xref: /petsc/src/snes/interface/snesj.c (revision e2df7a95c5ea77c899beea10ff9effd6061e7c8f)
1 #define PETSCSNES_DLL
2 
3 #include "src/snes/snesimpl.h"    /*I  "petscsnes.h"  I*/
4 
5 #undef __FUNCT__
6 #define __FUNCT__ "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 .  -snes_test_err - Square root of function error tolerance, default square root of machine
24                     epsilon (1.e-8 in double, 3.e-4 in single)
25 -  -mat_fd_type - Either wp or ds (see MATSNESMF_WP or MATSNESMF_DS)
26 
27    Notes:
28    This routine is slow and expensive, and is not currently optimized
29    to take advantage of sparsity in the problem.  Although
30    SNESDefaultComputeJacobian() is not recommended for general use
31    in large-scale applications, It can be useful in checking the
32    correctness of a user-provided Jacobian.
33 
34    An alternative routine that uses coloring to exploit matrix sparsity is
35    SNESDefaultComputeJacobianColor().
36 
37    Level: intermediate
38 
39 .keywords: SNES, finite differences, Jacobian
40 
41 .seealso: SNESSetJacobian(), SNESDefaultComputeJacobianColor(), MatCreateSNESMF()
42 @*/
43 PetscErrorCode PETSCSNES_DLLEXPORT SNESDefaultComputeJacobian(SNES snes,Vec x1,Mat *J,Mat *B,MatStructure *flag,void *ctx)
44 {
45   Vec            j1a,j2a,x2;
46   PetscErrorCode ierr;
47   PetscInt       i,N,start,end,j,value;
48   PetscScalar    dx,*y,scale,*xx,wscale;
49   PetscReal      amax,epsilon = PETSC_SQRT_MACHINE_EPSILON;
50   PetscReal      dx_min = 1.e-16,dx_par = 1.e-1,unorm;
51   MPI_Comm       comm;
52   PetscErrorCode (*eval_fct)(SNES,Vec,Vec)=0;
53   PetscTruth     assembled,use_wp = PETSC_TRUE,flg;
54   const char     *list[2] = {"ds","wp"};
55 
56   PetscFunctionBegin;
57   ierr = PetscOptionsGetReal(snes->prefix,"-snes_test_err",&epsilon,0);CHKERRQ(ierr);
58   eval_fct = SNESComputeFunction;
59 
60   ierr = PetscObjectGetComm((PetscObject)x1,&comm);CHKERRQ(ierr);
61   ierr = MatAssembled(*B,&assembled);CHKERRQ(ierr);
62   if (assembled) {
63     ierr = MatZeroEntries(*B);CHKERRQ(ierr);
64   }
65   if (!snes->nvwork) {
66     ierr = VecDuplicateVecs(x1,3,&snes->vwork);CHKERRQ(ierr);
67     snes->nvwork = 3;
68     ierr = PetscLogObjectParents(snes,3,snes->vwork);CHKERRQ(ierr);
69   }
70   j1a = snes->vwork[0]; j2a = snes->vwork[1]; x2 = snes->vwork[2];
71 
72   ierr = VecGetSize(x1,&N);CHKERRQ(ierr);
73   ierr = VecGetOwnershipRange(x1,&start,&end);CHKERRQ(ierr);
74   ierr = (*eval_fct)(snes,x1,j1a);CHKERRQ(ierr);
75 
76   ierr = PetscOptionsEList("-mat_fd_type","Algorithm to compute difference parameter","SNESDefaultComputeJacobian",list,2,"wp",&value,&flg);CHKERRQ(ierr);
77   if (flg && !value) {
78     use_wp = PETSC_FALSE;
79   }
80   if (use_wp) {
81     ierr = VecNorm(x1,NORM_2,&unorm);CHKERRQ(ierr);
82   }
83   /* Compute Jacobian approximation, 1 column at a time.
84       x1 = current iterate, j1a = F(x1)
85       x2 = perturbed iterate, j2a = F(x2)
86    */
87   for (i=0; i<N; i++) {
88     ierr = VecCopy(x1,x2);CHKERRQ(ierr);
89     if (i>= start && i<end) {
90       ierr = VecGetArray(x1,&xx);CHKERRQ(ierr);
91       if (use_wp) {
92         dx = 1.0 + unorm;
93       } else {
94         dx = xx[i-start];
95       }
96       ierr = VecRestoreArray(x1,&xx);CHKERRQ(ierr);
97 #if !defined(PETSC_USE_COMPLEX)
98       if (dx < dx_min && dx >= 0.0) dx = dx_par;
99       else if (dx < 0.0 && dx > -dx_min) dx = -dx_par;
100 #else
101       if (PetscAbsScalar(dx) < dx_min && PetscRealPart(dx) >= 0.0) dx = dx_par;
102       else if (PetscRealPart(dx) < 0.0 && PetscAbsScalar(dx) < dx_min) dx = -dx_par;
103 #endif
104       dx *= epsilon;
105       wscale = 1.0/dx;
106       ierr = VecSetValues(x2,1,&i,&dx,ADD_VALUES);CHKERRQ(ierr);
107     } else {
108       wscale = 0.0;
109     }
110     ierr = (*eval_fct)(snes,x2,j2a);CHKERRQ(ierr);
111     ierr = VecAXPY(j2a,-1.0,j1a);CHKERRQ(ierr);
112     /* Communicate scale to all processors */
113     ierr = MPI_Allreduce(&wscale,&scale,1,MPIU_SCALAR,PetscSum_Op,comm);CHKERRQ(ierr);
114     ierr = VecScale(j2a,scale);CHKERRQ(ierr);
115     ierr = VecNorm(j2a,NORM_INFINITY,&amax);CHKERRQ(ierr); amax *= 1.e-14;
116     ierr = VecGetArray(j2a,&y);CHKERRQ(ierr);
117     for (j=start; j<end; j++) {
118       if (PetscAbsScalar(y[j-start]) > amax) {
119         ierr = MatSetValues(*B,1,&j,1,&i,y+j-start,INSERT_VALUES);CHKERRQ(ierr);
120       }
121     }
122     ierr = VecRestoreArray(j2a,&y);CHKERRQ(ierr);
123   }
124   ierr  = MatAssemblyBegin(*B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
125   ierr  = MatAssemblyEnd(*B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
126   if (*B != *J) {
127     ierr  = MatAssemblyBegin(*J,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
128     ierr  = MatAssemblyEnd(*J,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
129   }
130   *flag =  DIFFERENT_NONZERO_PATTERN;
131   PetscFunctionReturn(0);
132 }
133 
134 
135