaplacian(self):
119 s = self.fft.local_slice()
120 N = self.fft.global_shape()
121 k = [self.xp.fft.fftfreq(n, 1.0 / n).astype(int) for n in N]
122 K = [ki[si] for ki, si in zip(k, s)]
123 Ks = list(self.xp.meshgrid(*K, indexing='ij', sparse=True))
124 Lp = 2 * np.pi / self.L
125 for i in range(self.ndim):
126 Ks[i] = (Ks[i] * Lp[i]).astype(float)
127 K = [self.xp.broadcast_to(k, self.fft.shape(True)) for k in Ks]
128 K = self.xp.array(K).astype(float)
129 self.K2 = self.xp.sum(K * K, 0, dtype=float)
131 def eval_f(self, u, t):
132 """
133 Routine to evaluate the right-hand side of the problem.
135 Parameters
136 ----------
137 u : dtype_u
138 Current values of the numerical solution.
139 t : float
140 Current time at which the numerical solution is computed.
142 Returns
143 -------
144 f : dtype_f
145 The right-hand side of the problem.
146 """
148 f = self.dtype_f(self.init)
150 f.impl[:] = self._eval_Laplacian(u, f.impl)
152 if self.spectral:
153 tmp = self.fft.backward(u)
154 tmp[:] = self._eval_explicit_part(tmp, t, tmp)
155 f.expl[:] = self.fft.forward(tmp)
157 else:
158 f.expl[:] = self._eval_explicit_part(u, t, f.expl)
160 self.work_counters['rhs']()
161 return f
163 def _eval_Laplacian(self, u, f_impl, alpha=None):
164 alpha = alpha if alpha else self.alpha
165 if self.spectral:
166 f_impl[:] = -alpha * self.K2 * u
167 else:
168 u_hat = self.fft.forward(u)
169 lap_u_hat = -alpha * self.K2 * u_hat
170 f_impl[:] = self.fft.backward(lap_u_hat, f_impl)
171 return f_impl
173 def _eval_explicit_part(self, u, t, f_expl):
174 return f_expl
176 def solve_system(self, rhs, factor, u0, t):
177 """
178 Simple FFT solver for the diffusion part.
180 Parameters
181 ----------
182 rhs : dtype_f
183 Right-hand side for the linear system.
184 factor : float
185 Abbrev. for the node-to-node stepsize (or any other factor required).
186 u0 : dtype_u
187 Initial guess for the iterative solver (not used here so far).
188 t : float
189 Current time (e.g. for time-dependent BCs).
191 Returns
192 -------
193 me : dtype_u
194 The solution as mesh.
195 """
196 me = self.dtype_u(self.init)
197 me[:] = self._invert_Laplacian(me, factor, rhs)
199 return me
201 def _invert_Laplacian(self, me, factor, rhs, alpha=None):
202 alpha = alpha if alpha else self.alpha
203 if self.spectral:
204 me[:] = rhs / (1.0 + factor * alpha * self.K2)
206 else:
207 rhs_hat = self.fft.forward(rhs)
208 rhs_hat /= 1.0 + factor * alpha * self.K2
209 me[:] = self.fft.backward(rhs_hat)
210 return me