xref: /petsc/src/snes/interface/snesj.c (revision 95d5f7c29374efcfd2ca44c2fe93981fbc2b4454)
1 /*$Id: snesj.c,v 1.64 2000/04/09 04:38:32 bsmith Exp bsmith $*/
2 
3 #include "src/snes/snesimpl.h"    /*I  "snes.h"  I*/
4 
5 #undef __FUNC__
6 #define __FUNC__ /*<a name=""></a>*/"SNESDefaultComputeJacobian"
7 /*@C
8    SNESDefaultComputeJacobian - Computes the Jacobian using finite differences.
9 
10    Collective on SNES
11 
12    Input Parameters:
13 +  x1 - compute Jacobian at this point
14 -  ctx - application's function context, as set with SNESSetFunction()
15 
16    Output Parameters:
17 +  J - Jacobian matrix (not altered in this routine)
18 .  B - newly computed Jacobian matrix to use with preconditioner (generally the same as J)
19 -  flag - flag indicating whether the matrix sparsity structure has changed
20 
21    Options Database Key:
22 .  -snes_fd - Activates SNESDefaultComputeJacobian()
23 
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.
30 
31    An alternative routine that uses coloring to explot matrix sparsity is
32    SNESDefaultComputeJacobianColor().
33 
34    Level: intermediate
35 
36 .keywords: SNES, finite differences, Jacobian
37 
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   Scalar    dx,mone = -1.0,*y,scale,*xx,wscale;
45   PetscReal amax,epsilon = 1.e-8; /* assumes PetscReal precision */
46   PetscReal dx_min = 1.e-16,dx_par = 1.e-1;
47   MPI_Comm  comm;
48   int      (*eval_fct)(SNES,Vec,Vec)=0;
49 
50   PetscFunctionBegin;
51   if (snes->method_class == SNES_NONLINEAR_EQUATIONS) eval_fct = SNESComputeFunction;
52   else if (snes->method_class == SNES_UNCONSTRAINED_MINIMIZATION) eval_fct = SNESComputeGradient;
53   else SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,0,"Invalid method class");
54 
55   ierr = PetscObjectGetComm((PetscObject)x1,&comm);CHKERRQ(ierr);
56   ierr = MatZeroEntries(*B);CHKERRQ(ierr);
57   if (!snes->nvwork) {
58     ierr = VecDuplicateVecs(x1,3,&snes->vwork);CHKERRQ(ierr);
59     snes->nvwork = 3;
60     PLogObjectParents(snes,3,snes->vwork);
61   }
62   j1a = snes->vwork[0]; j2a = snes->vwork[1]; x2 = snes->vwork[2];
63 
64   ierr = VecGetSize(x1,&N);CHKERRQ(ierr);
65   ierr = VecGetOwnershipRange(x1,&start,&end);CHKERRQ(ierr);
66   ierr = eval_fct(snes,x1,j1a);CHKERRQ(ierr);
67 
68   /* Compute Jacobian approximation, 1 column at a time.
69       x1 = current iterate, j1a = F(x1)
70       x2 = perturbed iterate, j2a = F(x2)
71    */
72   for (i=0; i<N; i++) {
73     ierr = VecCopy(x1,x2);CHKERRQ(ierr);
74     if (i>= start && i<end) {
75       ierr = VecGetArray(x1,&xx);CHKERRQ(ierr);
76       dx = xx[i-start];
77       ierr = VecRestoreArray(x1,&xx);CHKERRQ(ierr);
78 #if !defined(PETSC_USE_COMPLEX)
79       if (dx < dx_min && dx >= 0.0) dx = dx_par;
80       else if (dx < 0.0 && dx > -dx_min) dx = -dx_par;
81 #else
82       if (PetscAbsScalar(dx) < dx_min && PetscRealPart(dx) >= 0.0) dx = dx_par;
83       else if (PetscRealPart(dx) < 0.0 && PetscAbsScalar(dx) < dx_min) dx = -dx_par;
84 #endif
85       dx *= epsilon;
86       wscale = 1.0/dx;
87       ierr = VecSetValues(x2,1,&i,&dx,ADD_VALUES);CHKERRQ(ierr);
88     } else {
89       wscale = 0.0;
90     }
91     ierr = eval_fct(snes,x2,j2a);CHKERRQ(ierr);
92     ierr = VecAXPY(&mone,j1a,j2a);CHKERRQ(ierr);
93     /* Communicate scale to all processors */
94     ierr = MPI_Allreduce(&wscale,&scale,1,MPIU_SCALAR,PetscSum_Op,comm);CHKERRQ(ierr);
95     ierr = VecScale(&scale,j2a);CHKERRQ(ierr);
96     ierr = VecGetArray(j2a,&y);CHKERRQ(ierr);
97     ierr = VecNorm(j2a,NORM_INFINITY,&amax);CHKERRQ(ierr); amax *= 1.e-14;
98     for (j=start; j<end; j++) {
99       if (PetscAbsScalar(y[j-start]) > amax) {
100         ierr = MatSetValues(*B,1,&j,1,&i,y+j-start,INSERT_VALUES);CHKERRQ(ierr);
101       }
102     }
103     ierr = VecRestoreArray(j2a,&y);CHKERRQ(ierr);
104   }
105   ierr  = MatAssemblyBegin(*B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
106   ierr  = MatAssemblyEnd(*B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
107   *flag =  DIFFERENT_NONZERO_PATTERN;
108   PetscFunctionReturn(0);
109 }
110 
111 #undef __FUNC__
112 #define __FUNC__ /*<a name=""></a>*/"SNESDefaultComputeHessian"
113 /*@C
114    SNESDefaultComputeHessian - Computes the Hessian using finite differences.
115 
116    Collective on SNES
117 
118    Input Parameters:
119 +  x1 - compute Hessian at this point
120 -  ctx - application's gradient context, as set with SNESSetGradient()
121 
122    Output Parameters:
123 +  J - Hessian matrix (not altered in this routine)
124 .  B - newly computed Hessian matrix to use with preconditioner (generally the same as J)
125 -  flag - flag indicating whether the matrix sparsity structure has changed
126 
127    Options Database Key:
128 $  -snes_fd - Activates SNESDefaultComputeHessian()
129 
130 
131    Level: intermediate
132 
133    Notes:
134    This routine is slow and expensive, and is not currently optimized
135    to take advantage of sparsity in the problem.  Although
136    SNESDefaultComputeHessian() is not recommended for general use
137    in large-scale applications, It can be useful in checking the
138    correctness of a user-provided Hessian.
139 
140 .keywords: SNES, finite differences, Hessian
141 
142 .seealso: SNESSetHessian()
143 @*/
144 int SNESDefaultComputeHessian(SNES snes,Vec x1,Mat *J,Mat *B,MatStructure *flag,void *ctx)
145 {
146   int ierr;
147 
148   PetscFunctionBegin;
149   ierr = SNESDefaultComputeJacobian(snes,x1,J,B,flag,ctx);CHKERRQ(ierr);
150   PetscFunctionReturn(0);
151 }
152 
153 #undef __FUNC__
154 #define __FUNC__ /*<a name=""></a>*/"SNESDefaultComputeHessianColor"
155 /*@C
156    SNESDefaultComputeHessianColor - Computes the Hessian using colored finite differences.
157 
158    Collective on SNES
159 
160    Input Parameters:
161 +  x1 - compute Hessian at this point
162 -  ctx - application's gradient context, as set with SNESSetGradient()
163 
164    Output Parameters:
165 +  J - Hessian matrix (not altered in this routine)
166 .  B - newly computed Hessian matrix to use with preconditioner (generally the same as J)
167 -  flag - flag indicating whether the matrix sparsity structure has changed
168 
169     Options Database Keys:
170 .  -mat_fd_coloring_freq <freq> - Activates SNESDefaultComputeJacobianColor()
171 
172    Level: intermediate
173 
174  .keywords: SNES, finite differences, Hessian, coloring, sparse
175 
176 .seealso: SNESSetHessian()
177 @*/
178 int SNESDefaultComputeHessianColor(SNES snes,Vec x1,Mat *J,Mat *B,MatStructure *flag,void *ctx)
179 {
180   int ierr;
181 
182   PetscFunctionBegin;
183   ierr = SNESDefaultComputeJacobianColor(snes,x1,J,B,flag,ctx);CHKERRQ(ierr);
184   PetscFunctionReturn(0);
185 }
186 
187