xref: /petsc/src/ts/tutorials/autodiff/ex16adj.cxx (revision 58d68138c660dfb4e9f5b03334792cd4f2ffd7cc)
1 static char help[] = "Demonstrates automatic Jacobian generation using ADOL-C for an adjoint sensitivity analysis of the van der Pol equation.\n\
2 Input parameters include:\n\
3       -mu : stiffness parameter\n\n";
4 
5 /*
6    REQUIRES configuration of PETSc with option --download-adolc.
7 
8    For documentation on ADOL-C, see
9      $PETSC_ARCH/externalpackages/ADOL-C-2.6.0/ADOL-C/doc/adolc-manual.pdf
10 */
11 /* ------------------------------------------------------------------------
12    See ex16adj for a description of the problem being solved.
13   ------------------------------------------------------------------------- */
14 
15 #include <petscts.h>
16 #include <petscmat.h>
17 #include "adolc-utils/drivers.cxx"
18 #include <adolc/adolc.h>
19 
20 typedef struct _n_User *User;
21 struct _n_User {
22   PetscReal mu;
23   PetscReal next_output;
24   PetscReal tprev;
25 
26   /* Automatic differentiation support */
27   AdolcCtx *adctx;
28 };
29 
30 /*
31   'Passive' RHS function, used in residual evaluations during the time integration.
32 */
33 static PetscErrorCode RHSFunctionPassive(TS ts, PetscReal t, Vec X, Vec F, void *ctx) {
34   User               user = (User)ctx;
35   PetscScalar       *f;
36   const PetscScalar *x;
37 
38   PetscFunctionBeginUser;
39   PetscCall(VecGetArrayRead(X, &x));
40   PetscCall(VecGetArray(F, &f));
41   f[0] = x[1];
42   f[1] = user->mu * (1. - x[0] * x[0]) * x[1] - x[0];
43   PetscCall(VecRestoreArrayRead(X, &x));
44   PetscCall(VecRestoreArray(F, &f));
45   PetscFunctionReturn(0);
46 }
47 
48 /*
49   Trace RHS to mark on tape 1 the dependence of f upon x. This tape is used in generating the
50   Jacobian transform.
51 */
52 static PetscErrorCode RHSFunctionActive(TS ts, PetscReal t, Vec X, Vec F, void *ctx) {
53   User               user = (User)ctx;
54   PetscScalar       *f;
55   const PetscScalar *x;
56 
57   adouble f_a[2]; /* 'active' double for dependent variables */
58   adouble x_a[2]; /* 'active' double for independent variables */
59 
60   PetscFunctionBeginUser;
61   PetscCall(VecGetArrayRead(X, &x));
62   PetscCall(VecGetArray(F, &f));
63 
64   /* Start of active section */
65   trace_on(1);
66   x_a[0] <<= x[0];
67   x_a[1] <<= x[1]; /* Mark independence */
68   f_a[0] = x_a[1];
69   f_a[1] = user->mu * (1. - x_a[0] * x_a[0]) * x_a[1] - x_a[0];
70   f_a[0] >>= f[0];
71   f_a[1] >>= f[1]; /* Mark dependence */
72   trace_off();
73   /* End of active section */
74 
75   PetscCall(VecRestoreArrayRead(X, &x));
76   PetscCall(VecRestoreArray(F, &f));
77   PetscFunctionReturn(0);
78 }
79 
80 /*
81   Trace RHS again to mark on tape 2 the dependence of f upon the parameter mu. This tape is used in
82   generating JacobianP.
83 */
84 static PetscErrorCode RHSFunctionActiveP(TS ts, PetscReal t, Vec X, Vec F, void *ctx) {
85   User               user = (User)ctx;
86   PetscScalar       *f;
87   const PetscScalar *x;
88 
89   adouble f_a[2];       /* 'active' double for dependent variables */
90   adouble x_a[2], mu_a; /* 'active' double for independent variables */
91 
92   PetscFunctionBeginUser;
93   PetscCall(VecGetArrayRead(X, &x));
94   PetscCall(VecGetArray(F, &f));
95 
96   /* Start of active section */
97   trace_on(3);
98   x_a[0] <<= x[0];
99   x_a[1] <<= x[1];
100   mu_a <<= user->mu; /* Mark independence */
101   f_a[0] = x_a[1];
102   f_a[1] = mu_a * (1. - x_a[0] * x_a[0]) * x_a[1] - x_a[0];
103   f_a[0] >>= f[0];
104   f_a[1] >>= f[1]; /* Mark dependence */
105   trace_off();
106   /* End of active section */
107 
108   PetscCall(VecRestoreArrayRead(X, &x));
109   PetscCall(VecRestoreArray(F, &f));
110   PetscFunctionReturn(0);
111 }
112 
113 /*
114   Compute the Jacobian w.r.t. x using PETSc-ADOL-C driver for explicit TS.
115 */
116 static PetscErrorCode RHSJacobian(TS ts, PetscReal t, Vec X, Mat A, Mat B, void *ctx) {
117   User               user = (User)ctx;
118   const PetscScalar *x;
119 
120   PetscFunctionBeginUser;
121   PetscCall(VecGetArrayRead(X, &x));
122   PetscCall(PetscAdolcComputeRHSJacobian(1, A, x, user->adctx));
123   PetscCall(VecRestoreArrayRead(X, &x));
124   PetscFunctionReturn(0);
125 }
126 
127 /*
128   Compute the Jacobian w.r.t. mu using PETSc-ADOL-C driver for explicit TS.
129 */
130 static PetscErrorCode RHSJacobianP(TS ts, PetscReal t, Vec X, Mat A, void *ctx) {
131   User               user = (User)ctx;
132   const PetscScalar *x;
133 
134   PetscFunctionBeginUser;
135   PetscCall(VecGetArrayRead(X, &x));
136   PetscCall(PetscAdolcComputeRHSJacobianP(3, A, x, &user->mu, user->adctx));
137   PetscCall(VecRestoreArrayRead(X, &x));
138   PetscFunctionReturn(0);
139 }
140 
141 /*
142   Monitor timesteps and use interpolation to output at integer multiples of 0.1
143 */
144 static PetscErrorCode Monitor(TS ts, PetscInt step, PetscReal t, Vec X, void *ctx) {
145   const PetscScalar *x;
146   PetscReal          tfinal, dt, tprev;
147   User               user = (User)ctx;
148 
149   PetscFunctionBeginUser;
150   PetscCall(TSGetTimeStep(ts, &dt));
151   PetscCall(TSGetMaxTime(ts, &tfinal));
152   PetscCall(TSGetPrevTime(ts, &tprev));
153   PetscCall(VecGetArrayRead(X, &x));
154   PetscCall(PetscPrintf(PETSC_COMM_WORLD, "[%.1f] %" PetscInt_FMT " TS %.6f (dt = %.6f) X % 12.6e % 12.6e\n", (double)user->next_output, step, (double)t, (double)dt, (double)PetscRealPart(x[0]), (double)PetscRealPart(x[1])));
155   PetscCall(PetscPrintf(PETSC_COMM_WORLD, "t %.6f (tprev = %.6f) \n", (double)t, (double)tprev));
156   PetscCall(VecRestoreArrayRead(X, &x));
157   PetscFunctionReturn(0);
158 }
159 
160 int main(int argc, char **argv) {
161   TS             ts;   /* nonlinear solver */
162   Vec            x;    /* solution, residual vectors */
163   Mat            A;    /* Jacobian matrix */
164   Mat            Jacp; /* JacobianP matrix */
165   PetscInt       steps;
166   PetscReal      ftime   = 0.5;
167   PetscBool      monitor = PETSC_FALSE;
168   PetscScalar   *x_ptr;
169   PetscMPIInt    size;
170   struct _n_User user;
171   AdolcCtx      *adctx;
172   Vec            lambda[2], mu[2], r;
173 
174   /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
175      Initialize program
176      - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
177   PetscFunctionBeginUser;
178   PetscCall(PetscInitialize(&argc, &argv, NULL, help));
179   PetscCallMPI(MPI_Comm_size(PETSC_COMM_WORLD, &size));
180   PetscCheck(size == 1, PETSC_COMM_WORLD, PETSC_ERR_WRONG_MPI_SIZE, "This is a uniprocessor example only!");
181 
182   /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
183     Set runtime options and create AdolcCtx
184     - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
185   PetscCall(PetscNew(&adctx));
186   user.mu           = 1;
187   user.next_output  = 0.0;
188   adctx->m          = 2;
189   adctx->n          = 2;
190   adctx->p          = 2;
191   adctx->num_params = 1;
192   user.adctx        = adctx;
193 
194   PetscCall(PetscOptionsGetReal(NULL, NULL, "-mu", &user.mu, NULL));
195   PetscCall(PetscOptionsGetBool(NULL, NULL, "-monitor", &monitor, NULL));
196 
197   /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
198     Create necessary matrix and vectors, solve same ODE on every process
199     - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
200   PetscCall(MatCreate(PETSC_COMM_WORLD, &A));
201   PetscCall(MatSetSizes(A, PETSC_DECIDE, PETSC_DECIDE, 2, 2));
202   PetscCall(MatSetFromOptions(A));
203   PetscCall(MatSetUp(A));
204   PetscCall(MatCreateVecs(A, &x, NULL));
205 
206   PetscCall(MatCreate(PETSC_COMM_WORLD, &Jacp));
207   PetscCall(MatSetSizes(Jacp, PETSC_DECIDE, PETSC_DECIDE, 2, 1));
208   PetscCall(MatSetFromOptions(Jacp));
209   PetscCall(MatSetUp(Jacp));
210 
211   /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
212      Create timestepping solver context
213      - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
214   PetscCall(TSCreate(PETSC_COMM_WORLD, &ts));
215   PetscCall(TSSetType(ts, TSRK));
216   PetscCall(TSSetRHSFunction(ts, NULL, RHSFunctionPassive, &user));
217 
218   /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
219      Set initial conditions
220    - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
221   PetscCall(VecGetArray(x, &x_ptr));
222   x_ptr[0] = 2;
223   x_ptr[1] = 0.66666654321;
224   PetscCall(VecRestoreArray(x, &x_ptr));
225 
226   /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
227      Trace just once on each tape and put zeros on Jacobian diagonal
228      - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
229   PetscCall(VecDuplicate(x, &r));
230   PetscCall(RHSFunctionActive(ts, 0., x, r, &user));
231   PetscCall(RHSFunctionActiveP(ts, 0., x, r, &user));
232   PetscCall(VecSet(r, 0));
233   PetscCall(MatDiagonalSet(A, r, INSERT_VALUES));
234   PetscCall(VecDestroy(&r));
235 
236   /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
237      Set RHS Jacobian for the adjoint integration
238      - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
239   PetscCall(TSSetRHSJacobian(ts, A, A, RHSJacobian, &user));
240   PetscCall(TSSetMaxTime(ts, ftime));
241   PetscCall(TSSetExactFinalTime(ts, TS_EXACTFINALTIME_MATCHSTEP));
242   if (monitor) { PetscCall(TSMonitorSet(ts, Monitor, &user, NULL)); }
243   PetscCall(TSSetTimeStep(ts, .001));
244 
245   /*
246     Have the TS save its trajectory so that TSAdjointSolve() may be used
247   */
248   PetscCall(TSSetSaveTrajectory(ts));
249 
250   /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
251      Set runtime options
252    - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
253   PetscCall(TSSetFromOptions(ts));
254 
255   /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
256      Solve nonlinear system
257      - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
258   PetscCall(TSSolve(ts, x));
259   PetscCall(TSGetSolveTime(ts, &ftime));
260   PetscCall(TSGetStepNumber(ts, &steps));
261   PetscCall(PetscPrintf(PETSC_COMM_WORLD, "mu %g, steps %" PetscInt_FMT ", ftime %g\n", (double)user.mu, steps, (double)ftime));
262   PetscCall(VecView(x, PETSC_VIEWER_STDOUT_WORLD));
263 
264   /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
265      Start the Adjoint model
266      - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
267   PetscCall(MatCreateVecs(A, &lambda[0], NULL));
268   PetscCall(MatCreateVecs(A, &lambda[1], NULL));
269   /*   Reset initial conditions for the adjoint integration */
270   PetscCall(VecGetArray(lambda[0], &x_ptr));
271   x_ptr[0] = 1.0;
272   x_ptr[1] = 0.0;
273   PetscCall(VecRestoreArray(lambda[0], &x_ptr));
274   PetscCall(VecGetArray(lambda[1], &x_ptr));
275   x_ptr[0] = 0.0;
276   x_ptr[1] = 1.0;
277   PetscCall(VecRestoreArray(lambda[1], &x_ptr));
278 
279   PetscCall(MatCreateVecs(Jacp, &mu[0], NULL));
280   PetscCall(MatCreateVecs(Jacp, &mu[1], NULL));
281   PetscCall(VecGetArray(mu[0], &x_ptr));
282   x_ptr[0] = 0.0;
283   PetscCall(VecRestoreArray(mu[0], &x_ptr));
284   PetscCall(VecGetArray(mu[1], &x_ptr));
285   x_ptr[0] = 0.0;
286   PetscCall(VecRestoreArray(mu[1], &x_ptr));
287   PetscCall(TSSetCostGradients(ts, 2, lambda, mu));
288 
289   /*   Set RHS JacobianP */
290   PetscCall(TSSetRHSJacobianP(ts, Jacp, RHSJacobianP, &user));
291 
292   PetscCall(TSAdjointSolve(ts));
293 
294   PetscCall(VecView(lambda[0], PETSC_VIEWER_STDOUT_WORLD));
295   PetscCall(VecView(lambda[1], PETSC_VIEWER_STDOUT_WORLD));
296   PetscCall(VecView(mu[0], PETSC_VIEWER_STDOUT_WORLD));
297   PetscCall(VecView(mu[1], PETSC_VIEWER_STDOUT_WORLD));
298 
299   /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
300      Free work space.  All PETSc objects should be destroyed when they
301      are no longer needed.
302    - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
303   PetscCall(MatDestroy(&A));
304   PetscCall(MatDestroy(&Jacp));
305   PetscCall(VecDestroy(&x));
306   PetscCall(VecDestroy(&lambda[0]));
307   PetscCall(VecDestroy(&lambda[1]));
308   PetscCall(VecDestroy(&mu[0]));
309   PetscCall(VecDestroy(&mu[1]));
310   PetscCall(TSDestroy(&ts));
311   PetscCall(PetscFree(adctx));
312   PetscCall(PetscFinalize());
313   return 0;
314 }
315 
316 /*TEST
317 
318   build:
319     requires: double !complex adolc
320 
321   test:
322     suffix: 1
323     args: -ts_max_steps 10 -ts_monitor -ts_adjoint_monitor
324     output_file: output/ex16adj_1.out
325 
326   test:
327     suffix: 2
328     args: -ts_max_steps 10 -ts_monitor -ts_adjoint_monitor -mu 5
329     output_file: output/ex16adj_2.out
330 
331   test:
332     suffix: 3
333     args: -ts_max_steps 10 -monitor
334     output_file: output/ex16adj_3.out
335 
336   test:
337     suffix: 4
338     args: -ts_max_steps 10 -monitor -mu 5
339     output_file: output/ex16adj_4.out
340 
341 TEST*/
342