Actual source code: ex15f.F90
1: !
2: ! Solves a linear system in parallel with KSP. Also indicates
3: ! use of a user-provided preconditioner. Input parameters include:
4: ! -user_defined_pc : Activate a user-defined preconditioner
5: !
6: ! -------------------------------------------------------------------------
7: !
8: ! Module contains diag needed by shell preconditioner
9: !
10: #include <petsc/finclude/petscksp.h>
11: module ex15fmodule
12: use petscksp
13: implicit none
14: Vec diag
16: contains
18: !/***********************************************************************/
19: !/* Routines for a user-defined shell preconditioner */
20: !/***********************************************************************/
22: !
23: ! SampleShellPCSetUp - This routine sets up a user-defined
24: ! preconditioner context.
25: !
26: ! Input Parameters:
27: ! pc - preconditioner object
28: !
29: ! Output Parameter:
30: ! ierr - error code (nonzero if error has been detected)
31: !
32: ! Notes:
33: ! In this example, we define the shell preconditioner to be Jacobi
34: ! method. Thus, here we create a work vector for storing the reciprocal
35: ! of the diagonal of the matrix; this vector is then
36: ! used within the routine SampleShellPCApply().
37: !
38: subroutine SampleShellPCSetUp(pc, ierr)
40: PC pc
41: Mat pmat
42: PetscErrorCode, intent(out) :: ierr
44: PetscCallA(PCGetOperators(pc, PETSC_NULL_MAT, pmat, ierr))
45: PetscCallA(MatCreateVecs(pmat, diag, PETSC_NULL_VEC, ierr))
46: PetscCallA(MatGetDiagonal(pmat, diag, ierr))
47: PetscCallA(VecReciprocal(diag, ierr))
49: end
51: ! -------------------------------------------------------------------
52: !
53: ! SampleShellPCApply - This routine demonstrates the use of a
54: ! user-provided preconditioner.
55: !
56: ! Input Parameters:
57: ! pc - preconditioner object
58: ! x - input vector
59: !
60: ! Output Parameters:
61: ! y - preconditioned vector
62: ! ierr - error code (nonzero if error has been detected)
63: !
64: ! Notes:
65: ! This code implements the Jacobi preconditioner, merely as an
66: ! example of working with a PCSHELL. Note that the Jacobi method
67: ! is already provided within PETSc.
68: !
69: subroutine SampleShellPCApply(pc, x, y, ierr)
70: PC pc
71: Vec x, y
72: PetscErrorCode, intent(out) :: ierr
74: PetscCallA(VecPointwiseMult(y, x, diag, ierr))
76: end
78: !/***********************************************************************/
79: !/* Routines for a user-defined shell preconditioner */
80: !/***********************************************************************/
82: !
83: ! SampleShellPCDestroy - This routine destroys (frees the memory of) any
84: ! objects we made for the preconditioner
85: !
86: ! Input Parameters:
87: ! pc - for this example we use the actual PC as our shell context
88: !
89: ! Output Parameter:
90: ! ierr - error code (nonzero if error has been detected)
91: !
93: subroutine SampleShellPCDestroy(pc, ierr)
94: PC pc
95: PetscErrorCode, intent(out) :: ierr
97: ! Normally we would recommend storing all the work data (like diag) in
98: ! the context set with PCShellSetContext()
100: PetscCallA(VecDestroy(diag, ierr))
102: end
104: end module
106: program main
107: use ex15fmodule
108: implicit none
110: ! - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
111: ! Variable declarations
112: ! - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
113: !
114: ! Variables:
115: ! ksp - linear solver context
116: ! ksp - Krylov subspace method context
117: ! pc - preconditioner context
118: ! x, b, u - approx solution, right-hand side, exact solution vectors
119: ! A - matrix that defines linear system
120: ! its - iterations for convergence
121: ! norm - norm of solution error
123: Vec x, b, u
124: Mat A
125: PC pc
126: KSP ksp
127: PetscScalar v
128: PetscScalar, parameter :: one = 1.0, neg_one = -1.0
129: PetscReal norm
130: PetscReal, parameter :: tol = 1.e-7
131: PetscErrorCode ierr
132: PetscInt i, j, II, JJ, Istart, Iend, its
133: PetscInt m, n
134: PetscMPIInt rank
135: PetscBool user_defined_pc, flg
137: ! - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
138: ! Beginning of program
139: ! - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
141: PetscCallA(PetscInitialize(ierr))
143: m = 8
144: PetscCallA(PetscOptionsGetInt(PETSC_NULL_OPTIONS, PETSC_NULL_CHARACTER, '-m', m, flg, ierr))
145: n = 7
146: PetscCallA(PetscOptionsGetInt(PETSC_NULL_OPTIONS, PETSC_NULL_CHARACTER, '-n', n, flg, ierr))
147: PetscCallMPIA(MPI_Comm_rank(PETSC_COMM_WORLD, rank, ierr))
149: ! - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
150: ! Compute the matrix and right-hand-side vector that define
151: ! the linear system, Ax = b.
152: ! - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
154: ! Create parallel matrix, specifying only its global dimensions.
155: ! When using MatCreate(), the matrix format can be specified at
156: ! runtime. Also, the parallel partitioning of the matrix is
157: ! determined by PETSc at runtime.
159: PetscCallA(MatCreate(PETSC_COMM_WORLD, A, ierr))
160: PetscCallA(MatSetSizes(A, PETSC_DECIDE, PETSC_DECIDE, m*n, m*n, ierr))
161: PetscCallA(MatSetType(A, MATAIJ, ierr))
162: PetscCallA(MatSetFromOptions(A, ierr))
163: PetscCallA(MatMPIAIJSetPreallocation(A, 5_PETSC_INT_KIND, PETSC_NULL_INTEGER_ARRAY, 5_PETSC_INT_KIND, PETSC_NULL_INTEGER_ARRAY, ierr))
164: PetscCallA(MatSeqAIJSetPreallocation(A, 5_PETSC_INT_KIND, PETSC_NULL_INTEGER_ARRAY, ierr))
166: ! Currently, all PETSc parallel matrix formats are partitioned by
167: ! contiguous chunks of rows across the processors. Determine which
168: ! rows of the matrix are locally owned.
170: PetscCallA(MatGetOwnershipRange(A, Istart, Iend, ierr))
172: ! Set matrix elements for the 2-D, five-point stencil in parallel.
173: ! - Each processor needs to insert only elements that it owns
174: ! locally (but any non-local elements will be sent to the
175: ! appropriate processor during matrix assembly).
176: ! - Always specify global row and columns of matrix entries.
177: ! - Note that MatSetValues() uses 0-based row and column numbers
178: ! in Fortran as well as in C.
180: do II = Istart, Iend - 1
181: v = -1.0
182: i = II/n
183: j = II - i*n
184: if (i > 0) then
185: JJ = II - n
186: PetscCallA(MatSetValues(A, 1_PETSC_INT_KIND, [II], 1_PETSC_INT_KIND, [JJ], [v], ADD_VALUES, ierr))
187: end if
188: if (i < m - 1) then
189: JJ = II + n
190: PetscCallA(MatSetValues(A, 1_PETSC_INT_KIND, [II], 1_PETSC_INT_KIND, [JJ], [v], ADD_VALUES, ierr))
191: end if
192: if (j > 0) then
193: JJ = II - 1
194: PetscCallA(MatSetValues(A, 1_PETSC_INT_KIND, [II], 1_PETSC_INT_KIND, [JJ], [v], ADD_VALUES, ierr))
195: end if
196: if (j < n - 1) then
197: JJ = II + 1
198: PetscCallA(MatSetValues(A, 1_PETSC_INT_KIND, [II], 1_PETSC_INT_KIND, [JJ], [v], ADD_VALUES, ierr))
199: end if
200: v = 4.0
201: PetscCallA(MatSetValues(A, 1_PETSC_INT_KIND, [II], 1_PETSC_INT_KIND, [II], [v], ADD_VALUES, ierr))
202: end do
204: ! Assemble matrix, using the 2-step process:
205: ! MatAssemblyBegin(), MatAssemblyEnd()
206: ! Computations can be done while messages are in transition,
207: ! by placing code between these two statements.
209: PetscCallA(MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY, ierr))
210: PetscCallA(MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY, ierr))
212: ! Create parallel vectors.
213: ! - Here, the parallel partitioning of the vector is determined by
214: ! PETSc at runtime. We could also specify the local dimensions
215: ! if desired -- or use the more general routine VecCreate().
216: ! - When solving a linear system, the vectors and matrices MUST
217: ! be partitioned accordingly. PETSc automatically generates
218: ! appropriately partitioned matrices and vectors when MatCreate()
219: ! and VecCreate() are used with the same communicator.
220: ! - Note: We form 1 vector from scratch and then duplicate as needed.
222: PetscCallA(VecCreateFromOptions(PETSC_COMM_WORLD, PETSC_NULL_CHARACTER, 1_PETSC_INT_KIND, PETSC_DECIDE, m*n, u, ierr))
223: PetscCallA(VecDuplicate(u, b, ierr))
224: PetscCallA(VecDuplicate(b, x, ierr))
226: ! Set exact solution; then compute right-hand-side vector.
228: PetscCallA(VecSet(u, one, ierr))
229: PetscCallA(MatMult(A, u, b, ierr))
231: ! - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
232: ! Create the linear solver and set various options
233: ! - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
235: ! Create linear solver context
237: PetscCallA(KSPCreate(PETSC_COMM_WORLD, ksp, ierr))
239: ! Set operators. Here the matrix that defines the linear system
240: ! also serves as the matrix from which the preconditioner is constructed.
242: PetscCallA(KSPSetOperators(ksp, A, A, ierr))
244: ! Set linear solver defaults for this problem (optional).
245: ! - By extracting the KSP and PC contexts from the KSP context,
246: ! we can then directly call any KSP and PC routines
247: ! to set various options.
249: PetscCallA(KSPGetPC(ksp, pc, ierr))
250: PetscCallA(KSPSetTolerances(ksp, tol, PETSC_CURRENT_REAL, PETSC_CURRENT_REAL, PETSC_CURRENT_INTEGER, ierr))
252: !
253: ! Set a user-defined shell preconditioner if desired
254: !
255: PetscCallA(PetscOptionsHasName(PETSC_NULL_OPTIONS, PETSC_NULL_CHARACTER, '-user_defined_pc', user_defined_pc, ierr))
257: if (user_defined_pc) then
259: ! (Required) Indicate to PETSc that we are using a shell preconditioner
260: PetscCallA(PCSetType(pc, PCSHELL, ierr))
262: ! (Required) Set the user-defined routine for applying the preconditioner
263: PetscCallA(PCShellSetApply(pc, SampleShellPCApply, ierr))
265: ! (Optional) Do any setup required for the preconditioner
266: PetscCallA(PCShellSetSetUp(pc, SampleShellPCSetUp, ierr))
268: ! (Optional) Frees any objects we created for the preconditioner
269: PetscCallA(PCShellSetDestroy(pc, SampleShellPCDestroy, ierr))
271: else
272: PetscCallA(PCSetType(pc, PCJACOBI, ierr))
273: end if
275: ! Set runtime options, e.g.,
276: ! -ksp_type <type> -pc_type <type> -ksp_monitor -ksp_rtol <rtol>
277: ! These options will override those specified above as long as
278: ! KSPSetFromOptions() is called _after_ any other customization
279: ! routines.
281: PetscCallA(KSPSetFromOptions(ksp, ierr))
283: ! - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
284: ! Solve the linear system
285: ! - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
287: PetscCallA(KSPSolve(ksp, b, x, ierr))
289: ! - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
290: ! Check solution and clean up
291: ! - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
293: ! Check the error
295: PetscCallA(VecAXPY(x, neg_one, u, ierr))
296: PetscCallA(VecNorm(x, NORM_2, norm, ierr))
297: PetscCallA(KSPGetIterationNumber(ksp, its, ierr))
299: if (rank == 0) then
300: if (norm > 1.e-12) then
301: write (6, 100) norm, its
302: else
303: write (6, 110) its
304: end if
305: end if
306: 100 format('Norm of error ', 1pe11.4, ' iterations ', i5)
307: 110 format('Norm of error < 1.e-12,iterations ', i5)
309: ! Free work space. All PETSc objects should be destroyed when they
310: ! are no longer needed.
312: PetscCallA(KSPDestroy(ksp, ierr))
313: PetscCallA(VecDestroy(u, ierr))
314: PetscCallA(VecDestroy(x, ierr))
315: PetscCallA(VecDestroy(b, ierr))
316: PetscCallA(MatDestroy(A, ierr))
318: ! Always call PetscFinalize() before exiting a program.
320: PetscCallA(PetscFinalize(ierr))
321: end program
322: !
323: !/*TEST
324: !
325: ! test:
326: ! nsize: 2
327: ! args: -ksp_view -user_defined_pc -ksp_gmres_cgs_refinement_type refine_always
328: !
329: !TEST*/