Source code for sfepy.solvers.ls

from __future__ import absolute_import
import hashlib

import numpy as nm
import warnings

import scipy.sparse as sps
import six
from six.moves import range

warnings.simplefilter('ignore', sps.SparseEfficiencyWarning)

from sfepy.base.base import output, get_default, assert_, try_imports
from sfepy.base.timing import Timer
from sfepy.solvers.solvers import LinearSolver

[docs] def solve(mtx, rhs, solver_class=None, solver_conf=None): """ Solve the linear system with the matrix `mtx` and the right-hand side `rhs`. Convenience wrapper around the linear solver classes below. """ solver_class = get_default(solver_class, ScipyDirect) solver_conf = get_default(solver_conf, {}) solver = solver_class(solver_conf, mtx=mtx) solution = solver(rhs) return solution
def _get_cs_matrix_hash(mtx, chunk_size=100000): def _gen_array_chunks(arr): ii = 0 while len(arr[ii:]): yield arr[ii:ii+chunk_size].tobytes() ii += chunk_size sha1 = hashlib.sha1() for chunk in _gen_array_chunks(mtx.indptr): sha1.update(chunk) for chunk in _gen_array_chunks(mtx.indices): sha1.update(chunk) for chunk in _gen_array_chunks(mtx.data): sha1.update(chunk) digest = sha1.hexdigest() return digest def _is_new_matrix(mtx, mtx_digest, force_reuse=False): if not isinstance(mtx, sps.csr_matrix): return True, mtx_digest if force_reuse: return False, mtx_digest id0, digest0 = mtx_digest id1 = id(mtx) digest1 = _get_cs_matrix_hash(mtx) if (id1 == id0) and (digest1 == digest0): return False, (id1, digest1) return True, (id1, digest1)
[docs] def standard_call(call): """ Decorator handling argument preparation and timing for linear solvers. """ def _standard_call(self, rhs, x0=None, conf=None, eps_a=None, eps_r=None, i_max=None, mtx=None, status=None, context=None, **kwargs): timer = Timer(start=True) conf = get_default(conf, self.conf) mtx = get_default(mtx, self.mtx) status = get_default(status, self.status) context = get_default(context, self.context) assert_(mtx.shape[0] == mtx.shape[1] == rhs.shape[0]) if x0 is not None: assert_(x0.shape[0] == rhs.shape[0]) result = call(self, rhs, x0, conf, eps_a, eps_r, i_max, mtx, status, context=context, **kwargs) if isinstance(result, tuple): result, n_iter = result else: n_iter = -1 # Number of iterations is undefined/unavailable. elapsed = timer.stop() if status is not None: status['time'] = elapsed status['n_iter'] = n_iter return result return _standard_call
[docs] def petsc_call(call): """ Decorator handling argument preparation and timing for PETSc-based linear solvers. """ def _petsc_call(self, rhs, x0=None, conf=None, eps_a=None, eps_r=None, i_max=None, mtx=None, status=None, comm=None, context=None, **kwargs): timer = Timer(start=True) conf = get_default(conf, self.conf) mtx = get_default(mtx, self.mtx) status = get_default(status, self.status) context = get_default(context, self.context) comm = get_default(comm, self.comm) mshape = mtx.size if isinstance(mtx, self.petsc.Mat) else mtx.shape rshape = [rhs.size] if isinstance(rhs, self.petsc.Vec) else rhs.shape assert_(mshape[0] == mshape[1] == rshape[0]) if x0 is not None: xshape = [x0.size] if isinstance(x0, self.petsc.Vec) else x0.shape assert_(xshape[0] == rshape[0]) result = call(self, rhs, x0, conf, eps_a, eps_r, i_max, mtx, status, comm, context=context, **kwargs) elapsed = timer.stop() if status is not None: status['time'] = elapsed status['n_iter'] = self.ksp.getIterationNumber() return result return _petsc_call
[docs] class ScipyDirect(LinearSolver): """ Direct sparse solver from SciPy. """ name = 'ls.scipy_direct' _parameters = [ ('method', "{'auto', 'umfpack', 'superlu'}", 'auto', False, 'The actual solver to use.'), ('use_presolve', 'bool', False, False, 'If True, pre-factorize the matrix.'), ('use_mtx_digest', 'bool', True, False, """If True, determine automatically a reused matrix using its SHA1 digest. If False, .clear() has to be called manually whenever the matrix changes - expert use only!"""), ] def __init__(self, conf, method=None, **kwargs): LinearSolver.__init__(self, conf, solve=None, **kwargs) self.sls = None if method is None: method = self.conf.method aux = try_imports(['import scipy.sparse.linalg as sls', 'import scipy.sparse.linalg.dsolve as sls'], 'cannot import scipy sparse direct solvers!') if 'sls' in aux: self.sls = aux['sls'] else: raise ValueError('SuperLU not available!') if method in ['auto', 'umfpack']: aux = try_imports(['import scikits.umfpack as um']) is_umfpack = True if 'um' in aux\ and hasattr(aux['um'], 'UMFPACK_OK') else False if method == 'umfpack' and not is_umfpack: raise ValueError('UMFPACK not available!') elif method == 'superlu': is_umfpack = False else: raise ValueError('uknown solution method! (%s)' % method) if is_umfpack: self.sls.use_solver(useUmfpack=True, assumeSortedIndices=True) else: self.sls.use_solver(useUmfpack=False) self.clear() @standard_call def __call__(self, rhs, x0=None, conf=None, eps_a=None, eps_r=None, i_max=None, mtx=None, status=None, **kwargs): if not conf.use_presolve: self.clear() if conf.use_presolve: self.presolve(mtx, use_mtx_digest=conf.use_mtx_digest) # Matrix is already prefactorized. return self.solve(rhs) else: return self.sls.spsolve(mtx, rhs)
[docs] def clear(self): if self.solve is not None: del self.solve self.solve = None
[docs] def presolve(self, mtx, use_mtx_digest=True): if use_mtx_digest: is_new, mtx_digest = _is_new_matrix(mtx, self.mtx_digest) else: is_new, mtx_digest = False, None if is_new or (self.solve is None): self.solve = self.sls.factorized(mtx.tocsc()) self.mtx_digest = mtx_digest
[docs] class ScipySuperLU(ScipyDirect): """ SuperLU - direct sparse solver from SciPy. """ name = 'ls.scipy_superlu' _parameters = [ ('use_presolve', 'bool', False, False, 'If True, pre-factorize the matrix.'), ('use_mtx_digest', 'bool', True, False, """If True, determine automatically a reused matrix using its SHA1 digest. If False, .clear() has to be called manually whenever the matrix changes - expert use only!"""), ] def __init__(self, conf, **kwargs): ScipyDirect.__init__(self, conf, method='superlu', **kwargs)
[docs] class ScipyUmfpack(ScipyDirect): """ UMFPACK - direct sparse solver from SciPy. """ name = 'ls.scipy_umfpack' _parameters = [ ('use_presolve', 'bool', False, False, 'If True, pre-factorize the matrix.'), ('use_mtx_digest', 'bool', True, False, """If True, determine automatically a reused matrix using its SHA1 digest. If False, .clear() has to be called manually whenever the matrix changes - expert use only!"""), ] def __init__(self, conf, **kwargs): ScipyDirect.__init__(self, conf, method='umfpack', **kwargs)
[docs] class ScipyIterative(LinearSolver): """ Interface to SciPy iterative solvers. The `eps_r` tolerance is both absolute and relative - the solvers stop when either the relative or the absolute residual is below it. """ name = 'ls.scipy_iterative' _parameters = [ ('method', 'str', 'cg', False, 'The actual solver to use.'), ('setup_precond', 'callable', lambda mtx, context: None, False, """User-supplied function for the preconditioner initialization/setup. It is called as setup_precond(mtx, context), where mtx is the matrix, context is a user-supplied context, and should return one of {sparse matrix, dense matrix, LinearOperator}. """), ('callback', 'callable', None, False, """User-supplied function to call after each iteration. It is called as callback(xk), where xk is the current solution vector, except the gmres method, where the argument is the residual. """), ('i_max', 'int', 100, False, 'The maximum number of iterations.'), ('eps_a', 'float', 1e-8, False, 'The absolute tolerance for the residual.'), ('eps_r', 'float', 1e-8, False, 'The relative tolerance for the residual.'), ('*', '*', None, False, 'Additional parameters supported by the method.'), ] # All iterative solvers in scipy.sparse.linalg pass a solution vector into # a callback except those below, that take a residual vector. _callbacks_res = ['gmres'] def __init__(self, conf, context=None, **kwargs): import scipy.sparse.linalg as la LinearSolver.__init__(self, conf, context=context, **kwargs) try: solver = getattr(la, self.conf.method) except AttributeError: output('scipy solver %s does not exist!' % self.conf.method) output('using cg instead') solver = la.cg self.solver = solver self.converged_reasons = { 0 : 'successful exit', 1 : 'number of iterations', -1 : 'illegal input or breakdown', } @standard_call def __call__(self, rhs, x0=None, conf=None, eps_a=None, eps_r=None, i_max=None, mtx=None, status=None, context=None, **kwargs): solver_kwargs = self.build_solver_kwargs(conf) eps_a = get_default(eps_a, self.conf.eps_a) eps_r = get_default(eps_r, self.conf.eps_r) i_max = get_default(i_max, self.conf.i_max) setup_precond = get_default(kwargs.get('setup_precond', None), self.conf.setup_precond) callback = get_default(kwargs.get('callback', lambda sol: None), self.conf.callback) self.iter = 0 def iter_callback(sol): self.iter += 1 msg = '%s: iteration %d' % (self.conf.name, self.iter) if conf.verbose > 2: if conf.method not in self._callbacks_res: res = mtx * sol - rhs else: res = sol rnorm = nm.linalg.norm(res) msg += ': |Ax-b| = %e' % rnorm output(msg, verbose=conf.verbose > 1) # Call an optional user-defined callback. callback(sol) precond = setup_precond(mtx, context) if conf.method == 'qmr': prec_args = {'M1' : precond, 'M2' : precond} else: prec_args = {'M' : precond} solver_kwargs.update(prec_args) if conf.method == 'gmres': from pkg_resources import parse_version import scipy as sp if parse_version(sp.__version__) >= parse_version('1.4.0'): solver_kwargs.update({'callback_type' : 'legacy'}) try: sol, info = self.solver(mtx, rhs, x0=x0, atol=eps_a, tol=eps_r, maxiter=i_max, callback=iter_callback, **solver_kwargs) except TypeError: sol, info = self.solver(mtx, rhs, x0=x0, tol=eps_r, maxiter=i_max, callback=iter_callback, **solver_kwargs) output('%s: %s convergence: %s (%s, %d iterations)' % (self.conf.name, self.conf.method, info, self.converged_reasons[nm.sign(info)], self.iter), verbose=conf.verbose) return sol, self.iter
[docs] class PyAMGSolver(LinearSolver): """ Interface to PyAMG solvers. The `method` parameter can be one of: 'smoothed_aggregation_solver', 'ruge_stuben_solver'. The `accel` parameter specifies the Krylov solver name, that is used as an accelerator for the multigrid solver. """ name = 'ls.pyamg' _parameters = [ ('method', 'str', 'smoothed_aggregation_solver', False, 'The actual solver to use.'), ('accel', 'str', None, False, 'The accelerator.'), ('callback', 'callable', None, False, """User-supplied function to call after each iteration. It is called as callback(xk), where xk is the current solution vector, except the gmres accelerator, where the argument is the residual norm. """), ('i_max', 'int', 100, False, 'The maximum number of iterations.'), ('eps_r', 'float', 1e-8, False, 'The relative tolerance for the residual.'), ('force_reuse', 'bool', False, False, """If True, skip the check whether the MG solver object corresponds to the `mtx` argument: it is always reused."""), ('*', '*', None, False, """Additional parameters supported by the method. Use the 'method:' prefix for arguments of the method construction function (e.g. 'method:max_levels' : 5), and the 'solve:' prefix for the subsequent solver call."""), ] # All iterative solvers in pyamg.krylov pass a solution vector into # a callback except those below, that take a residual vector norm. _callbacks_res = ['gmres'] def __init__(self, conf, **kwargs): try: import pyamg except ImportError: msg = 'cannot import pyamg!' raise ImportError(msg) LinearSolver.__init__(self, conf, mg=None, **kwargs) try: solver = getattr(pyamg, self.conf.method) except AttributeError: output('pyamg.%s does not exist!' % self.conf.method) output('using pyamg.smoothed_aggregation_solver instead') solver = pyamg.smoothed_aggregation_solver self.solver = solver @standard_call def __call__(self, rhs, x0=None, conf=None, eps_a=None, eps_r=None, i_max=None, mtx=None, status=None, **kwargs): solver_kwargs = self.build_solver_kwargs(conf) eps_r = get_default(eps_r, self.conf.eps_r) i_max = get_default(i_max, self.conf.i_max) callback = get_default(kwargs.get('callback', lambda sol: None), self.conf.callback) self.iter = 0 def iter_callback(sol): self.iter += 1 msg = '%s: iteration %d' % (self.conf.name, self.iter) if conf.verbose > 2: if conf.accel not in self._callbacks_res: res = mtx * sol - rhs else: res = sol rnorm = nm.linalg.norm(res) msg += ': |Ax-b| = %e' % rnorm output(msg, verbose=conf.verbose > 1) # Call an optional user-defined callback. callback(sol) is_new, mtx_digest = _is_new_matrix(mtx, self.mtx_digest, force_reuse=conf.force_reuse) if is_new or (self.mg is None): _kwargs = {key[7:] : val for key, val in six.iteritems(solver_kwargs) if key.startswith('method:')} self.mg = self.solver(mtx, **_kwargs) self.mtx_digest = mtx_digest _kwargs = {key[6:] : val for key, val in six.iteritems(solver_kwargs) if key.startswith('solve:')} sol = self.mg.solve(rhs, x0=x0, accel=conf.accel, tol=eps_r, maxiter=i_max, callback=iter_callback, **_kwargs) return sol, self.iter
[docs] class PyAMGKrylovSolver(LinearSolver): """ Interface to PyAMG Krylov solvers. """ name = 'ls.pyamg_krylov' _parameters = [ ('method', 'str', 'cg', False, 'The actual solver to use.'), ('setup_precond', 'callable', lambda mtx, context: None, False, """User-supplied function for the preconditioner initialization/setup. It is called as setup_precond(mtx, context), where mtx is the matrix, context is a user-supplied context, and should return one of {sparse matrix, dense matrix, LinearOperator}. """), ('callback', 'callable', None, False, """User-supplied function to call after each iteration. It is called as callback(xk), where xk is the current solution vector, except the gmres method, where the argument is the residual norm. """), ('i_max', 'int', 100, False, 'The maximum number of iterations.'), ('eps_r', 'float', 1e-8, False, 'The relative tolerance for the residual.'), ('*', '*', None, False, 'Additional parameters supported by the method.'), ] # All iterative solvers in pyamg.krylov pass a solution vector into # a callback except those below, that take a residual vector norm. _callbacks_res = ['gmres'] def __init__(self, conf, context=None, **kwargs): try: import pyamg.krylov as krylov except ImportError: msg = 'cannot import pyamg.krylov!' raise ImportError(msg) LinearSolver.__init__(self, conf, mg=None, context=context, **kwargs) try: solver = getattr(krylov, self.conf.method) except AttributeError: output('pyamg.krylov.%s does not exist!' % self.conf.method) raise self.solver = solver self.converged_reasons = { 0 : 'successful exit', 1 : 'number of iterations', -1 : 'illegal input or breakdown', } @standard_call def __call__(self, rhs, x0=None, conf=None, eps_a=None, eps_r=None, i_max=None, mtx=None, status=None, context=None, **kwargs): solver_kwargs = self.build_solver_kwargs(conf) eps_r = get_default(eps_r, self.conf.eps_r) i_max = get_default(i_max, self.conf.i_max) setup_precond = get_default(kwargs.get('setup_precond', None), self.conf.setup_precond) callback = get_default(kwargs.get('callback', lambda sol: None), self.conf.callback) self.iter = 0 def iter_callback(sol): self.iter += 1 msg = '%s: iteration %d' % (self.conf.name, self.iter) if conf.verbose > 2: if conf.method not in self._callbacks_res: res = mtx * sol - rhs else: res = sol rnorm = nm.linalg.norm(res) msg += ': |Ax-b| = %e' % rnorm output(msg, verbose=conf.verbose > 1) # Call an optional user-defined callback. callback(sol) precond = setup_precond(mtx, context) sol, info = self.solver(mtx, rhs, x0=x0, tol=eps_r, maxiter=i_max, M=precond, callback=iter_callback, **solver_kwargs) output('%s: %s convergence: %s (%s, %d iterations)' % (self.conf.name, self.conf.method, info, self.converged_reasons[nm.sign(info)], self.iter), verbose=conf.verbose) return sol, self.iter
[docs] class PETScKrylovSolver(LinearSolver): """ PETSc Krylov subspace solver. The solver supports parallel use with a given MPI communicator (see `comm` argument of :func:`PETScKrylovSolver.__init__()`) and allows passing in PETSc matrices and vectors. Returns a (global) PETSc solution vector instead of a (local) numpy array, when given a PETSc right-hand side vector. The solver and preconditioner types are set upon the solver object creation. Tolerances can be overridden when called by passing a `conf` object. Convergence is reached when `rnorm < max(eps_r * rnorm_0, eps_a)`, where, in PETSc, `rnorm` is by default the norm of *preconditioned* residual. """ name = 'ls.petsc' _parameters = [ ('method', 'str', 'cg', False, 'The actual solver to use.'), ('setup_precond', 'callable', None, False, """User-supplied function for the preconditioner initialization/setup. It is called as setup_precond(mtx, context), where mtx is the matrix, context is a user-supplied context, and should return an object with `setUp(self, pc)` and `apply(self, pc, x, y)` methods. Has precedence over the `precond`/`sub_precond` parameters. """), ('precond', 'str', 'icc', False, 'The preconditioner.'), ('sub_precond', 'str', 'none', False, 'The preconditioner for matrix blocks (in parallel runs).'), ('precond_side', "{'left', 'right', 'symmetric', None}", None, False, 'The preconditioner side.'), ('i_max', 'int', 100, False, 'The maximum number of iterations.'), ('eps_a', 'float', 1e-8, False, 'The absolute tolerance for the residual.'), ('eps_r', 'float', 1e-8, False, 'The relative tolerance for the residual.'), ('eps_d', 'float', 1e5, False, 'The divergence tolerance for the residual.'), ('force_reuse', 'bool', False, False, """If True, skip the check whether the KSP solver object corresponds to the `mtx` argument: it is always reused."""), ('*', '*', None, False, """Additional parameters supported by the method. Can be used to pass all PETSc options supported by :func:`petsc.Options()`."""), ] _precond_sides = {None : None, 'left' : 0, 'right' : 1, 'symmetric' : 2} def __init__(self, conf, comm=None, context=None, **kwargs): if comm is None: from sfepy.parallel.parallel import init_petsc_args; init_petsc_args from petsc4py import PETSc as petsc converged_reasons = {} for key, val in six.iteritems(petsc.KSP.ConvergedReason.__dict__): if isinstance(val, int): converged_reasons[val] = key LinearSolver.__init__(self, conf, petsc=petsc, comm=comm, converged_reasons=converged_reasons, fields=None, ksp=None, pmtx=None, context=context, **kwargs)
[docs] def set_field_split(self, field_ranges, comm=None): """ Setup local PETSc ranges for fields to be used with 'fieldsplit' preconditioner. This function must be called before solving the linear system. """ comm = get_default(comm, self.comm) self.fields = [] for key, rng in six.iteritems(field_ranges): if isinstance(rng, slice): rng = rng.start, rng.stop size = rng[1] - rng[0] field_is = self.petsc.IS().createStride(size, first=rng[0], step=1, comm=comm) self.fields.append((key, field_is))
[docs] def create_ksp(self, options=None, comm=None): optDB = self.petsc.Options() optDB['sub_pc_type'] = self.conf.sub_precond if options is not None: for key, val in six.iteritems(options): optDB[key] = val ksp = self.petsc.KSP() ksp.create(comm) ksp.setType(self.conf.method) pc = ksp.getPC() if self.conf.setup_precond is None: pc.setType(self.conf.precond) else: pc.setType(pc.Type.PYTHON) ksp.setFromOptions() if (pc.type == 'fieldsplit'): if self.fields is not None: pc.setFieldSplitIS(*self.fields) else: msg = 'PETScKrylovSolver.set_field_split() has to be called!' raise ValueError(msg) side = self._precond_sides[self.conf.precond_side] if side is not None: ksp.setPCSide(side) return ksp
[docs] def create_petsc_matrix(self, mtx, comm=None): if isinstance(mtx, self.petsc.Mat): pmtx = mtx else: mtx = sps.csr_matrix(mtx) pmtx = self.petsc.Mat() pmtx.createAIJ(mtx.shape, csr=(mtx.indptr, mtx.indices, mtx.data), comm=comm) return pmtx
@petsc_call def __call__(self, rhs, x0=None, conf=None, eps_a=None, eps_r=None, i_max=None, mtx=None, status=None, comm=None, context=None, **kwargs): solver_kwargs = self.build_solver_kwargs(conf) eps_a = get_default(eps_a, self.conf.eps_a) eps_r = get_default(eps_r, self.conf.eps_r) i_max = get_default(i_max, self.conf.i_max) eps_d = self.conf.eps_d is_new, mtx_digest = _is_new_matrix(mtx, self.mtx_digest, force_reuse=conf.force_reuse) if (not is_new) and self.ksp is not None: ksp = self.ksp pmtx = self.pmtx else: pmtx = self.create_petsc_matrix(mtx, comm=comm) ksp = self.create_ksp(options=solver_kwargs, comm=comm) ksp.setOperators(pmtx) ksp.setTolerances(atol=eps_a, rtol=eps_r, divtol=eps_d, max_it=i_max) setup_precond = self.conf.setup_precond if setup_precond is not None: ksp.pc.setPythonContext(setup_precond(mtx, context)) ksp.setFromOptions() self.mtx_digest = mtx_digest self.ksp = ksp self.pmtx = pmtx if isinstance(rhs, self.petsc.Vec): prhs = rhs else: prhs = pmtx.getVecLeft() prhs[...] = rhs if x0 is not None: if isinstance(x0, self.petsc.Vec): psol = x0 else: psol = pmtx.getVecRight() psol[...] = x0 ksp.setInitialGuessNonzero(True) else: psol = pmtx.getVecRight() ksp.setInitialGuessNonzero(False) ksp.solve(prhs, psol) output('%s(%s, %s/proc) convergence: %s (%s, %d iterations)' % (ksp.getType(), ksp.getPC().getType(), self.conf.sub_precond, ksp.reason, self.converged_reasons[ksp.reason], ksp.getIterationNumber()), verbose=conf.verbose) if isinstance(rhs, self.petsc.Vec): sol = psol else: sol = psol[...].copy() return sol
[docs] class MUMPSSolver(LinearSolver): """ Interface to MUMPS solver. """ name = 'ls.mumps' _parameters = [ ('use_presolve', 'bool', False, False, 'If True, pre-factorize the matrix.'), ('use_mtx_digest', 'bool', True, False, """If True, determine automatically a reused matrix using its SHA1 digest. If False, .clear() has to be called manually whenever the matrix changes - expert use only!"""), ('memory_relaxation', 'int', 20, False, 'The percentage increase in the estimated working space.'), ] def __init__(self, conf, **kwargs): import sfepy.solvers.ls_mumps as mumps LinearSolver.__init__(self, conf, mumps=mumps, mumps_ls=None, mumps_presolved=False, **kwargs) mumps.load_mumps_libraries() # try to load MUMPS libraries if not mumps.use_mpi: raise AttributeError('No mpi4py found! Required by MUMPS solver.') self.clear() @standard_call def __call__(self, rhs, x0=None, conf=None, eps_a=None, eps_r=None, i_max=None, mtx=None, status=None, **kwargs): if not conf.use_presolve: self.clear() self.presolve(mtx, use_mtx_digest=conf.use_mtx_digest) out = rhs.copy(order='F') self.mumps_ls.set_rhs(out) self.mumps_ls(3) # solve return out
[docs] def clear(self): if self.mumps_ls is not None: del(self.mumps_ls) self.mumps_ls = None
[docs] def presolve(self, mtx, use_mtx_digest=True): if use_mtx_digest: is_new, mtx_digest = _is_new_matrix(mtx, self.mtx_digest) else: is_new, mtx_digest = False, None if is_new or (self.mumps_ls is None): if not isinstance(mtx, sps.coo_matrix): mtx = mtx.tocoo() if self.mumps_ls is None: system = 'complex' if mtx.dtype.name.startswith('complex')\ else 'real' is_sym = self.mumps.coo_is_symmetric(mtx) mem_relax = self.conf.memory_relaxation self.mumps_ls = self.mumps.MumpsSolver(system=system, is_sym=is_sym, mem_relax=mem_relax) if self.conf.verbose: self.mumps_ls.set_verbose() self.mumps_ls.set_mtx_centralized(mtx) self.mumps_ls(4) # analyze + factorize self.mtx_digest = mtx_digest
def __del__(self): self.clear()
[docs] class MUMPSParallelSolver(LinearSolver): """ Interface to MUMPS parallel solver. """ name = 'ls.mumps_par' _parameters = [ ('memory_relaxation', 'int', 20, False, 'The percentage increase in the estimated working space.'), ] def __init__(self, conf, **kwargs): import multiprocessing import sfepy.solvers.ls_mumps as mumps mumps.load_mumps_libraries() # try to load MUMPS libraries LinearSolver.__init__(self, conf, mumps=mumps, mumps_ls=None, number_of_cpu=multiprocessing.cpu_count(), mumps_presolved=False, **kwargs) @standard_call def __call__(self, rhs, x0=None, conf=None, eps_a=None, eps_r=None, i_max=None, mtx=None, status=None, **kwargs): from mpi4py import MPI import sys from sfepy import data_dir import os.path as op from tempfile import gettempdir def tmpfile(fname): return op.join(gettempdir(), fname) if not isinstance(mtx, sps.coo_matrix): mtx = mtx.tocoo() is_sym = self.mumps.coo_is_symmetric(mtx) rr, cc, data = mtx.row + 1, mtx.col + 1, mtx.data if is_sym: idxs = nm.where(cc >= rr)[0] # upper triangular matrix rr, cc, data = rr[idxs], cc[idxs], data[idxs] n = mtx.shape[0] nz = rr.shape[0] flags = nm.memmap(tmpfile('vals_flags.array'), dtype='int32', mode='w+', shape=(4,)) flags[0] = n flags[1] = 1 if data.dtype.name.startswith('complex') else 0 flags[2] = int(is_sym) flags[3] = int(self.conf.verbose) idxs = nm.memmap(tmpfile('idxs.array'), dtype='int32', mode='w+', shape=(2, nz)) idxs[0, :] = rr idxs[1, :] = cc dtype = {0: 'float64', 1: 'complex128'}[flags[1]] vals_mtx = nm.memmap(tmpfile('vals_mtx.array'), dtype=dtype, mode='w+', shape=(nz,)) vals_rhs = nm.memmap(tmpfile('vals_rhs.array'), dtype=dtype, mode='w+', shape=(n,)) vals_mtx[:] = data vals_rhs[:] = rhs mumps_call = op.join(data_dir, 'sfepy', 'solvers', 'ls_mumps_parallel.py') comm = MPI.COMM_SELF.Spawn(sys.executable, args=[mumps_call], maxprocs=self.number_of_cpu) comm.Disconnect() out = nm.memmap(tmpfile('vals_x.array'), dtype=dtype, mode='r') return out
[docs] class SchurMumps(MUMPSSolver): r""" Mumps Schur complement solver. """ name = 'ls.schur_mumps' _parameters = MUMPSSolver._parameters + [ ('schur_variables', 'list', None, True, 'The list of Schur variables.'), ] @standard_call def __call__(self, rhs, x0=None, conf=None, eps_a=None, eps_r=None, i_max=None, mtx=None, status=None, **kwargs): import scipy.linalg as sla if not isinstance(mtx, sps.coo_matrix): mtx = mtx.tocoo() system = 'complex' if mtx.dtype.name.startswith('complex') else 'real' self.mumps_ls = self.mumps.MumpsSolver(system=system) if self.conf.verbose: self.mumps_ls.set_verbose() schur_list = [] for schur_var in conf.schur_variables: slc = self.context.equations.variables.adi.indx[schur_var] schur_list.append(nm.arange(slc.start, slc.stop, slc.step, dtype='i')) self.mumps_ls.set_mtx_centralized(mtx) out = rhs.copy(order='F') self.mumps_ls.set_rhs(out) S, y2 = self.mumps_ls.get_schur(nm.hstack(schur_list)) x2 = sla.solve(S.T, y2) # solve the dense Schur system using scipy.linalg return self.mumps_ls.expand_schur(x2)
[docs] class MultiProblem(ScipyDirect): r""" Conjugate multiple problems. Allows to define conjugate multiple problems. """ name = 'ls.cm_pb' _parameters = ScipyDirect._parameters + [ ('others', 'list', None, True, 'The list of auxiliary problem definition files.'), ('coupling_variables', 'list', None, True, 'The list of coupling variables.'), ] def __init__(self, conf, context=None, **kwargs): ScipyDirect.__init__(self, conf, context=context, **kwargs)
[docs] def init_subproblems(self, conf, **kwargs): from sfepy.discrete import Problem from sfepy.base.conf import ProblemConf, get_standard_keywords from scipy.spatial import cKDTree as KDTree # init subproblems problem = self.context pb_vars = problem.get_variables() # get "master" DofInfo and last index pb_adi_indx = problem.equations.variables.adi.indx self.adi_indx = pb_adi_indx.copy() last_indx = -1 for ii in six.itervalues(self.adi_indx): last_indx = nm.max([last_indx, ii.stop]) # coupling variables self.cvars_to_pb = {} for jj in conf.coupling_variables: self.cvars_to_pb[jj] = [None, None] if jj in pb_vars.names: if pb_vars[jj].dual_var_name is not None: self.cvars_to_pb[jj][0] = -1 else: self.cvars_to_pb[jj][1] = -1 # init subproblems self.subpb = [] required, other = get_standard_keywords() master_prefix = output.get_output_prefix() for ii, ifname in enumerate(conf.others): sub_prefix = master_prefix[:-1] + '-sub%d:' % (ii + 1) output.set_output_prefix(sub_prefix) kwargs['master_problem'] = problem confi = ProblemConf.from_file(ifname, required, other, define_args=kwargs) pbi = Problem.from_conf(confi, init_equations=True) pbi.setup_default_output(conf=confi) pbi.set_output_dir(problem.output_dir) pbi.equations.set_data(None, ignore_unknown=True) pbi.time_update() pbi.update_materials() sti = pbi.get_initial_state() sti.apply_ebc() pbi_vars = pbi.get_variables() output.set_output_prefix(master_prefix) self.subpb.append([pbi, sti, None]) # append "slave" DofInfo for jj in pbi_vars.names: if not(pbi_vars[jj].is_state()): continue didx = pbi.equations.variables.adi.indx[jj] ndof = didx.stop - didx.start if jj in self.adi_indx: if ndof != \ (self.adi_indx[jj].stop - self.adi_indx[jj].start): raise ValueError('DOFs do not match!') else: self.adi_indx.update({ jj: slice(last_indx, last_indx + ndof, None)}) last_indx += ndof for jj in conf.coupling_variables: if jj in pbi_vars.names: if pbi_vars[jj].dual_var_name is not None: self.cvars_to_pb[jj][0] = ii else: self.cvars_to_pb[jj][1] = ii self.subpb.append([problem, None, None]) self.cvars_to_pb_map = {} for varname, pbs in six.iteritems(self.cvars_to_pb): # match field nodes coors = [] for ii in pbs: pbi = self.subpb[ii][0] pbi_vars = pbi.get_variables() fcoors = pbi_vars[varname].field.coors dc = nm.abs(nm.max(fcoors, axis=0)\ - nm.min(fcoors, axis=0)) ax = nm.where(dc > 1e-9)[0] coors.append(fcoors[:,ax]) if len(coors[0]) != len(coors[1]): raise ValueError('number of nodes does not match!') kdtree = KDTree(coors[0]) map_12 = kdtree.query(coors[1])[1] pbi1 = self.subpb[pbs[0]][0] pbi1_vars = pbi1.get_variables() eq_map_1 = pbi1_vars[varname].eq_map pbi2 = self.subpb[pbs[1]][0] pbi2_vars = pbi2.get_variables() eq_map_2 = pbi2_vars[varname].eq_map dpn = eq_map_2.dpn nnd = map_12.shape[0] map_12_nd = nm.zeros((nnd * dpn,), dtype=nm.int32) if dpn > 1: for ii in range(dpn): map_12_nd[ii::dpn] = map_12 * dpn + ii else: map_12_nd = map_12 idx = nm.where(eq_map_2.eq >= 0)[0] self.cvars_to_pb_map[varname] = eq_map_1.eq[map_12[idx]]
[docs] def sparse_submat(self, Ad, Ar, Ac, gr, gc, S): """ A[gr,gc] = S """ if type(gr) is slice: gr = nm.arange(gr.start, gr.stop) if type(gc) is slice: gc = nm.arange(gc.start, gc.stop) for ii, lrow in enumerate(S): m = lrow.indices.shape[0] idxrow = nm.ones((m,), dtype=nm.int32) * gr[ii] Ar = nm.hstack([Ar, idxrow]) Ac = nm.hstack([Ac, gc[lrow.indices]]) Ad = nm.hstack([Ad, lrow.data]) return Ad, Ar, Ac
@standard_call def __call__(self, rhs, x0=None, conf=None, eps_a=None, eps_r=None, i_max=None, mtx=None, status=None, **kwargs): self.init_subproblems(self.conf, **kwargs) max_indx = 0 hst = nm.hstack for ii in six.itervalues(self.adi_indx): max_indx = nm.max([max_indx, ii.stop]) new_rhs = nm.zeros((max_indx,), dtype=rhs.dtype) new_rhs[:rhs.shape[0]] = rhs # copy "master" matrices pbi = self.subpb[-1][0] adi_indxi = pbi.equations.variables.adi.indx mtxc = mtx.tocsc() aux_data = nm.array([], dtype=mtxc.dtype) aux_rows = nm.array([], dtype=nm.int32) aux_cols = nm.array([], dtype=nm.int32) for jk, jv in six.iteritems(adi_indxi): if jk in self.cvars_to_pb: if not(self.cvars_to_pb[jk][0] == -1): continue gjv = self.adi_indx[jk] ii = gjv.start for jj in nm.arange(jv.start, jv.stop): ptr = mtxc.indptr[jj] nn = mtxc.indptr[jj + 1] - ptr sl = slice(ptr, ptr + nn, None) aux_data = hst([aux_data, mtxc.data[sl]]) aux_rows = hst([aux_rows, mtxc.indices[sl]]) aux_cols = hst([aux_cols, nm.ones((nn,), dtype=nm.int32) * ii]) ii += 1 # copy "slave" (sub)matricies mtxs = [] for kk, (pbi, sti0, _) in enumerate(self.subpb[:-1]): x0i = sti0.get_state(pbi.active_only) evi = pbi.get_evaluator() mtxi = evi.eval_tangent_matrix(x0i, mtx=pbi.mtx_a) rhsi = evi.eval_residual(x0i) mtxs.append(mtxi) adi_indxi = pbi.equations.variables.adi.indx for ik, iv in six.iteritems(adi_indxi): if ik in self.cvars_to_pb: if not(self.cvars_to_pb[ik][0] == kk): continue giv = self.adi_indx[ik] for jk, jv in six.iteritems(adi_indxi): gjv = self.adi_indx[jk] if jk in self.cvars_to_pb: if not(self.cvars_to_pb[jk][0] == kk): continue aux_data, aux_rows, aux_cols =\ self.sparse_submat(aux_data, aux_rows, aux_cols, giv, gjv, mtxi[iv, jv]) new_rhs[giv] = rhsi[iv] mtxs.append(mtx) # copy "coupling" (sub)matricies for varname, pbs in six.iteritems(self.cvars_to_pb): idx = pbs[1] pbi = self.subpb[idx][0] mtxi = mtxs[idx] gjv = self.adi_indx[varname] jv = pbi.equations.variables.adi.indx[varname] adi_indxi = pbi.equations.variables.adi.indx for ik, iv in six.iteritems(adi_indxi): if ik == varname: continue giv = self.adi_indx[ik] aux_mtx = mtxi[iv,:].tocsc() for ll, jj in enumerate(nm.arange(jv.start, jv.stop)): ptr = aux_mtx.indptr[jj] nn = aux_mtx.indptr[jj + 1] - ptr if nn < 1: continue sl = slice(ptr, ptr + nn, None) aux_data = hst([aux_data, aux_mtx.data[sl]]) aux_rows = hst([aux_rows, aux_mtx.indices[sl] + giv.start]) jjr = gjv.start + self.cvars_to_pb_map[varname][ll] aux_cols = hst([aux_cols, nm.ones((nn,), dtype=nm.int32) * jjr]) # create new matrix new_mtx = sps.coo_matrix((aux_data, (aux_rows, aux_cols))).tocsr() res0 = ScipyDirect.__call__(self, new_rhs, mtx=new_mtx) res = [] for kk, (pbi, sti0, _) in enumerate(self.subpb): adi_indxi = pbi.equations.variables.adi.indx max_indx = 0 for ii in six.itervalues(adi_indxi): max_indx = nm.max([max_indx, ii.stop]) resi = nm.zeros((max_indx,), dtype=res0.dtype) for ik, iv in six.iteritems(adi_indxi): giv = self.adi_indx[ik] if ik in self.cvars_to_pb: if pbi is self.subpb[self.cvars_to_pb[ik][1]][0]: giv = self.cvars_to_pb_map[ik] + giv.start resi[iv] = res0[giv] if sti0 is not None: sti = sti0.copy() sti.set_state(-resi, pbi.active_only) pbi.save_state(pbi.get_output_name(), sti) self.subpb[kk][-1] = sti res.append(resi) return res[-1]
[docs] class RMMSolver(LinearSolver): """ Special solver for explicit transient elastodynamics. The solver uses the reciprocal mass matrix algorithm [1]_, [2]_ to directly construct a sparse inverse mass matrix. Instead of solving a linear system, calling the solver simply performs a sparse matrix multiplication. Limitations: - Assumes that the density is constant in time. - Uses the direct EBC application, i.e., no EBC projection matrix. .. [1] González, J.A., Kolman, R., Cho, S.S., Felippa, C.A., Park, K.C., 2018. Inverse mass matrix via the method of localized Lagrange multipliers. International Journal for Numerical Methods in Engineering 113, 277–295. https://doi.org/10.1002/nme.5613 .. [2] González, J.A., Kopačka, J., Kolman, R., Cho, S.S., Park, K.C., 2019. Inverse mass matrix for isogeometric explicit transient analysis via the method of localized Lagrange multipliers. International Journal for Numerical Methods in Engineering 117, 939–966. https://doi.org/10.1002/nme.5986 """ name = 'ls.rmm' _parameters = [ ('rmm_term', 'str', None, True, """The RMM term definition, see :class:`MassTerm <sfepy.terms.terms_mass.MassTerm>`."""), ('debug', 'bool', False, False, 'If True, run in debug mode.'), ] def __init__(self, conf, context=None, **kwargs): LinearSolver.__init__(self, conf, context=context, mtx_im=None, a0=None, **kwargs)
[docs] def init_rmm(self, mtx): from sfepy.discrete.evaluate import eval_equations, apply_ebc_to_matrix problem = self.context equations, variables = problem.create_evaluable( self.conf.rmm_term, preserve_caches=True, copy_materials=False, mode='weak', active_only=problem.active_only, ) vu = next(variables.iter_state()) mtx_a = eval_equations(equations, variables, preserve_caches=True, mode='weak', dw_mode='matrix', term_mode='DPM', active_only=problem.active_only) if not problem.active_only: apply_ebc_to_matrix(mtx_a, vu.eq_map.eq_ebc, (vu.eq_map.master, vu.eq_map.slave)) mtx_a.eliminate_zeros() mtx_ia = mtx_a.copy() mtx_ia.setdiag(1.0 / mtx_a.diagonal()) mtx_c = eval_equations(equations, variables, preserve_caches=True, mode='weak', dw_mode='matrix', term_mode='RMM', active_only=problem.active_only) mtx_c.eliminate_zeros() mtx_im = mtx_ia @ (mtx_c @ mtx_ia) if not problem.active_only: apply_ebc_to_matrix(mtx_im, vu.eq_map.eq_ebc, (vu.eq_map.master, vu.eq_map.slave)) if self.conf.debug: mtx_m = eval_equations( equations, variables, preserve_caches=True, mode='weak', dw_mode='matrix', term_mode=None, active_only=problem.active_only, ) if not problem.active_only: mtx_r = vu.eq_map.get_operator() mtx_imr = mtx_r.T @ mtx_im @ mtx_r mtx_mr = mtx_r.T @ mtx_m @ mtx_r mtx_mor = mtx_r.T @ mtx @ mtx_r else: mtx_imr = mtx_im mtx_mr = mtx_m mtx_mor = mtx dim = problem.domain.shape.dim output('total mass check: AMM:', mtx_mr.sum() / dim, 'RMM:', nm.linalg.inv(mtx_imr.toarray()).sum() / dim, 'M:', mtx_mor.sum() / dim) return mtx_im
@standard_call def __call__(self, rhs, x0=None, conf=None, eps_a=None, eps_r=None, i_max=None, mtx=None, status=None, **kwargs): if self.mtx_im is None: self.mtx_im = self.init_rmm(mtx) sol = self.mtx_im @ rhs if self.a0 is not None: # To make RMMSolver work with the standard Newton solver, a0 has to # be set to the previous acceleration and M term has to be # nullified (use dw_zero). This option is not used in the current # implementation of CentralDifferenceTS. sol += self.a0 return sol