Actual source code: snesj.c
1: /*$Id: snesj.c,v 1.75 2001/09/11 18:06:40 bsmith Exp $*/
3: #include src/snes/snesimpl.h
5: /*@C
6: SNESDefaultComputeJacobian - Computes the Jacobian using finite differences.
8: Collective on SNES
10: Input Parameters:
11: + x1 - compute Jacobian at this point
12: - ctx - application's function context, as set with SNESSetFunction()
14: Output Parameters:
15: + J - Jacobian matrix (not altered in this routine)
16: . B - newly computed Jacobian matrix to use with preconditioner (generally the same as J)
17: - flag - flag indicating whether the matrix sparsity structure has changed
19: Options Database Key:
20: + -snes_fd - Activates SNESDefaultComputeJacobian()
21: - -snes_test_err - Square root of function error tolerance, default square root of machine
22: epsilon (1.e-8 in double, 3.e-4 in single)
24: Notes:
25: This routine is slow and expensive, and is not currently optimized
26: to take advantage of sparsity in the problem. Although
27: SNESDefaultComputeJacobian() is not recommended for general use
28: in large-scale applications, It can be useful in checking the
29: correctness of a user-provided Jacobian.
31: An alternative routine that uses coloring to explot matrix sparsity is
32: SNESDefaultComputeJacobianColor().
34: Level: intermediate
36: .keywords: SNES, finite differences, Jacobian
38: .seealso: SNESSetJacobian(), SNESDefaultComputeJacobianColor()
39: @*/
40: int SNESDefaultComputeJacobian(SNES snes,Vec x1,Mat *J,Mat *B,MatStructure *flag,void *ctx)
41: {
42: Vec j1a,j2a,x2;
43: int i,ierr,N,start,end,j;
44: PetscScalar dx,mone = -1.0,*y,scale,*xx,wscale;
45: PetscReal amax,epsilon = PETSC_SQRT_MACHINE_EPSILON;
46: PetscReal dx_min = 1.e-16,dx_par = 1.e-1;
47: MPI_Comm comm;
48: int (*eval_fct)(SNES,Vec,Vec)=0;
51: PetscOptionsGetReal(snes->prefix,"-snes_test_err",&epsilon,0);
52: if (snes->method_class == SNES_NONLINEAR_EQUATIONS) eval_fct = SNESComputeFunction;
53: else if (snes->method_class == SNES_UNCONSTRAINED_MINIMIZATION) eval_fct = SNESComputeGradient;
54: else SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Invalid method class");
56: PetscObjectGetComm((PetscObject)x1,&comm);
57: MatZeroEntries(*B);
58: if (!snes->nvwork) {
59: VecDuplicateVecs(x1,3,&snes->vwork);
60: snes->nvwork = 3;
61: PetscLogObjectParents(snes,3,snes->vwork);
62: }
63: j1a = snes->vwork[0]; j2a = snes->vwork[1]; x2 = snes->vwork[2];
65: VecGetSize(x1,&N);
66: VecGetOwnershipRange(x1,&start,&end);
67: (*eval_fct)(snes,x1,j1a);
69: /* Compute Jacobian approximation, 1 column at a time.
70: x1 = current iterate, j1a = F(x1)
71: x2 = perturbed iterate, j2a = F(x2)
72: */
73: for (i=0; i<N; i++) {
74: VecCopy(x1,x2);
75: if (i>= start && i<end) {
76: VecGetArray(x1,&xx);
77: dx = xx[i-start];
78: VecRestoreArray(x1,&xx);
79: #if !defined(PETSC_USE_COMPLEX)
80: if (dx < dx_min && dx >= 0.0) dx = dx_par;
81: else if (dx < 0.0 && dx > -dx_min) dx = -dx_par;
82: #else
83: if (PetscAbsScalar(dx) < dx_min && PetscRealPart(dx) >= 0.0) dx = dx_par;
84: else if (PetscRealPart(dx) < 0.0 && PetscAbsScalar(dx) < dx_min) dx = -dx_par;
85: #endif
86: dx *= epsilon;
87: wscale = 1.0/dx;
88: VecSetValues(x2,1,&i,&dx,ADD_VALUES);
89: } else {
90: wscale = 0.0;
91: }
92: (*eval_fct)(snes,x2,j2a);
93: VecAXPY(&mone,j1a,j2a);
94: /* Communicate scale to all processors */
95: MPI_Allreduce(&wscale,&scale,1,MPIU_SCALAR,PetscSum_Op,comm);
96: VecScale(&scale,j2a);
97: VecNorm(j2a,NORM_INFINITY,&amax); amax *= 1.e-14;
98: VecGetArray(j2a,&y);
99: for (j=start; j<end; j++) {
100: if (PetscAbsScalar(y[j-start]) > amax) {
101: MatSetValues(*B,1,&j,1,&i,y+j-start,INSERT_VALUES);
102: }
103: }
104: VecRestoreArray(j2a,&y);
105: }
106: ierr = MatAssemblyBegin(*B,MAT_FINAL_ASSEMBLY);
107: ierr = MatAssemblyEnd(*B,MAT_FINAL_ASSEMBLY);
108: *flag = DIFFERENT_NONZERO_PATTERN;
109: return(0);
110: }
112: /*@C
113: SNESDefaultComputeHessian - Computes the Hessian using finite differences.
115: Collective on SNES
117: Input Parameters:
118: + x1 - compute Hessian at this point
119: - ctx - application's gradient context, as set with SNESSetGradient()
121: Output Parameters:
122: + J - Hessian matrix (not altered in this routine)
123: . B - newly computed Hessian matrix to use with preconditioner (generally the same as J)
124: - flag - flag indicating whether the matrix sparsity structure has changed
126: Options Database Key:
127: $ -snes_fd - Activates SNESDefaultComputeHessian()
130: Level: intermediate
132: Notes:
133: This routine is slow and expensive, and is not currently optimized
134: to take advantage of sparsity in the problem. Although
135: SNESDefaultComputeHessian() is not recommended for general use
136: in large-scale applications, It can be useful in checking the
137: correctness of a user-provided Hessian.
139: .keywords: SNES, finite differences, Hessian
141: .seealso: SNESSetHessian()
142: @*/
143: int SNESDefaultComputeHessian(SNES snes,Vec x1,Mat *J,Mat *B,MatStructure *flag,void *ctx)
144: {
148: SNESDefaultComputeJacobian(snes,x1,J,B,flag,ctx);
149: return(0);
150: }
152: /*@C
153: SNESDefaultComputeHessianColor - Computes the Hessian using colored finite differences.
155: Collective on SNES
157: Input Parameters:
158: + x1 - compute Hessian at this point
159: - ctx - application's gradient context, as set with SNESSetGradient()
161: Output Parameters:
162: + J - Hessian matrix (not altered in this routine)
163: . B - newly computed Hessian matrix to use with preconditioner (generally the same as J)
164: - flag - flag indicating whether the matrix sparsity structure has changed
166: Options Database Keys:
167: . -mat_fd_coloring_freq <freq> - Activates SNESDefaultComputeJacobianColor()
169: Level: intermediate
171: .keywords: SNES, finite differences, Hessian, coloring, sparse
173: .seealso: SNESSetHessian()
174: @*/
175: int SNESDefaultComputeHessianColor(SNES snes,Vec x1,Mat *J,Mat *B,MatStructure *flag,void *ctx)
176: {
180: SNESDefaultComputeJacobianColor(snes,x1,J,B,flag,ctx);
181: return(0);
182: }